State Space of Arxiv papers on LLMs, 2025

{
  "core_topic": {
    "domain": "Contemporary AI Research Landscape",
    "abstract": "The field spans efforts to extend model capabilities through architectural innovation, improve efficiency and deployment practicality, and develop more autonomous and reliable systems. Key tensions exist between scaling-based approaches versus efficiency optimization, between end-to-end learning versus compositional reasoning, and between pure prediction versus interactive planning. Research simultaneously pushes capability frontiers while addressing fundamental bottlenecks in computation, memory, and alignment.",
    "central_question": "What architectural, training, and deployment strategies best advance AI systems toward general-purpose intelligence while remaining practical and controllable?"
  },
  "state_features": {
    "model_architecture": [
      "world-model",
      "retrieval-augmented",
      "diffusion-based",
      "transformer-standard",
      "transformer-modified",
      "hybrid-architecture"
    ],
    "training_paradigm": [
      "reinforcement-learning",
      "self-evolution",
      "pretraining-only",
      "supervised-finetuning",
      "distillation",
      "test-time-adaptation"
    ],
    "modality_focus": [
      "embodied-3d",
      "language-only",
      "video-temporal",
      "vision-language",
      "multimodal-unified",
      "cross-modal"
    ],
    "computational_strategy": [
      "scaling-capability",
      "memory-optimization",
      "architecture-innovation",
      "inference-speed",
      "efficiency-compression",
      "distributed-decentralized"
    ],
    "autonomy_level": [
      "agentic-planning",
      "self-improving",
      "passive-prediction",
      "agentic-reactive",
      "tool-augmented",
      "autonomous-exploration"
    ],
    "generalization_mechanism": [
      "transfer-learning",
      "continual-learning",
      "zero-shot-emergence",
      "meta-learning",
      "few-shot-adaptation",
      "in-domain-fitting"
    ],
    "reasoning_approach": [
      "model-based-simulation",
      "iterative-refinement",
      "direct-prediction",
      "chain-of-thought",
      "search-based",
      "test-time-compute"
    ],
    "representation_type": [
      "latent-diffusion",
      "hierarchical-memory",
      "continuous-embeddings",
      "discrete-tokens",
      "hybrid-mixed",
      "structured-symbolic"
    ],
    "research_objective": [
      "capability-extension",
      "safety-alignment",
      "evaluation-benchmarking",
      "efficiency-optimization",
      "interpretability-analysis",
      "robustness-reliability"
    ],
    "interaction_model": [
      "fully-autonomous",
      "feedback-driven",
      "conditional-sampling",
      "human-in-loop",
      "interactive-controllable",
      "static-generation"
    ],
    "data_dependency": [
      "self-supervised-unlabeled",
      "synthetic-generated",
      "human-feedback",
      "supervised-labeled",
      "environment-grounded",
      "zero-shot-prompted"
    ],
    "temporal_horizon": [
      "long-horizon-episodic",
      "continual-indefinite",
      "fixed-sequence",
      "variable-length",
      "single-step",
      "meta-across-tasks"
    ],
    "primary_bottleneck": [
      "compute-cost",
      "data-quality",
      "alignment-specification",
      "exploration-coverage",
      "architecture-expressiveness",
      "memory-capacity"
    ]
  },
  "positions": {
    "dreamer_4_world_models": {
      "model_architecture": "world-model",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "compute-cost"
    },
    "reasoningbank_memory": {
      "model_architecture": "retrieval-augmented",
      "training_paradigm": "self-evolution",
      "modality_focus": "language-only",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "self-improving",
      "generalization_mechanism": "continual-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "hierarchical-memory",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "continual-indefinite",
      "primary_bottleneck": "data-quality"
    },
    "veo3_video_reasoning": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "video-temporal",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "zero-shot-emergence",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "compute-cost"
    },
    "rlhi_conversation_learning": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "continual-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "safety-alignment",
      "interaction_model": "human-in-loop",
      "data_dependency": "human-feedback",
      "temporal_horizon": "continual-indefinite",
      "primary_bottleneck": "alignment-specification"
    },
    "mcpmark_tool_agents": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "discrete-tokens",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "didi_instruct_fast_generation": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "distillation",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "compute-cost"
    },
    "lvr_latent_visual_reasoning": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "truthrl_hallucination_reduction": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "safety-alignment",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "memgen_self_evolving_agents": {
      "model_architecture": "retrieval-augmented",
      "training_paradigm": "self-evolution",
      "modality_focus": "language-only",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "self-improving",
      "generalization_mechanism": "continual-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "hierarchical-memory",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "memory-capacity"
    },
    "swax_long_context_hybrid": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "hybrid-mixed",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "memory-capacity"
    },
    "knapsack_rl_budget_allocation": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "meta-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "transformer_multiplication_circuits": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "in-domain-fitting",
      "reasoning_approach": "direct-prediction",
      "representation_type": "structured-symbolic",
      "research_objective": "interpretability-analysis",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "realunify_multimodal_synergy": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "hybrid-mixed",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "muon_optimizer_tail_learning": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "data-quality"
    },
    "gem_agentic_rl_framework": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "brorl_scaling_exploration": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "autonomous-exploration",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "omniretarget_humanoid_motion": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "zero-shot-emergence",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "data-quality"
    },
    "sft_generalization_debunk": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "few-shot-adaptation",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "discrete-tokens",
      "research_objective": "interpretability-analysis",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "data-quality"
    },
    "dc_videogen_compression": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "video-temporal",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "compute-cost"
    },
    "dparallel_diffusion_llm": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "deepscientist_autonomous_research": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "autonomous-exploration",
      "generalization_mechanism": "meta-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "exploration-coverage"
    },
    "vla_rft_world_sim": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "continuous-embeddings",
      "research_objective": "robustness-reliability",
      "interaction_model": "fully-autonomous",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "data-quality"
    },
    "eliciting_secret_knowledge": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "in-domain-fitting",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "alignment-specification"
    },
    "rewatch_r1_video_reasoning": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "video-temporal",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "data-quality"
    },
    "infllm_v2_sparse_attention": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "memory-capacity"
    },
    "deepsearch_mcts_rlvr": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "tooluniverse_ai_scientist": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "exploration-coverage"
    },
    "thinking_sparks_attention_heads": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "parallax_decentralized_inference": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "distributed-decentralized",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "compute-cost"
    },
    "longlive_streaming_video": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "video-temporal",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "continual-indefinite",
      "primary_bottleneck": "memory-capacity"
    },
    "entropy_after_think_early_exit": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "test-time-compute",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "conditional-sampling",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "sparseserve_hierarchical_kv": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "memory-capacity"
    },
    "ttt3r_test_time_3d": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "test-time-adaptation",
      "modality_focus": "embodied-3d",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "continual-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "memory-capacity"
    },
    "dragon_hatchling_brain_bridge": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "polychromic_rl_objectives": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "robustness-reliability",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "diffusion_memorization_mechanism": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "in-domain-fitting",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "interpretability-analysis",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "data-quality"
    },
    "nvfp4_low_precision": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "compute-cost"
    },
    "salt_frozen_teacher": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "self-evolution",
      "modality_focus": "video-temporal",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "compute-cost"
    },
    "gat_scalable_gans": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "pretraining-only",
      "modality_focus": "vision-language",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "compute-cost"
    },
    "llava_onevision_open_training": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "data-quality"
    },
    "regression_language_models": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "cross-modal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "ra3_action_abstractions": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "meta-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "structured-symbolic",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "visual_jigsaw_post_training": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "data-quality"
    },
    "spherear_hyperspherical_latents": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "pretraining-only",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "univid_unified_video": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "video-temporal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "sequential_diffusion_llm": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "evoworld_panoramic_generation": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "pretraining-only",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "hybrid-mixed",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "memory-capacity"
    },
    "da2_panoramic_depth": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "zero-shot-emergence",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "data-quality"
    },
    "rolling_forcing_autoregressive": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "video-temporal",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "latent-diffusion",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "continual-indefinite",
      "primary_bottleneck": "compute-cost"
    },
    "vggt_x_dense_nvs": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "memory-capacity"
    },
    "love_r1_adaptive_zoom": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "video-temporal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "query_kontext_multimodal": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "hybrid-mixed",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "kv_cache_compression_pitfalls": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "robustness-reliability",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "memory-capacity"
    },
    "reseek_search_agents": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "alpharl_predictable_dynamics": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "dc_gen_compressed_latent": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "vision-language",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "memory-capacity"
    },
    "fast_dllm_v2_block_diffusion": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "distillation",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "ar3po_adaptive_rollout": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "piper_env_setup": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "concrete_score_distillation": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "distillation",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "compute-cost"
    },
    "dyva_world_language_models": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "rlmt_thinking_chat": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "human-feedback",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "vapo_vision_anchored": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "robustness-reliability",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "mem_alpha_memory_construction": {
      "model_architecture": "retrieval-augmented",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "continual-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "hierarchical-memory",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "memory-capacity"
    },
    "believing_without_seeing_quality": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "safety-alignment",
      "interaction_model": "human-in-loop",
      "data_dependency": "human-feedback",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "growing_winning_subnetworks": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "vision-language",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "img_implicit_guidance": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "latent-diffusion",
      "research_objective": "safety-alignment",
      "interaction_model": "feedback-driven",
      "data_dependency": "human-feedback",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "preconditioning_feature_learning": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "crostata_cross_state_attention": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "data-quality"
    },
    "mobilellm_r1_sub_billion": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "conditional-sampling",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "rfg_reward_free_guidance": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "test-time-adaptation",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "test-time-compute",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "conditional-sampling",
      "data_dependency": "zero-shot-prompted",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "vlm_fo1_fine_grained_perception": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "hybrid-mixed",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "icl_scaling_laws_theory": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "few-shot-adaptation",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "conditional-sampling",
      "data_dependency": "zero-shot-prompted",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "quasar_quantum_assembly": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "data-quality"
    },
    "rover_random_policy_valuation": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "drivee2e_closed_loop": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "continuous-embeddings",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "fully-autonomous",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "data-quality"
    },
    "airoa_moma_dataset": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "fully-autonomous",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "data-quality"
    },
    "dvla_diffusion_vla": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "compute-cost"
    },
    "hamlet_history_aware_vla": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "memory-optimization",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "hierarchical-memory",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "memory-capacity"
    },
    "longcodezip_context_compression": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "memory-capacity"
    },
    "normal_abnormal_gad": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "data-quality"
    },
    "ttt_in_context_learning": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "test-time-adaptation",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "few-shot-adaptation",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "zero-shot-prompted",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "point_it_out_grounding": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "synchrony_stability_chatbots": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "continual-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "safety-alignment",
      "interaction_model": "human-in-loop",
      "data_dependency": "human-feedback",
      "temporal_horizon": "continual-indefinite",
      "primary_bottleneck": "alignment-specification"
    },
    "cardiobench_echo_foundation": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "pretraining-only",
      "modality_focus": "video-temporal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "zero-shot-emergence",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "data-quality"
    },
    "sana_video_block_linear": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "video-temporal",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "compute-cost"
    },
    "fastforward_camera_localization": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "zero-shot-emergence",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "compute-cost"
    },
    "softcfg_autoregressive_guidance": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "test-time-adaptation",
      "modality_focus": "vision-language",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "robustness-reliability",
      "interaction_model": "conditional-sampling",
      "data_dependency": "zero-shot-prompted",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "infoagent_deep_research": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "oceangym_underwater_agents": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "continuous-embeddings",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "deepsketcher_visual_manipulation": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "data-quality"
    },
    "hearing_order_audio_llm": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "cross-modal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "robustness-reliability",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "cogito_ergo_ludo_reasoning": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "self-evolution",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "autonomous-exploration",
      "generalization_mechanism": "meta-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "structured-symbolic",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "diffusion_bridge_flow_matching": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "pretraining-only",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "interpretability-analysis",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "editscore_rl_reward": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "continuous-embeddings",
      "research_objective": "safety-alignment",
      "interaction_model": "feedback-driven",
      "data_dependency": "human-feedback",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "up2you_unconstrained_3d": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "hybrid-mixed",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "compute-cost"
    },
    "physicsminions_olympiad": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "self-evolution",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "meta-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "visualoverload_dense_scenes": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "spinbench_spatial_reasoning": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "dolphin_ultrasound_foundation": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "data-quality"
    },
    "llm_fingerprint_robustness": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "robustness-reliability",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "alignment-specification"
    },
    "generalized_info_bottleneck": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "bayesian_influence_functions": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "pretraining-only",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "compute-cost"
    },
    "rl_guided_data_selection": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "meta-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "feedback-driven",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "data-quality"
    },
    "clip_entropy_dynamics": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "interpretability-analysis",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "code2video_educational": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "video-temporal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "structured-symbolic",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "instant4d_casual_video": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "socratic_zero_coevolution": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "self-evolution",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "autonomous-exploration",
      "generalization_mechanism": "meta-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "meta-across-tasks",
      "primary_bottleneck": "data-quality"
    },
    "cj_grpo_consistency_trajectory": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "world_env_vla_simulator": {
      "model_architecture": "world-model",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "model-based-simulation",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "fully-autonomous",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "data-quality"
    },
    "agent_matrix_self_replication": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "autonomous-exploration",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "safety-alignment",
      "interaction_model": "fully-autonomous",
      "data_dependency": "environment-grounded",
      "temporal_horizon": "continual-indefinite",
      "primary_bottleneck": "alignment-specification"
    },
    "agent_failure_analysis": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "tool-augmented",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "robustness-reliability",
      "interaction_model": "interactive-controllable",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "long-horizon-episodic",
      "primary_bottleneck": "exploration-coverage"
    },
    "deepfake_gaze_detection": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "video-temporal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "robustness-reliability",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "fixed-sequence",
      "primary_bottleneck": "data-quality"
    },
    "triangle_splatting_plus": {
      "model_architecture": "hybrid-architecture",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "embodied-3d",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "hybrid-mixed",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "noisy_label_detection": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "robustness-reliability",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "data-quality"
    },
    "calibrating_autoraters": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "safety-alignment",
      "interaction_model": "feedback-driven",
      "data_dependency": "human-feedback",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "authentic_discrete_diffusion": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "cmt_consistency_midtraining": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "efficiency-compression",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "latent-diffusion",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "align_visual_encoders": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "compute-cost"
    },
    "learn2pd_adaptive_decoding": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "efficiency-optimization",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "tr2_d2_tree_search_diffusion": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "language-only",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-planning",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "search-based",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "imageedit_video_subject": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "test-time-adaptation",
      "modality_focus": "video-temporal",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "zero-shot-emergence",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "latent-diffusion",
      "research_objective": "capability-extension",
      "interaction_model": "interactive-controllable",
      "data_dependency": "zero-shot-prompted",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "critpt_physics_reasoning": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "language-only",
      "computational_strategy": "scaling-capability",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "chain-of-thought",
      "representation_type": "discrete-tokens",
      "research_objective": "evaluation-benchmarking",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "exploration-coverage"
    },
    "rethinking_reward_models": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "reinforcement-learning",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "agentic-reactive",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "test-time-compute",
      "representation_type": "continuous-embeddings",
      "research_objective": "interpretability-analysis",
      "interaction_model": "feedback-driven",
      "data_dependency": "synthetic-generated",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "alignment-specification"
    },
    "fusion_of_n": {
      "model_architecture": "transformer-standard",
      "training_paradigm": "test-time-adaptation",
      "modality_focus": "language-only",
      "computational_strategy": "inference-speed",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "few-shot-adaptation",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "zero-shot-prompted",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "compute-cost"
    },
    "molm_lora_markers": {
      "model_architecture": "diffusion-based",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "latent-diffusion",
      "research_objective": "robustness-reliability",
      "interaction_model": "static-generation",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "alignment-specification"
    },
    "jepa_t_joint_embedding": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "supervised-finetuning",
      "modality_focus": "multimodal-unified",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "direct-prediction",
      "representation_type": "continuous-embeddings",
      "research_objective": "capability-extension",
      "interaction_model": "conditional-sampling",
      "data_dependency": "supervised-labeled",
      "temporal_horizon": "single-step",
      "primary_bottleneck": "architecture-expressiveness"
    },
    "visual_self_refinement": {
      "model_architecture": "transformer-modified",
      "training_paradigm": "test-time-adaptation",
      "modality_focus": "vision-language",
      "computational_strategy": "architecture-innovation",
      "autonomy_level": "passive-prediction",
      "generalization_mechanism": "transfer-learning",
      "reasoning_approach": "iterative-refinement",
      "representation_type": "discrete-tokens",
      "research_objective": "capability-extension",
      "interaction_model": "static-generation",
      "data_dependency": "self-supervised-unlabeled",
      "temporal_horizon": "variable-length",
      "primary_bottleneck": "architecture-expressiveness"
    }
  },
  "positions_metadata": {
    "dreamer_4_world_models": {
      "description": "Trains agents inside scalable world models to achieve complex long-horizon tasks like diamond acquisition in Minecraft purely from offline data. Enables real-time simulation of game mechanics on single GPU for policy learning without environment interaction.",
      "reference": "Training Agents Inside of Scalable World Models - Google DeepMind"
    },
    "reasoningbank_memory": {
      "description": "Memory framework enabling LLM agents to continuously learn by distilling generalizable reasoning strategies from both successful and failed experiences. Improves success rates and reduces interaction steps through accumulated strategic knowledge.",
      "reference": "ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory"
    },
    "veo3_video_reasoning": {
      "description": "Large-scale generative video model demonstrating emergent zero-shot learning and reasoning across diverse visual tasks. Performs perception, world modeling, and chain-of-frames reasoning through text prompts and initial images without task-specific training.",
      "reference": "Video models are zero-shot learners and reasoners - Google DeepMind"
    },
    "rlhi_conversation_learning": {
      "description": "Reinforcement learning paradigm that continuously aligns and personalizes LLMs by learning directly from natural user conversations. Leverages in-the-wild feedback and long-term personas to improve instruction-following and reasoning without curated datasets.",
      "reference": "The Era of Real-World Human Interaction: RL from User Conversations - FAIR at Meta"
    },
    "mcpmark_tool_agents": {
      "description": "Benchmark evaluating LLM agents' ability to interact with external systems using Model Context Protocol across 127 complex multi-step tasks. Tests agentic capabilities in realistic environments like Notion, GitHub, and databases.",
      "reference": "MCPMark: A Benchmark for Stress-Testing Realistic and Comprehensive MCP Usage"
    },
    "didi_instruct_fast_generation": {
      "description": "Training-based method accelerating language generation from discrete diffusion models by 64x through distillation. Achieves strong perplexity with minimal entropy loss and significantly reduced training time compared to standard approaches.",
      "reference": "Ultra-Fast Language Generation via Discrete Diffusion Divergence Instruct"
    },
    "lvr_latent_visual_reasoning": {
      "description": "Paradigm enabling multimodal LLMs to perform autoregressive reasoning directly within visual embedding space. Achieves state-of-the-art performance on perception-intensive visual question answering through enhanced fine-grained visual understanding.",
      "reference": "Latent Visual Reasoning - Advanced Micro Devices, UC Davis"
    },
    "truthrl_hallucination_reduction": {
      "description": "Reinforcement learning framework optimizing LLM truthfulness through ternary reward system that encourages abstention when uncertain. Reduces hallucination rates by 29% while improving truthfulness scores across benchmarks.",
      "reference": "TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning - Meta Reality Labs"
    },
    "memgen_self_evolving_agents": {
      "description": "Generative latent memory framework enabling LLM agents to dynamically interweave memory synthesis with ongoing reasoning. Mitigates catastrophic forgetting through dedicated memory weaver while keeping core model frozen.",
      "reference": "MemGen: Weaving Generative Latent Memory for Self-Evolving Agents - National University of Singapore"
    },
    "swax_long_context_hybrid": {
      "description": "Hybrid architecture combining sliding window attention with xLSTM layers for efficient long-sequence processing. Discovers that shorter attention windows during training paradoxically improve long-term memorization up to 131k tokens.",
      "reference": "Short window attention enables long-term memorization - Meta FAIR"
    },
    "knapsack_rl_budget_allocation": {
      "description": "Formulates LLM exploration budget allocation as knapsack optimization problem. Adaptively distributes computational resources based on task difficulty, increasing effective gradient signals by 20-40% during training.",
      "reference": "Knapsack RL: Unlocking Exploration of LLMs via Optimizing Budget Allocation"
    },
    "transformer_multiplication_circuits": {
      "description": "Reverse-engineers transformer model successfully learning multiplication via implicit chain-of-thought. Reveals mechanisms using attention to construct directed graphs for caching partial products with Fourier-basis digit representations.",
      "reference": "Why Can't Transformers Learn Multiplication? - Google DeepMind"
    },
    "realunify_multimodal_synergy": {
      "description": "Benchmark evaluating whether unified multimodal models achieve true synergy between understanding and generation capabilities. Tests bidirectional enhancement where reasoning guides generation and simulation facilitates comprehension.",
      "reference": "RealUnify: Do Unified Models Truly Benefit from Unification?"
    },
    "muon_optimizer_tail_learning": {
      "description": "Discovers Muon optimizer's superior LLM performance stems from fostering isotropic representations in transformer components. Enables more balanced acquisition of rare, tail-end knowledge compared to Adam optimization.",
      "reference": "Muon Outperforms Adam in Tail-End Associative Memory Learning - NUS"
    },
    "gem_agentic_rl_framework": {
      "description": "Open-source environment simulator providing standardized platform for reinforcement learning research with agentic LLMs. Facilitates development of multi-turn, long-horizon agents capable of tool use.",
      "reference": "GEM: A Gym for Agentic LLMs - NUS, Stanford"
    },
    "brorl_scaling_exploration": {
      "description": "Scales reinforcement learning by increasing rollouts per example rather than just training steps. Enables continuous performance gains where prior methods plateau, doubling generation throughput for reasoning tasks.",
      "reference": "BroRL: Scaling Reinforcement Learning via Broadened Exploration - NVIDIA, Stanford"
    },
    "omniretarget_humanoid_motion": {
      "description": "Framework generating high-fidelity, interaction-preserving kinematic references for humanoid robots. Enables zero-shot sim-to-real transfer of complex whole-body manipulation skills with significantly reduced motion artifacts.",
      "reference": "OmniRetarget: Interaction-Preserving Data Generation for Humanoid Loco-Manipulation - Amazon FAR"
    },
    "sft_generalization_debunk": {
      "description": "Challenges prevailing view that supervised fine-tuning fails to generalize compared to reinforcement learning. Shows prompt diversity during training breaks frozen-prompt artifacts, achieving strong generalization without RL complexity.",
      "reference": "Debunk the Myth of SFT Generalization - Boston University, LinkedIn"
    },
    "dc_videogen_compression": {
      "description": "Acceleration framework enabling efficient video generation through deep compression video autoencoder. Achieves 14.8x lower latency and 230x reduced training costs while maintaining or improving quality.",
      "reference": "DC-VideoGen: Efficient Video Generation with Deep Compression Video Autoencoder"
    },
    "dparallel_diffusion_llm": {
      "description": "Introduces learnable parallel decoding strategy for diffusion language models addressing sequential certainty convergence. Achieves 10.5x speedup while maintaining or improving benchmark performance through adaptive parallelization.",
      "reference": "dParallel: Learnable Parallel Decoding for dLLMs - National University of Singapore"
    },
    "deepscientist_autonomous_research": {
      "description": "System conducting goal-oriented, fully autonomous scientific discovery over month-long timelines through Bayesian optimization. Generated 5000 ideas, validated 1100 experiments, surpassing human SOTA methods by up to 184%.",
      "reference": "DeepScientist: Advancing Frontier-Pushing Scientific Findings Progressively - Westlake University"
    },
    "vla_rft_world_sim": {
      "description": "Reinforcement fine-tuning framework for vision-language-action models leveraging data-driven world model as controllable simulator. Enables efficient policy learning with trajectory-level rewards derived from goal-achieving references.",
      "reference": "VLA-RFT: Vision-Language-Action Reinforcement Fine-tuning - Fudan University"
    },
    "eliciting_secret_knowledge": {
      "description": "Develops methods to elicit hidden knowledge from LLMs fine-tuned to conceal specific information. Demonstrates both black-box prefill attacks and white-box interpretability techniques successfully uncover concealed knowledge.",
      "reference": "Eliciting Secret Knowledge from Language Models - Anthropic, Oxford"
    },
    "rewatch_r1_video_reasoning": {
      "description": "Multi-agent synthesis pipeline and reinforcement learning framework for complex video reasoning. Introduces observation and reasoning reward mechanism directly penalizing hallucination while evaluating trajectory correctness.",
      "reference": "ReWatch-R1: Boosting Complex Video Reasoning via Agentic Data Synthesis"
    },
    "infllm_v2_sparse_attention": {
      "description": "Dense-sparse switchable attention framework enabling efficient long-sequence processing while maintaining standard pretrain-finetune compatibility. Achieves 9.3x speedup with performance matching or exceeding full attention.",
      "reference": "InfLLM-V2: Dense-Sparse Switchable Attention for Seamless Short-to-Long Adaptation"
    },
    "deepsearch_mcts_rlvr": {
      "description": "Integrates Monte Carlo Tree Search into reinforcement learning training loop for LLMs. Addresses exploration bottleneck through systematic search and fine-grained credit assignment, achieving SOTA for 1.5B reasoning models.",
      "reference": "DeepSearch: Overcome RLVR Bottleneck via Monte Carlo Tree Search - UW, Berkeley"
    },
    "tooluniverse_ai_scientist": {
      "description": "Unified ecosystem with standardized AI-tool interaction protocol enabling dynamic discovery, creation, and composition of 600+ scientific tools. Facilitates end-to-end research workflows from hypothesis to therapeutic candidate identification.",
      "reference": "ToolUniverse: Democratizing AI scientists using ToolUniverse"
    },
    "thinking_sparks_attention_heads": {
      "description": "Circuit analysis demonstrating post-training for complex reasoning sparks emergence of novel, functionally specialized attention heads. Reveals different training regimes produce distinct head evolution patterns supporting structured computation.",
      "reference": "Thinking Sparks!: Emergent Attention Heads in Reasoning Models - Korea University, Upstage AI"
    },
    "parallax_decentralized_inference": {
      "description": "Decentralized LLM serving system turning heterogeneous GPU pools into efficient inference platform. Two-phase scheduler optimizes model allocation and request-time pipeline selection under bandwidth constraints.",
      "reference": "Parallax: Efficient LLM Inference Service over Decentralized Environment - NUS, HKUST"
    },
    "longlive_streaming_video": {
      "description": "Frame-level autoregressive framework for real-time interactive long video generation. KV-recache mechanism enables smooth prompt transitions while maintaining 20.7 FPS on single GPU for minute-long sequences.",
      "reference": "LongLive: Real-time Interactive Long Video Generation - NVIDIA, HKUST"
    },
    "entropy_after_think_early_exit": {
      "description": "Proposes entropy monitoring after stop-thinking token as signal for reasoning model early exit. Reduces token usage by 13-21% without harming accuracy through adaptive compute allocation.",
      "reference": "Entropy After </Think> for reasoning model early exiting - Netflix Research"
    },
    "ttt3r_test_time_3d": {
      "description": "Improves length generalization of RNN-based 3D reconstruction through confidence-guided state update rule at test time. Enables robust reconstruction over thousands of images with constant memory and real-time speed.",
      "reference": "TTT3R: 3D Reconstruction as Test-Time Training - Westlake University"
    },
    "dragon_hatchling_brain_bridge": {
      "description": "Bridges large language models and biologically plausible brain models through locally interacting neuron particles. GPU-optimized variant achieves transformer-comparable performance while demonstrating emergent modularity and inherent interpretability.",
      "reference": "The Dragon Hatchling: The Missing Link between Transformer and Brain Models - Pathway"
    },
    "polychromic_rl_objectives": {
      "description": "Set reinforcement learning framework with polychromic objectives mitigating entropy collapse during RL fine-tuning. Fosters diverse, robust policies demonstrating improved pass@k rates across BabyAI and algorithmic creativity environments.",
      "reference": "Polychromic Objectives for Reinforcement Learning - Stanford University"
    },
    "diffusion_memorization_mechanism": {
      "description": "Clarifies memorization in diffusion models stems from classifier-free guidance causing early training sample overestimation. This premature latent collapse forces denoising trajectories to memorized images rather than simple overfitting.",
      "reference": "How Diffusion Models Memorize - Yonsei University"
    },
    "nvfp4_low_precision": {
      "description": "Enhanced 4-bit floating point format and training methodology enabling stable pretraining of 12B parameter LLMs. Achieves 2-3x arithmetic speedup and 50% memory reduction with FP8-comparable performance.",
      "reference": "Pretraining Large Language Models with NVFP4 - NVIDIA"
    },
    "salt_frozen_teacher": {
      "description": "Two-stage video representation learning using frozen teacher from pixel reconstruction, then training student for masked latent prediction. Increases transparency and efficiency while outperforming V-JEPA under frozen evaluation.",
      "reference": "Rethinking JEPA: Compute-Efficient Video SSL with Frozen Teachers - Apple"
    },
    "gat_scalable_gans": {
      "description": "Framework leveraging latent space training and pure transformer architectures for scalable GANs. Achieves state-of-the-art FID of 2.96 on ImageNet-256 while training approximately six times faster than competitive methods.",
      "reference": "Scalable GANs with Transformers - Sungkyunkwan University"
    },
    "llava_onevision_open_training": {
      "description": "Family of large multimodal models achieving competitive SOTA performance against proprietary systems while trainable from scratch under $16,000. Provides open-source framework with model weights and curated datasets.",
      "reference": "LLaVA-OneVision-1.5: Fully Open Framework for Democratized Multimodal Training"
    },
    "regression_language_models": {
      "description": "Directly predicts numeric metrics from raw textual code and computational graph representations, eliminating manual feature engineering. Achieves >0.9 Spearman correlation on memory prediction, outperforming GNNs on architecture search.",
      "reference": "Regression Language Models for Code"
    },
    "ra3_action_abstractions": {
      "description": "Mid-training algorithm extracting temporal action abstractions from expert demonstrations to improve code generation. Accelerates post-training RLVR convergence while achieving 4-point accuracy improvement over next-token baselines.",
      "reference": "Learning to Reason as Action Abstractions with Scalable Mid-Training RL"
    },
    "visual_jigsaw_post_training": {
      "description": "Self-supervised RL-based post-training framework enhancing MLLMs' fine-grained visual understanding through jigsaw-style ordering tasks. Shows consistent performance gains across images, videos, and 3D data benchmarks.",
      "reference": "Visual Jigsaw Post-Training Improves MLLMs - Nanyang Technological University"
    },
    "spherear_hyperspherical_latents": {
      "description": "Introduces hyperspherical latents and token-level diffusion head to stabilize continuous-token autoregressive image generation. Achieves FID of 1.34 on ImageNet 256x256 with substantially fewer parameters than competitors.",
      "reference": "Hyperspherical Latents Improve Continuous-Token Autoregressive Generation - Peking University"
    },
    "univid_unified_video": {
      "description": "Unified architecture coupling MLLM with diffusion decoder through lightweight adapter for both video understanding and generation. Temperature modality alignment improves prompt adherence while pyramid reflection enables efficient temporal reasoning.",
      "reference": "UniVid: The Open-Source Unified Video Model - Peking University"
    },
    "sequential_diffusion_llm": {
      "description": "Framework unifying next-token and next-block prediction, enabling dynamic generation length while efficiently retrofitting pretrained autoregressive models. Achieves near-SFT accuracy with 2.1x inference acceleration.",
      "reference": "Sequential Diffusion Language Models - Shanghai AI Laboratory"
    },
    "evoworld_panoramic_generation": {
      "description": "World model bridging panoramic video generation with evolving 3D memory for spatially consistent long-horizon exploration. Evolving reconstruction provides explicit spatial guidance through geometric reprojections to target viewpoints.",
      "reference": "EvoWorld: Evolving Panoramic World Generation with Explicit 3D Memory - Johns Hopkins"
    },
    "da2_panoramic_depth": {
      "description": "Achieves zero-shot generalizable panoramic depth estimation by curating large-scale dataset from perspective data and introducing SphereViT for spherical distortions. Outperforms prior zero-shot and many in-domain techniques.",
      "reference": "DA²: Depth Anything in Any Direction - UC San Diego, HKUST"
    },
    "rolling_forcing_autoregressive": {
      "description": "Enables real-time, high-quality, long-duration video streaming by addressing error accumulation in autoregressive generation. Dramatically reduces temporal drift while maintaining visual consistency over multi-minute sequences.",
      "reference": "Rolling Forcing: Autoregressive Long Video Diffusion in Real Time"
    },
    "vggt_x_dense_nvs": {
      "description": "Memory-efficient 3D foundation model implementation scaling to 1,000+ images with adaptive global alignment for output enhancement. Achieves SOTA results in dense COLMAP-free novel view synthesis and pose estimation.",
      "reference": "VGGT-X: When VGGT Meets Dense Novel View Synthesis - Chinese Academy of Sciences"
    },
    "love_r1_adaptive_zoom": {
      "description": "Introduces adaptive zoom-in mechanism via multi-step reasoning for long video understanding. Dynamically focuses on relevant clips at higher resolution when needed, outperforming baseline by 3.1% average.",
      "reference": "LOVE-R1: Advancing Long Video Understanding with Adaptive Zoom-in - Alibaba Group"
    },
    "query_kontext_multimodal": {
      "description": "Bridges VLM and diffusion model via multimodal kontext composed of semantic cues and coarse-grained conditions. Delegates multimodal generative reasoning to VLM while reserving diffusion for high-quality synthesis.",
      "reference": "Query-Kontext: Unified Multimodal Model for Image Generation and Editing - NUS"
    },
    "kv_cache_compression_pitfalls": {
      "description": "Identifies pitfalls in KV cache compression including certain instructions degrading rapidly, system prompt leakage, and issues in multi-instruction scenarios. Proposes simple eviction policy changes improving multi-instruction performance.",
      "reference": "The Pitfalls of KV Cache Compression - UCLA"
    },
    "reseek_search_agents": {
      "description": "Self-correcting framework for search agents with instructive rewards. Introduces JUDGE action enabling dynamic identification and recovery from erroneous search paths through correctness and utility reward decomposition.",
      "reference": "ReSeek: A Self-Correcting Framework for Search Agents - Tsinghua, Tencent"
    },
    "alpharl_predictable_dynamics": {
      "description": "Discovers RL parameter updates in LLMs exhibit predictable low-dimensional patterns with dominant rank-1 subspace evolving linearly. Enables AlphaRL acceleration framework speeding training 2.5x while retaining 96%+ performance.",
      "reference": "On Predictability of Reinforcement Learning Dynamics for Large Language Models - USTC"
    },
    "dc_gen_compressed_latent": {
      "description": "Post-training framework adapting pretrained text-to-image diffusion models to deeply compressed latent spaces. Enables 4K generation with 56x throughput improvement and 53x latency reduction on H100 GPUs.",
      "reference": "DC-Gen: Post-Training Diffusion Acceleration with Deeply Compressed Latent Space"
    },
    "fast_dllm_v2_block_diffusion": {
      "description": "Efficient block-diffusion framework accelerating LLM inference up to 2.5x by adapting pretrained autoregressive models with only 1B fine-tuning tokens. Integrates hierarchical caching and confidence-aware parallel decoding.",
      "reference": "Fast-dLLM v2: Efficient Block-Diffusion LLM - NVIDIA, HKU, MIT"
    },
    "ar3po_adaptive_rollout": {
      "description": "Sampling efficient RLVR algorithm introducing adaptive rollout dynamically allocating responses to difficult prompts and response reuse leveraging previously correct answers. Reduces rollout cost up to 4.2x.",
      "reference": "Improving Sampling Efficiency in RLVR through Adaptive Rollout - UIUC, Amazon"
    },
    "piper_env_setup": {
      "description": "Enables small open-source LLMs to automate environment setup efficiently on consumer hardware using two-stage supervised fine-tuning and online RL. Achieves 9x improvement in success rate over base model.",
      "reference": "PIPer: On-Device Environment Setup via Online Reinforcement Learning"
    },
    "concrete_score_distillation": {
      "description": "Discrete score-matching objective for LLM knowledge distillation overcoming softmax smoothing and logit shift restrictions. Aligns relative logit differences across vocabulary pairs with flexible weighting for better fidelity-diversity tradeoffs.",
      "reference": "Distillation of Large Language Models via Concrete Score Matching"
    },
    "dyva_world_language_models": {
      "description": "Re-purposes video diffusion model as generative encoder performing single denoising step, treating resulting latents as visual embeddings. Significantly enhances spatial reasoning and enables single-image models for multi-frame reasoning.",
      "reference": "Can World Models Benefit VLMs for World Dynamics? - UCSB, HKUST"
    },
    "rlmt_thinking_chat": {
      "description": "Integrates explicit chain-of-thought reasoning into general-purpose chat models through reinforcement learning with model-rewarded thinking. Enables 8B models to rival much larger frontier LLMs on chat benchmarks.",
      "reference": "Language Models that Think, Chat Better - Princeton University"
    },
    "vapo_vision_anchored": {
      "description": "Addresses dual nature of multimodal reasoning where extended reasoning impairs perceptual grounding. Explicitly steers reasoning toward visually grounded trajectories, strengthening reliance on visual information for SOTA results.",
      "reference": "More Thought, Less Accuracy? On the Dual Nature of Reasoning in VLMs"
    },
    "mem_alpha_memory_construction": {
      "description": "Reinforcement learning framework training LLM agents to learn effective memory management strategies for complex external systems. Achieves 0.592 on MemoryAgentBench with robust generalization to 13x longer sequences.",
      "reference": "Mem-α: Learning Memory Construction via Reinforcement Learning - Stanford, UC San Diego"
    },
    "believing_without_seeing_quality": {
      "description": "Proposes visual fidelity and contrastiveness quality scores for VLM explanations to contextualize predictions for users without visual access. Showing scores improves accuracy at predicting correctness by 11.1%.",
      "reference": "Believing without Seeing: Quality Scores for VLM Explanations - CMU, USC"
    },
    "growing_winning_subnetworks": {
      "description": "Constructive sparse-to-dense training paradigm growing networks from sparse seed rather than pruning. Path weight magnitude product-biased random growth with step normalization automatically discovers operating density.",
      "reference": "Growing Winning Subnetworks, Not Pruning Them - Georgia Institute of Technology"
    },
    "img_implicit_guidance": {
      "description": "Re-generation-based multimodal alignment framework requiring no extra data or editing operations. Utilizes MLLM to identify misalignments, then manipulates diffusion conditioning features to reduce misalignments during re-generation.",
      "reference": "IMG: Calibrating Diffusion Models via Implicit Multimodal Guidance - Tsinghua, SHI Labs"
    },
    "preconditioning_feature_learning": {
      "description": "Shows preconditioning methods define input similarity geometry inducing controllable spectral bias profoundly impacting feature learning. Generalization improves when spectral bias aligns with underlying data structure and teacher signal.",
      "reference": "How Does Preconditioning Guide Feature Learning in DNNs? - Nanyang Technological University"
    },
    "crostata_cross_state_attention": {
      "description": "Proposes cross-state transition attention transformer employing novel state transition attention mechanism modulating weights based on learned evolution patterns. Combines with temporal masking encouraging reasoning from historical context.",
      "reference": "CroSTAta: Cross-State Transition Attention Transformer - Istituto Italiano di Tecnologia"
    },
    "mobilellm_r1_sub_billion": {
      "description": "Sub-billion parameter language models achieving strong reasoning capabilities outperforming prior open-sourced models using only 4.2T training tokens. Leverages open training recipes and advanced data curation for efficient on-device deployment.",
      "reference": "MobileLLM-R1: Exploring Limits of Sub-Billion Language Model Reasoners - Meta AI"
    },
    "rfg_reward_free_guidance": {
      "description": "Principled method for guiding diffusion LLM reasoning trajectory without explicit process reward. Parameterizes reward by log-likelihood ratios of enhanced and reference dLLMs, achieving 9.2% accuracy gains.",
      "reference": "RFG: Test-Time Scaling for Diffusion LLM Reasoning with Reward-Free Guidance - Stanford"
    },
    "vlm_fo1_fine_grained_perception": {
      "description": "Reframes object-centric perception from coordinate generation to feature retrieval task. Hybrid fine-grained region encoder with dual vision generates powerful region tokens enabling seamless reasoning about visual regions.",
      "reference": "VLM-FO1: Bridging Gap Between High-Level Reasoning and Fine-Grained Perception - Zhejiang University"
    },
    "icl_scaling_laws_theory": {
      "description": "Characterizes in-context learning of linear regression in deep linear self-attention model analyzing performance dependence on width, depth, training steps, batch size, and data. Provides solvable toy model of neural scaling laws.",
      "reference": "Theory of Scaling Laws for In-Context Regression - Harvard University"
    },
    "quasar_quantum_assembly": {
      "description": "Agentic reinforcement learning framework for quantum circuit generation and optimization using tool-augmented LLMs. Achieves 99.31% validity in Pass@1, outperforming GPT-4o, GPT-5 and DeepSeek-V3 on quantum assembly tasks.",
      "reference": "QUASAR: Quantum Assembly Code Generation Using Tool-Augmented LLMs - Southampton"
    },
    "rover_random_policy_valuation": {
      "description": "RLVR method leveraging uniform random policy's Q-values to improve LLM reasoning. Significantly enhances both quality and diversity by maintaining higher policy entropy while simplifying RL training process.",
      "reference": "Random Policy Valuation is Enough for LLM Reasoning - HKUST, Kuaishou"
    },
    "drivee2e_closed_loop": {
      "description": "Closed-loop benchmark for end-to-end autonomous driving leveraging real-world traffic scenarios captured via infrastructure sensors integrated into CARLA digital twins. Provides realistic and challenging evaluation environment.",
      "reference": "DriveE2E: Closed-Loop Benchmark through Real-to-Simulation - AIR, Tsinghua"
    },
    "airoa_moma_dataset": {
      "description": "Large-scale real-world multimodal dataset for mobile manipulation integrating synchronized dual-view RGB, proprioceptive, and 6-axis wrist force-torque signals. Hierarchical annotation schema for long-horizon tasks with explicit failure cases.",
      "reference": "The AIRoA MoMa Dataset: Large-Scale Hierarchical Dataset for Mobile Manipulation"
    },
    "dvla_diffusion_vla": {
      "description": "Diffusion-based VLA leveraging multimodal chain-of-thought to unify visual perception, language reasoning, and robotic control. Jointly optimizes all modalities under single diffusion objective for stronger cross-modal reasoning.",
      "reference": "dVLA: Diffusion Vision-Language-Action Model with Multimodal CoT - Peking, Shanghai Jiaotong"
    },
    "hamlet_history_aware_vla": {
      "description": "Scalable framework adapting VLAs to attend to historical context during action prediction using moment tokens compactly encoding perceptual information. Lightweight memory module integrates tokens across timesteps for long-horizon tasks.",
      "reference": "HAMLET: Switch VLA Model into History-Aware Policy - KAIST, UC Berkeley"
    },
    "longcodezip_context_compression": {
      "description": "Plug-and-play code compression framework for code LLMs employing dual-stage strategy identifying function-level chunks then selecting optimal blocks. Achieves 5.6x compression ratio without degrading task performance.",
      "reference": "LongCodeZip: Compress Long Context for Code Language Models - Shanghai Jiao Tong, Stanford"
    },
    "normal_abnormal_gad": {
      "description": "Generalist anomaly detection leveraging both normal and anomalous samples as references. Residual mining extracts abnormal patterns while anomaly feature learning adaptively identifies instance-aware anomalies through residual mapping.",
      "reference": "Normal-Abnormal Guided Generalist Anomaly Detection - Xi'an Jiaotong-Liverpool, Liverpool"
    },
    "ttt_in_context_learning": {
      "description": "Analyzes test-time training combination with in-context learning for single-index models. Proves single-layer transformers with TTT adapt to both feature vectors and link functions, unlike ICL alone.",
      "reference": "Test time training enhances in-context learning of nonlinear functions - University of Tokyo"
    },
    "point_it_out_grounding": {
      "description": "Benchmark systematically assessing embodied reasoning abilities of VLMs through precise visual grounding. Hierarchical evaluation protocol spanning referred-object localization, task-driven pointing, and visual trace prediction.",
      "reference": "Point-It-Out: Benchmarking Embodied Reasoning for VLMs in Multi-Stage Grounding"
    },
    "synchrony_stability_chatbots": {
      "description": "Evaluation framework making explicit tension between moment-to-moment linguistic synchrony and long-term persona stability in adaptive chatbots. Maps Pareto frontier where bounded policies achieve stability gains with modest synchrony costs.",
      "reference": "Navigating the Synchrony-Stability Frontier in Adaptive Chatbots - University of Minnesota"
    },
    "cardiobench_echo_foundation": {
      "description": "Comprehensive benchmark for echocardiography foundation models unifying eight public datasets into standardized suite spanning regression and classification tasks. Evaluates cardiac-specific, biomedical, and general-purpose encoders under consistent protocols.",
      "reference": "CardioBench: Do Echocardiography Foundation Models Generalize Beyond Lab? - MBZUAI"
    },
    "sana_video_block_linear": {
      "description": "Block linear diffusion transformer for efficient high-resolution minute-long video generation drastically reducing training costs and accelerating inference. Achieves 16x faster inference than prior art with robust on-device deployment.",
      "reference": "SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer"
    },
    "fastforward_camera_localization": {
      "description": "Creates map representation and relocalizes query image on-the-fly in single feed-forward pass. Represents multiple mapping images as collection of features anchored in 3D, predicting correspondences for camera pose estimation.",
      "reference": "FastForward: Feed-Forward Camera Localization from Image Features - Oxford, Niantic"
    },
    "softcfg_autoregressive_guidance": {
      "description": "Uncertainty-guided inference method distributing adaptive perturbations across all tokens in autoregressive sequence. Each token contributes certainty-weighted guidance ensuring signal persists while resolving text-visual conflicts.",
      "reference": "SoftCFG: Uncertainty-guided Stable Guidance for Visual Autoregressive Model - KU Leuven"
    },
    "infoagent_deep_research": {
      "description": "Deep research agent with innovative data synthesis pipeline and orchestrated web search tools. Entity tree sub-tree sampling with fuzzification systematically increases question difficulty for challenging information-seeking tasks.",
      "reference": "InfoAgent: Advancing Autonomous Information-Seeking Agents"
    },
    "oceangym_underwater_agents": {
      "description": "First comprehensive benchmark for ocean underwater embodied agents spanning eight realistic task domains. MLLM-driven unified framework integrating perception, memory, and sequential decision-making under extreme perceptual challenges.",
      "reference": "OceanGym: Benchmark Environment for Underwater Embodied Agents - NUS, Zhejiang"
    },
    "deepsketcher_visual_manipulation": {
      "description": "Comprehensive suite with image-text interleaved dataset and self-contained model for visual manipulation reasoning. Dataset contains 31k CoT trajectories with diverse tool calls and edited images across data types.",
      "reference": "DeepSketcher: Internalizing Visual Manipulation for Multimodal Reasoning - Wuhan University"
    },
    "hearing_order_audio_llm": {
      "description": "First systematic investigation of selection bias in large audio-language models showing no model is immune. Shuffling answer order causes up to 24% performance fluctuations, raising concerns about evaluation reliability.",
      "reference": "Hearing the Order: Investigating Selection Bias in Large Audio-Language Models - National Taiwan"
    },
    "cogito_ergo_ludo_reasoning": {
      "description": "Agent architecture leveraging LLM to build explicit language-based understanding of environment mechanics and strategy. Starting tabula rasa, operates on cycle of interaction and reflection performing concurrent rule induction and strategy summarization.",
      "reference": "Cogito, Ergo Ludo: Agent that Learns to Play by Reasoning and Planning"
    },
    "diffusion_bridge_flow_matching": {
      "description": "Unified theoretical and experimental validation of diffusion bridge and flow matching through stochastic optimal control lens. Proves diffusion bridge cost function is lower, guiding system toward more stable trajectories.",
      "reference": "Diffusion Bridge or Flow Matching? Unifying Framework and Analysis - ShanghaiTech"
    },
    "editscore_rl_reward": {
      "description": "State-of-the-art specialized reward models for evaluating instruction-guided image editing quality. Introduces EditReward-Bench benchmark and EditScore series enabling efficient robust policy optimization for image editing.",
      "reference": "EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward - USTC, Zhejiang"
    },
    "up2you_unconstrained_3d": {
      "description": "First tuning-free solution reconstructing high-fidelity 3D clothed portraits from extremely unconstrained in-the-wild photos. Data rectifier paradigm converts unconstrained inputs into clean orthogonal multi-view images in seconds.",
      "reference": "UP2You: Fast Reconstruction from Unconstrained Photo Collections - Nanjing, Westlake"
    },
    "physicsminions_olympiad": {
      "description": "Coevolutionary multimodal multi-agent framework designed to solve complex physics Olympiad problems. Elevated open-source models to gold medals and enabled closed-source model to achieve gold in all seven evaluated Olympiads.",
      "reference": "PhysicsMinions: Winning Gold Medals in Latest Physics Olympiads - Shanghai AI Lab"
    },
    "visualoverload_dense_scenes": {
      "description": "VQA benchmark with 2,720 question-answer pairs challenging models to perform simple knowledge-free vision tasks in densely populated scenes. High-resolution scans of public-domain paintings with manually annotated questions across six categories.",
      "reference": "VisualOverload: Probing Visual Understanding of VLMs in Really Dense Scenes"
    },
    "spinbench_spatial_reasoning": {
      "description": "Cognitively grounded diagnostic benchmark evaluating spatial reasoning in VLMs designed around perspective taking. Progressive structure targets translation, rotation, relative pose, and viewpoint change revealing systematic weaknesses.",
      "reference": "SpinBench: Perspective and Rotation as Lens on Spatial Reasoning in VLMs"
    },
    "dolphin_ultrasound_foundation": {
      "description": "First large-scale multimodal ultrasound foundation models unifying diverse clinical tasks in single vision-language framework. Curated 2-million-scale dataset combining textbook knowledge, public data, and synthetic samples for robust perception.",
      "reference": "Dolphin v1.0 Technical Report"
    },
    "llm_fingerprint_robustness": {
      "description": "Critical look at model fingerprinting schemes identifying fundamental vulnerabilities to adversarial attacks. Demonstrates adaptive attacks can bypass authentication completely for ten recent schemes while maintaining high utility.",
      "reference": "Are Robust LLM Fingerprints Adversarially Robust?"
    },
    "generalized_info_bottleneck": {
      "description": "Reformulates information bottleneck principle through synergy lens using average interaction information. Demonstrates synergistic functions achieve superior generalization and provides computable framework addressing original IB limitations.",
      "reference": "A Generalized Information Bottleneck Theory of Deep Learning - University College London"
    },
    "bayesian_influence_functions": {
      "description": "Extension of classical influence functions replacing Hessian inversion with loss landscape statistics estimated via stochastic-gradient MCMC sampling. Hessian-free approach captures higher-order interactions and scales to billions of parameters.",
      "reference": "Bayesian Influence Functions for Hessian-Free Data Attribution - Melbourne, TU Munich"
    },
    "rl_guided_data_selection": {
      "description": "Reformulates data selection for finetuning LLMs as tractable MDP, training RL agents to learn optimal policies guided by proxy-model reward. Training on 5% selected subset matches full dataset performance while cutting training time 2x.",
      "reference": "RL-Guided Data Selection for Language Model Finetuning - Stanford University"
    },
    "clip_entropy_dynamics": {
      "description": "Reveals clipping mechanism in PPO and GRPO induces biases on entropy. Clip-low increases entropy while clip-high decreases it, with clip-high dominating under standard parameters resulting in overall entropy reduction.",
      "reference": "Clip-Low Increases Entropy and Clip-High Decreases Entropy in RL of LLMs - UCLA, Stanford"
    },
    "code2video_educational": {
      "description": "Code-centric agent framework for generating educational videos via executable Python code. Three collaborative agents structure content, convert to code with auto-fix, and refine spatial layout using vision-language models.",
      "reference": "Code2Video: Code-centric Paradigm for Educational Video Generation - NUS"
    },
    "instant4d_casual_video": {
      "description": "Monocular reconstruction system leveraging native 4D representation to efficiently process casual video sequences within minutes. Grid pruning reduces model size to under 10% while maintaining geometric integrity.",
      "reference": "Instant4D: 4D Gaussian Splatting in Minutes - Pittsburgh, CMU, Sichuan"
    },
    "socratic_zero_coevolution": {
      "description": "Fully autonomous framework generating high-quality training data from minimal seed examples through co-evolution of Teacher, Solver, and Generator agents. Solver refines reasoning, Teacher crafts challenging questions, Generator distills question-design strategy.",
      "reference": "Socratic-Zero: Bootstrapping Reasoning via Data-Free Agent Co-evolution"
    },
    "cj_grpo_consistency_trajectory": {
      "description": "Introduces novel decoding strategies and specialized RL algorithm for masked diffusion language models. Resolves trajectory inconsistencies enabling efficient training while reducing decoding steps from O(L) to O(log_2 L).",
      "reference": "Taming Masked Diffusion via Consistency Trajectory RL - Shanghai AI Lab, Fudan"
    },
    "world_env_vla_simulator": {
      "description": "RL-based post-training framework replacing physical interaction with low-cost world model-based virtual simulator. Video-based simulator generates temporally consistent observations while VLM-guided reflector provides continuous rewards and termination prediction.",
      "reference": "World-Env: Leveraging World Model as Virtual Environment for VLA - Alibaba, Sun Yat-Sen"
    },
    "agent_matrix_self_replication": {
      "description": "Comprehensive evaluation framework for quantifying self-replication risks in LLM agents driven by objective misalignment. Authentic production environments and realistic tasks enable scenario-driven assessment of spontaneous replication behaviors.",
      "reference": "Dive into the Agent Matrix: Realistic Evaluation of Self-Replication Risk - Shanghai AI Lab"
    },
    "agent_failure_analysis": {
      "description": "Introduces unified taxonomy and benchmark for LLM agent failure analysis alongside iterative debugging framework. Enables agents to identify and recover from root-cause errors, achieving 26% relative improvements in task success.",
      "reference": "Where LLM Agents Fail and How They can Learn From Failures"
    },
    "deepfake_gaze_detection": {
      "description": "First reported method utilizing point-of-gaze tracking for deepfake detection in real-time video calls. Leverages screen content and estimated gaze to detect deepfakes incapable of mimicking subtle nonverbal communication.",
      "reference": "DeepFake Detection in Dyadic Video Calls using Point of Gaze Tracking"
    },
    "triangle_splatting_plus": {
      "description": "Generates high-quality opaque semi-connected triangle meshes directly from 2D images. Output is immediately compatible with traditional graphics engines, eliminating post-processing and enabling real-time interactive applications.",
      "reference": "Triangle Splatting+: Differentiable Rendering with Opaque Triangles - Toronto, UBC"
    },
    "noisy_label_detection": {
      "description": "Similarity-based approach detecting and rectifying noisy labels in deep learning datasets by leveraging penultimate layer feature representations. Demonstrates robust error detection accuracy across noise types improving model generalization.",
      "reference": "Detecting and Rectifying Noisy Labels: A Similarity-based Approach - RIKEN, JAIST"
    },
    "calibrating_autoraters": {
      "description": "General framework for calibrating probabilistic autoraters to preference distributions. Formalizes problem and presents supervised fine-tuning for dense labels and RL approach for sparse labels achieving better-calibrated predictions.",
      "reference": "Judging with Confidence: Calibrating Autoraters to Preference Distributions - Google DeepMind"
    },
    "authentic_discrete_diffusion": {
      "description": "Redefines discrete diffusion by preserving core characteristics directly in one-hot space through coordinated mechanisms. Timestep-conditioned cross-entropy loss establishes bridge between discriminative and generative learning.",
      "reference": "Authentic Discrete Diffusion Model - Sun Yat-Sen University"
    },
    "cmt_consistency_midtraining": {
      "description": "First concept and practical method inserting lightweight intermediate stage between diffusion pre-training and flow map post-training. Trains trajectory-consistent initialization enabling fast robust convergence using 98% less data.",
      "reference": "CMT: Mid-Training for Efficient Learning of Consistency, Mean Flow, Flow Map Models"
    },
    "align_visual_encoders": {
      "description": "Three-stage progressive alignment method integrating pretrained visual foundation encoders into continuous visual tokenizers for diffusion models. Leads to models training 5x faster with superior generation quality.",
      "reference": "Aligning Visual Foundation Encoders to Tokenizers for Diffusion Models - UW, MIT, Adobe"
    },
    "learn2pd_adaptive_decoding": {
      "description": "Introduces Learn2PD and EoTP for adaptive parallel decoding and end-of-text prediction to accelerate diffusion LLMs. Achieves 22.58x speedup and 57.51x when combined with KV-Cache.",
      "reference": "Learning to Parallel: Accelerating Diffusion LLMs via Adaptive Parallel Decoding - UCF, HKUST"
    },
    "tr2_d2_tree_search_diffusion": {
      "description": "Framework integrating Monte Carlo Tree Search into discrete diffusion fine-tuning using stochastic optimal control. Generates replay buffers through MCTS subsequently used for trajectory-aware fine-tuning.",
      "reference": "TR2-D2: Tree Search Guided Trajectory-Aware Fine-Tuning - Georgia Tech, UPenn"
    },
    "imageedit_video_subject": {
      "description": "Training-free framework for any number of video subject editing manipulating appearances while preserving non-target regions. Provides robust multimodal conditioning and precise masks through prompt-guided alignment and prior-based retargeting.",
      "reference": "IMAGEdit: Let Any Subject Transform - NUS, Nanjing University"
    },
    "critpt_physics_reasoning": {
      "description": "Benchmark comprising 71 composite research challenges and 190 checkpoint tasks crafted by 50+ physicists evaluating LLMs' reasoning on frontier physics research. Current LLMs achieve only 4% accuracy demonstrating significant limitations.",
      "reference": "Probing the Critical Point of AI Reasoning: Frontier Physics Research Benchmark - Toronto, Caltech"
    },
    "rethinking_reward_models": {
      "description": "First unified evaluation of discriminative and generative ORM and PRM across 14 diverse domains. Finds generative ORM is most robust, challenging assumption that fine-grained supervision is always better.",
      "reference": "Rethinking Reward Models for Multi-Domain Test-Time Scaling - KAIST, Microsoft"
    },
    "fusion_of_n": {
      "description": "Method using general LLM judge to synthesize most informative elements of each sample from diverse pool into single final answer. Consistently outperforms Best-of-N showing versatility in test-time scaling and synthetic data generation.",
      "reference": "Making, not Taking, the Best of N - Cohere"
    },
    "molm_lora_markers": {
      "description": "General watermarking framework formulating encoding as key-dependent perturbation of generative model parameters. Routing-based instantiation with LoRA adapters achieves imperceptibility, fidelity, verifiability, and robustness against attacks.",
      "reference": "MOLM: Mixture of LoRA Markers - Michigan State, MBZUAI"
    },
    "jepa_t_joint_embedding": {
      "description": "Unified multimodal framework encoding images and captions into discrete tokens processed by joint-embedding predictive transformer. Late architectural fusion with objective-level alignment effectively balances conditioning strength with backbone generality.",
      "reference": "JEPA-T: Joint-Embedding Predictive Architecture with Text Fusion - USC, UC San Diego"
    },
    "visual_self_refinement": {
      "description": "Plug-and-play refinement module enhancing complex spatial correspondence modeling within generated visual sequence. Operates as post-pretraining step jointly refining all tokens, leveraging global context to mitigate error accumulation.",
      "reference": "Visual Self-Refinement for Autoregressive Models - Rochester Institute of Technology"
    }
  }
}