d6fc7c1dfa
- Work type = area x role (fixed vs varied), vocabulary in scripts/worktypes.json - LLM-read labels for 2,678 agent-related papers (data/labels), rule labels for the rest - Per-type synthesis notes (why-same, sub-variants, benchmarks, workflow) - Pages: /work-types matrix and detail (timeline, benchmarks, workflow, score charts), /benchmarks - build_worktypes.py adds work-type tables to data/papers.db Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bhs9LU8eMixk6g1kb1d3j7
61 lines
29 KiB
JSON
61 lines
29 KiB
JSON
{"id": 2583, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Retrieves cell complexes capturing cycles for textual graph QA RAG"}], "pattern": "vary-scaffold", "fixes": ["LLM", "graph QA tasks"], "varies": ["retrieval structure"], "benchmarks": [], "results": [], "workflow": ["Build cell complex", "Retrieve cycles", "Augment LLM generation"], "summary": "TopoRAG augments RAG on textual graphs with cell-complex topology so cycles are retrieved for reasoning."}
|
|
{"id": 13909, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Calibrated teacher token rewards for robust RL policy optimization"}], "pattern": "vary-data", "fixes": ["RL recipe", "student model"], "varies": ["token reward calibration"], "benchmarks": [], "results": [], "workflow": ["Get teacher token signals", "Calibrate teacher influence", "Optimize policy"], "summary": "TRAC modulates noisy teacher token-level rewards with adaptive calibration to improve RL for LLM reasoning."}
|
|
{"id": 13179, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Agent society framework sustaining coherent long-horizon narrative simulation"}, {"area": "multi-agent", "role": "scaffold", "why": "Stratified narrative memory for LLM multi-agent societies"}], "pattern": "fix-env-vary-agent", "fixes": ["LLM"], "varies": ["memory and spatial structure"], "benchmarks": [], "results": [], "workflow": ["Stratified narrative memory", "Spatial-plot alignment", "Simulate society"], "summary": "EvoSpark sustains coherent long-horizon narratives in LLM agent societies via structured memory and spatial consistency."}
|
|
{"id": 6291, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL with evolving rubrics to train long-form deep research agent"}], "pattern": "fix-env-vary-agent", "fixes": ["long-form research tasks"], "varies": ["reward rubrics", "model weights"], "benchmarks": [], "results": [], "workflow": ["Construct rubrics", "Co-evolve with policy", "Train with RL"], "summary": "DR Tulu trains an open deep research agent with RL using rubrics that co-evolve with the policy."}
|
|
{"id": 13742, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "Multilingual chart QA benchmark across 10 languages"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated LVLMs"], "varies": ["languages", "charts"], "benchmarks": [{"name": "PolyChartQA", "purpose": "multilingual chart QA", "is_new": true}], "results": [], "workflow": ["Translate chart data", "Reuse code to render", "Quality control"], "summary": "PolyChartQA is a large multilingual chart question answering benchmark for evaluating vision-language models."}
|
|
{"id": 12151, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Synthesize multi-turn tool-use trajectories from text corpora"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": ["Relevance filter", "Extract workflows", "Generate trajectories"], "summary": "GEM synthesizes multi-turn tool-use training trajectories from text corpora."}
|
|
{"id": 11961, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free decoding using vision token semantics to reduce hallucination"}], "pattern": "vary-scaffold", "fixes": ["LVLM weights"], "varies": ["decoding method"], "benchmarks": [], "results": [], "workflow": ["Analyze vision tokens", "Reference vision tokens", "Guide decoding"], "summary": "ReVisiT is a training-free decoding method that references vision-token semantics to guide LVLM generation."}
|
|
{"id": 8358, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Two-stage GRPO jointly improving high-level planning and CoT reasoning"}], "pattern": "vary-data", "fixes": ["benchmarks"], "varies": ["RL training with plans"], "benchmarks": [], "results": [], "workflow": ["Generate high-level plan", "Train CoT with plan", "GRPO optimization"], "summary": "PTA-GRPO adds plan guidance to GRPO training for improved LLM reasoning."}
|
|
{"id": 10535, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Label-free 3D segmentation using foundation-model rewards"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["reward modules"], "benchmarks": [], "results": [], "workflow": ["Superpoint merging agent", "Semantic and geometric rewards"], "summary": "FoundObj does label-free 3D object segmentation using foundation-model-based rewards."}
|
|
{"id": 11057, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Gradient projection to stabilize long-horizon test-time adaptation"}], "pattern": "n/a", "fixes": ["pretrained models"], "varies": ["adaptation update rule"], "benchmarks": [], "results": [], "workflow": ["Find reliable subspace", "Project gradients"], "summary": "Manifold-aware gradient projection stabilizes long-horizon test-time adaptation."}
|
|
{"id": 16935, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "method", "why": "Discourse graph guided document translation with LLMs"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["discourse graph conditioning"], "benchmarks": [], "results": [], "workflow": ["Build discourse graph", "Condition on neighbors", "Translate chunks"], "summary": "TransGraph uses discourse graphs to condition document translation on relevant segments."}
|
|
{"id": 10636, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Hierarchical multi-agent planning with coordinator and parallel day executors"}, {"area": "multi-agent", "role": "train-agent", "why": "Single policy trained for coordinator and executors"}], "pattern": "fix-env-vary-agent", "fixes": ["travel planning tasks"], "varies": ["agent hierarchy", "policy training"], "benchmarks": [], "results": [], "workflow": ["Coordinator allocates", "Parallel day planning", "Monitor constraints"], "summary": "HiMAP-Travel hierarchically plans constrained long-horizon travel with parallel agents and a trained policy."}
|
|
{"id": 5759, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "MCP benchmark with 300+ interdependent tools in stateful sandboxes"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated agents"], "varies": ["tool sandboxes", "API failures"], "benchmarks": [{"name": "ComplexMCP", "purpose": "evaluate agents on complex tools", "is_new": true}], "results": [], "workflow": ["Build sandboxes", "Seed dynamic states", "Evaluate agents"], "summary": "ComplexMCP benchmarks agents on interdependent MCP tools in dynamic stateful sandboxes with failures."}
|
|
{"id": 11243, "in_agent_scope": false, "types": [{"area": "train/distill", "role": "method", "why": "Distill dense teacher into MoE student with routing learning"}], "pattern": "n/a", "fixes": ["teacher"], "varies": ["distillation method"], "benchmarks": [], "results": [], "workflow": ["Cluster teacher neurons", "Warm up experts", "Online distillation"], "summary": "PADD distills dense teachers into MoE students while learning routing."}
|
|
{"id": 11651, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Pretraining-based self-correction for discrete diffusion LMs"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["diffusion objective"], "benchmarks": [], "results": [], "workflow": ["Self-correcting objective", "Pretrain diffusion model"], "summary": "Proposes a self-correcting pretraining objective for discrete diffusion language models."}
|
|
{"id": 12272, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Behavioral watermarking of agent planning for provenance"}], "pattern": "vary-scaffold", "fixes": ["agent utility"], "varies": ["watermark mechanism"], "benchmarks": [], "results": [], "workflow": ["Embed behavior watermark", "Preserve utility", "Detect"], "summary": "AgentMark watermarks agent planning behaviors for provenance while preserving utility."}
|
|
{"id": 11062, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Agentic system for extreme weather diagnosis with hypothesis-verification-replanning"}], "pattern": "vary-scaffold", "fixes": ["general LLM"], "varies": ["diagnosis pipeline"], "benchmarks": [], "results": [], "workflow": ["Hypothesize", "Verify with tools", "Replan"], "summary": "HVR-Met is an agentic system diagnosing extreme weather via hypothesis, verification and replanning."}
|
|
{"id": 17850, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Irish minimal-pair linguistic benchmark with human baseline"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated LLMs"], "varies": ["linguistic features"], "benchmarks": [{"name": "Irish-BLiMP", "purpose": "Irish syntax minimal pairs", "is_new": true}], "results": [], "workflow": ["Construct 1020 minimal pairs", "Evaluate LLMs and humans"], "summary": "Irish-BLiMP is a minimal-pair benchmark of Irish syntax comparing humans and LLMs."}
|
|
{"id": 10420, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "theory", "why": "Proves recursive decomposition reduces needed context for long-horizon reasoning"}, {"area": "planning-decision", "role": "method", "why": "Recursive models invoking themselves on subtasks"}], "pattern": "vary-scaffold", "fixes": ["bounded context"], "varies": ["recursive invocation"], "benchmarks": [], "results": [], "workflow": ["Decompose task", "Recurse in isolated context"], "summary": "Recursive models invoke themselves on subtasks, with proofs of context reduction for long-horizon reasoning."}
|
|
{"id": 6233, "in_agent_scope": true, "types": [{"area": "science-data", "role": "train-agent", "why": "Train forecasting agent with bootstrapped exploration and causal reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["forecasting datasets"], "varies": ["training paradigm"], "benchmarks": [], "results": [], "workflow": ["Bootstrap exploration", "Causal reasoning", "Accumulate strategy knowledge"], "summary": "A training paradigm for adaptive time-series forecasting agents using bootstrapped exploration and causal reasoning."}
|
|
{"id": 2384, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent quantitative trading system decoupling strategy development and execution"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent architecture"], "benchmarks": [], "results": [], "workflow": ["Develop strategy", "Decouple deployment", "Execute trades"], "summary": "TiMi is a rationality-driven multi-agent system for quantitative trading."}
|
|
{"id": 11038, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Training-free self-regulating Darwinian memory for GUI agents"}, {"area": "web-gui", "role": "scaffold", "why": "Memory for long-horizon cross-app GUI tasks"}], "pattern": "vary-scaffold", "fixes": ["MLLM weights"], "varies": ["memory system"], "benchmarks": [], "results": [], "workflow": ["Build memory ecosystem", "Survival-of-fittest pruning", "Retrieve for tasks"], "summary": "Darwinian Memory System evolves GUI agent memory training-free by selecting useful experiences."}
|
|
{"id": 161, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "RL-trained memory agent with revisitable callback memory for long context"}], "pattern": "fix-env-vary-agent", "fixes": ["long-context QA"], "varies": ["memory mechanism", "RL training"], "benchmarks": [], "results": [], "workflow": ["Scan document", "Store memory", "Callback retrieval"], "summary": "ReMemR1 is a memory agent with revisitable memory trained by RL for long-context QA."}
|
|
{"id": 14252, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Deployed multi-agent pipeline for terminology-constrained translation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["pipeline configurations"], "benchmarks": [{"name": "WMT25 Terminology Translation", "purpose": "ablate configurations", "is_new": false}], "results": [], "workflow": ["Pre-analysis", "Translate", "Post-edit", "Review"], "summary": "AIDAterm is a multi-agent LLM pipeline for terminology-constrained translation, evaluated via ablations."}
|
|
{"id": 18487, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Voice-interactive LLM system for dementia history-taking interviews"}], "pattern": "vary-scaffold", "fixes": ["system"], "varies": ["prompting strategy"], "benchmarks": [], "results": [], "workflow": ["Scripted branching", "Voice interaction", "User study"], "summary": "A voice LLM interviewer for dementia pre-visit history-taking, compared across two prompting strategies in a user study."}
|
|
{"id": 15632, "in_agent_scope": true, "types": [{"area": "search-research", "role": "analysis", "why": "Analyzes reranking vs reasoning budget tradeoffs in deep search"}, {"area": "search-research", "role": "eval-method", "why": "Effective token cost metric"}], "pattern": "vary-conditions", "fixes": ["BrowseComp-Plus tasks"], "varies": ["model scale", "reasoning effort", "reranking depth"], "benchmarks": [{"name": "BrowseComp-Plus", "purpose": "analyze tradeoffs", "is_new": false}], "results": [], "workflow": ["Vary reranking", "Vary reasoning", "Compute ETC"], "summary": "Analyzes reranking depth versus reasoning effort in deep search agents using a new effective token cost metric."}
|
|
{"id": 6168, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Benchmark of agent compliance with copyright law in commercial tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated agents"], "varies": ["tasks"], "benchmarks": [{"name": "Copyright-Bench", "purpose": "copyright compliance of agents", "is_new": true}], "results": [], "workflow": ["Build commercial tasks", "Public vs copyrighted content"], "summary": "Copyright-Bench evaluates whether LLM agents choose lawful content in commercial tasks."}
|
|
{"id": 14858, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Dependency-aware graph retrieval for efficient consistent agentic editing"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval"], "benchmarks": [], "results": [], "workflow": ["Build dependency graph", "Retrieve context", "Edit"], "summary": "LEDGER uses dependency-aware graph retrieval to scale agentic document editing."}
|
|
{"id": 174, "in_agent_scope": true, "types": [{"area": "science-data", "role": "train-agent", "why": "Curriculum RL with synthetic data for geometry agent InternGeometry"}, {"area": "science-data", "role": "data", "why": "Synthetic geometry problems with complexity curriculum"}], "pattern": "fix-env-vary-agent", "fixes": ["geometry problems"], "varies": ["training curriculum", "data"], "benchmarks": [], "results": [], "workflow": ["Propose constructions", "Verify", "Curriculum RL"], "summary": "InternGeometry is an LLM geometry agent trained by complexity-curriculum RL on synthetic data."}
|
|
{"id": 9380, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RL-trained small orchestrator coordinating models and tools"}], "pattern": "fix-env-vary-agent", "fixes": ["tools"], "varies": ["orchestrator training reward"], "benchmarks": [{"name": "Humanity's Last Exam", "purpose": "agentic hard problem solving", "is_new": false}], "results": [], "workflow": ["Orchestrator selects tools", "RL with efficiency rewards"], "summary": "ToolOrchestra trains an 8B orchestrator with RL to coordinate models and tools efficiently."}
|
|
{"id": 7113, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Multi-agent RL point prompt optimization for SAM few-shot segmentation"}], "pattern": "n/a", "fixes": ["DINOv2", "SAM"], "varies": ["prompt construction"], "benchmarks": [], "results": [], "workflow": ["Hierarchical MARL", "Optimize point prompts"], "summary": "PromptPilot uses game-theoretic multi-agent RL to optimize point prompts for SAM."}
|
|
{"id": 14021, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent framework for ICD coding on Chinese EHRs"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["multi-agent workflow"], "benchmarks": [], "results": [], "workflow": ["Disease coding", "Procedure coding", "Multi-stage workflow"], "summary": "JointCoder is a multi-agent framework for ICD coding on Chinese clinical records."}
|
|
{"id": 13182, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "analysis", "why": "Hierarchical planning analysis of web agent failures"}], "pattern": "vary-conditions", "fixes": ["web agents"], "varies": ["plan formats", "layers"], "benchmarks": [], "results": [], "workflow": ["Analyze planning", "Analyze execution", "Analyze replanning"], "summary": "Analyzes why web agents fail across planning, execution and replanning layers, comparing PDDL plans."}
|
|
{"id": 11763, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Retrieval-augmented multi-agent framework for sarcasm detection"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["multi-agent reasoning"], "benchmarks": [], "results": [], "workflow": ["Route reasoning strategy", "Retrieve knowledge", "Aggregate"], "summary": "RAM-SD uses retrieval-augmented multi-agent reasoning for sarcasm detection."}
|
|
{"id": 11702, "in_agent_scope": true, "types": [{"area": "reason/math", "role": "benchmark", "why": "Hard Mode reannotated MiniF2F-Hard and FIMO-Hard"}, {"area": "reason/math", "role": "scaffold", "why": "Agentic framework for Lean 4 hard-mode theorem proving"}], "pattern": "vary-both", "fixes": ["ATP models"], "varies": ["benchmark mode", "framework"], "benchmarks": [{"name": "MiniF2F-Hard", "purpose": "answer-discovery ATP", "is_new": true}, {"name": "FIMO-Hard", "purpose": "answer-discovery ATP", "is_new": true}], "results": [], "workflow": ["Discover answer", "Construct formal proof"], "summary": "Releases Hard Mode ATP benchmarks and the DAP agentic framework for Lean 4."}
|
|
{"id": 18649, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "method", "why": "Guideline-centered annotation methodology"}], "pattern": "vary-evaluation", "fixes": ["tasks"], "varies": ["annotation protocol"], "benchmarks": [], "results": [], "workflow": ["Record guidelines per instance"], "summary": "GCAM annotates data with associated guidelines to reduce information loss and enable reuse."}
|
|
{"id": 12035, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Diffusion LM with sufficiency-guided remasking for argument summarization"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["generation method"], "benchmarks": [], "results": [], "workflow": ["Mask", "Regenerate"], "summary": "Arg-LLaDA iteratively refines argument summaries via sufficiency-guided remasking."}
|
|
{"id": 8270, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Self-correcting search agent trained with instructive rewards"}], "pattern": "fix-env-vary-agent", "fixes": ["QA tasks"], "varies": ["reward", "training"], "benchmarks": [], "results": [], "workflow": ["Detect errors", "Recover", "RL"], "summary": "ReSeek trains search agents to self-correct erroneous paths using instructive rewards."}
|
|
{"id": 1605, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Research code development benchmark with multi-turn human feedback"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "RECODE", "purpose": "multi-turn research coding", "is_new": true}], "results": [], "workflow": ["102 tasks", "Feedback hierarchy"], "summary": "RECODE benchmarks LLMs on research code with interactive human feedback."}
|
|
{"id": 9955, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Generalizable 3D shape deformation learning"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["Template deformation"], "summary": "Template deformation framework with geometry-guided features for monocular 3D recovery."}
|
|
{"id": 10454, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Agentic graph reasoning with neighbor retrieval and code interpreter"}, {"area": "tool-use", "role": "train-agent", "why": "Staged training"}], "pattern": "fix-env-vary-agent", "fixes": ["graph tasks"], "varies": ["agentic probing"], "benchmarks": [], "results": [], "workflow": ["Probe neighbors", "Compute with code"], "summary": "GRASP lets LLMs explore graphs and compute with a code interpreter."}
|
|
{"id": 13492, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Plug-and-play proxy policy controlling retrieval via KG reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLM", "retriever"], "varies": ["retrieval policy"], "benchmarks": [], "results": [], "workflow": ["Proxy decides retrieval"], "summary": "SPARKLE uses a proxy model for structured adaptive retrieval decisions."}
|
|
{"id": 13781, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "analysis", "why": "How exposing reasoning chains alters LLM judgments of factuality"}], "pattern": "vary-conditions", "fixes": ["judge LLMs"], "varies": ["reasoning exposure"], "benchmarks": [], "results": [], "workflow": ["Vary reasoning presence"], "summary": "Studies how reasoning chains influence LLM judges of answer factuality."}
|
|
{"id": 13435, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Agentic learning for proactive tool retrieval and grounded execution"}], "pattern": "fix-env-vary-agent", "fixes": ["tool repos"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["Proactive retrieval", "Grounded execution"], "summary": "ToolOmni trains LLMs for open-world tool retrieval and execution."}
|
|
{"id": 15499, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Agentic verifier with forward and backward agents trained by RL"}], "pattern": "fix-env-vary-agent", "fixes": ["reasoning tasks"], "varies": ["verifier"], "benchmarks": [], "results": [], "workflow": ["Forward agent", "Backward agent", "Tool-augmented"], "summary": "AgentV-RL turns reward modeling into a tool-augmented multi-turn verifier."}
|
|
{"id": 4231, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "LLM-guided MCTS for knowledge graph QA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search"], "benchmarks": [], "results": [], "workflow": ["MCTS", "LLM guidance"], "summary": "DAMR uses LLM-guided MCTS for adaptive KGQA."}
|
|
{"id": 15733, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "LLM-based constraint optimization to reduce in-context reward hacking"}], "pattern": "vary-scaffold", "fixes": ["agent model"], "varies": ["constraint mechanism"], "benchmarks": [], "results": [], "workflow": ["Optimize constraints"], "summary": "LCO reduces in-context reward hacking in agents without finetuning."}
|
|
{"id": 5329, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Flow-matching distributional rewards to train explanation LLMs"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": ["CNF rewards", "RLAIF"], "summary": "Trains explanation LLMs via RL with CNF-generated distributional rewards."}
|
|
{"id": 18543, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Simulated group motivational interviewing environment"}, {"area": "domain-agents", "role": "train-agent", "why": "RL dialogue manager for group MI"}], "pattern": "fix-env-vary-agent", "fixes": ["LLM patients"], "varies": ["dialogue policy"], "benchmarks": [], "results": [], "workflow": ["SAC selects acts", "LLM generates"], "summary": "Simulation environment and RL facilitator for group motivational interviewing."}
|
|
{"id": 12896, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Red-teaming that hijacks agent reasoning and memory without prompt change"}], "pattern": "fix-agent-vary-env", "fixes": ["agent"], "varies": ["attack"], "benchmarks": [], "results": [], "workflow": ["Trigger extraction", "Reasoning hijacking", "Constraint tightening"], "summary": "JailAgent red-teams agents by manipulating reasoning and memory retrieval."}
|
|
{"id": 15572, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Experience-based warm-up for tool-integrated reasoning training"}], "pattern": "fix-env-vary-agent", "fixes": ["TIR tasks"], "varies": ["training paradigm"], "benchmarks": [], "results": [], "workflow": ["Integrate experiences", "Warm-up"], "summary": "E3-TIR is a warm-up training paradigm for tool-integrated reasoning agents."}
|
|
{"id": 3941, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Benchmark measuring propensity for harmful actions in agentic scenarios"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["scenarios"], "benchmarks": [{"name": "PropensityBench", "purpose": "latent safety propensity", "is_new": true}], "results": [], "workflow": ["Agentic scenarios"], "summary": "PropensityBench evaluates models' propensity to pursue dangerous actions."}
|
|
{"id": 11086, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Neuro-symbolic formal verification for agent oversight"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["oversight"], "benchmarks": [], "results": [], "workflow": ["Translate to specs", "Verify"], "summary": "FormalJudge uses formal verification instead of LLM-as-judge to oversee agents."}
|
|
{"id": 7459, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Student-generated trajectories with teacher correcting earliest error"}, {"area": "tool-use", "role": "train-agent", "why": "Distillation to small agents"}], "pattern": "vary-data", "fixes": ["student training"], "varies": ["trajectory data"], "benchmarks": [], "results": [], "workflow": ["Student rollout", "Teacher corrects"], "summary": "SCoRe distills agentic ability to small LLMs using student-centered corrected trajectories."}
|
|
{"id": 12475, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent framework for legal consultation"}, {"area": "domain-agents", "role": "benchmark", "why": "JurisCQAD legal consultation dataset"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["framework"], "benchmarks": [{"name": "JurisCQAD", "purpose": "legal consultation QA", "is_new": true}], "results": [], "workflow": ["Element graph", "Modular agents"], "summary": "JurisMA multi-agent framework and JurisCQAD dataset for Chinese legal consultation."}
|
|
{"id": 9261, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Expansion-aware router updates for scaling agent pools"}], "pattern": "vary-scaffold", "fixes": ["agents"], "varies": ["router updating"], "benchmarks": [], "results": [], "workflow": ["Familiarization tasks", "Harvest evidence"], "summary": "MonoScale scales multi-agent systems monotonically as new agents are added."}
|
|
{"id": 8700, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Interactive benchmark for autonomous LLM fine-tuning agents"}, {"area": "code-swe", "role": "environment", "why": "Sandbox environment for fine-tuning tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "FT-Dojo", "purpose": "autonomous fine-tuning", "is_new": true}], "results": [], "workflow": ["13 tasks", "5 domains"], "summary": "FT-Dojo is an interactive benchmark environment for agents performing LLM fine-tuning."}
|
|
{"id": 4885, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Teacher-assisted distillation to super-tiny function-calling models"}], "pattern": "fix-env-vary-agent", "fixes": ["function calling"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["Similarity guided", "Teacher refinement"], "summary": "STAR distills function calling into tiny models."}
|
|
{"id": 10049, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Embodied perception-execution loop for scientific software agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent loop"], "benchmarks": [], "results": [], "workflow": ["Perceive", "Execute"], "summary": "EmbodiedAct turns scientific software into active embodied agents."}
|
|
{"id": 15962, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Recurrence-based lazy memory consolidation to cut tokens"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory"], "benchmarks": [], "results": [], "workflow": ["Store lightly", "Consolidate on recurrence"], "summary": "RecMem consolidates memory based on recurrence for efficient long-running agents."}
|