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
39 lines
24 KiB
JSON
39 lines
24 KiB
JSON
{"id": 16020, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Self-generated error-reasoning data fine-tunes small models for Text-to-SQL, plus retrieval at inference"}], "pattern": "vary-data", "fixes": ["benchmarks", "base LLMs"], "varies": ["error-reasoning training data", "retrieval of corrections"], "benchmarks": [{"name": "SPIDER", "purpose": "evaluate execution accuracy", "is_new": false}, {"name": "BIRD", "purpose": "evaluate execution accuracy", "is_new": false}], "results": [], "workflow": ["Collect model errors", "Analyze via feedback prompts", "Fine-tune on error reasoning", "Retrieve corrections at inference"], "summary": "Self-improving Text-to-SQL framework that fine-tunes on its own error-reasoning data and retrieves correction examples at inference."}
|
|
{"id": 15116, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Data-efficient RL with few labeled easy samples and pseudo-labeled hard data for self-evolution"}], "pattern": "vary-data", "fixes": ["base LLMs"], "varies": ["RL data and reward scheme"], "benchmarks": [], "results": [], "workflow": ["Warm-up RL on few labeled", "Pseudo-label hard unlabeled data", "Progressive RL training"], "summary": "EasyRL self-evolves LLMs via RL using few easy labeled samples and divide-and-conquer pseudo-labeling on hard unlabeled data."}
|
|
{"id": 14263, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Dual-horizon advantage normalization for turn and session rewards in multi-turn RL sales agents"}, {"area": "social-sim", "role": "application", "why": "Industrial sales dialogue agent evaluated with a user simulator"}], "pattern": "fix-env-vary-agent", "fixes": ["user simulator environment"], "varies": ["credit assignment method", "reward normalization"], "benchmarks": [], "results": [], "workflow": ["Compute turn and session advantages", "Normalize each separately", "Fuse for policy update"], "summary": "DuCA separately normalizes turn-level and session-level advantages to stabilize multi-turn RL for industrial sales agents."}
|
|
{"id": 9076, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "data", "why": "Agent-native mid-training trajectories (contextual and environmental) for software engineering"}, {"area": "code-swe", "role": "train-agent", "why": "Mid-training recipe outperforming Kimi-Dev on SWE-Bench Verified"}], "pattern": "vary-data", "fixes": ["base model", "post-training setting"], "varies": ["mid-training data"], "benchmarks": [{"name": "SWE-Bench Verified", "purpose": "compare against Kimi-Dev", "is_new": false}], "results": [], "workflow": ["Build contextually-native trajectories", "Collect environmentally-native trajectories", "Mid-train then post-train"], "summary": "Agentic mid-training on agent-native trajectories improves software engineering agents over prior open mid-training recipes."}
|
|
{"id": 15593, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Evolutionary search over agentic workflows guided by human preferences with experience memory"}], "pattern": "vary-scaffold", "fixes": ["LLM backbones"], "varies": ["workflow structure", "prompts"], "benchmarks": [], "results": [], "workflow": ["Mutate and cross over workflows", "Guide with human preferences", "Store hierarchical experience memory"], "summary": "HFlow evolves agentic workflows via evolutionary search with human preference guidance and hierarchical experience memory."}
|
|
{"id": 1258, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Learned context generators produce per-round instructions for coherent multi-agent discussion"}, {"area": "multi-agent", "role": "train-agent", "why": "Trains context generators for each agent"}], "pattern": "vary-scaffold", "fixes": ["task benchmarks"], "varies": ["discussion context generation"], "benchmarks": [], "results": [], "workflow": ["Generate context per round", "Train generators for coherence", "Discuss and converge"], "summary": "M2CL trains per-agent context generators so multi-agent discussion avoids inconsistency and majority noise."}
|
|
{"id": 13993, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Initializer-Restorer starts diffusion NER from candidate spans to restore iterative refinement"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["diffusion initialization"], "benchmarks": [{"name": "ACE2004", "purpose": "analysis of inference steps", "is_new": false}], "results": [], "workflow": ["Get candidate spans from NER model", "Initialize diffusion from them", "Iteratively refine"], "summary": "Initializes diffusion-based NER from a standard model's candidate spans so iterative refinement helps again."}
|
|
{"id": 8867, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Live self-evolving benchmark of fresh Linux kernel crashes"}, {"area": "code-swe", "role": "environment", "why": "kEnv standardized crash-resolution environment"}], "pattern": "fix-agent-vary-env", "fixes": ["agent frameworks"], "varies": ["kernel bugs", "evaluation environment"], "benchmarks": [{"name": "Live-kBench", "purpose": "live kernel crash benchmark", "is_new": true}], "results": [], "workflow": ["Scrape fresh kernel bugs", "Run in kEnv", "Evaluate agents"], "summary": "Live-kBench and kEnv give a contamination-resistant, continuously updated benchmark and environment for kernel crash resolution agents."}
|
|
{"id": 13200, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "data", "why": "Automated synthetic data generation from a threat model for agent safety"}, {"area": "agent-safety", "role": "defense", "why": "Safety alignment of agents with self-reflective safe responses"}], "pattern": "vary-data", "fixes": ["safety benchmarks"], "varies": ["synthetic safety data"], "benchmarks": [], "results": [], "workflow": ["Decompose risk via OTS threat model", "Stress-test agents for failures", "Generate safe responses"], "summary": "SafeAgent synthesizes scenario-specific risk data from a threat model to improve LLM agent safety."}
|
|
{"id": 391, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "Induces fine-grained agent metrics from open-ended human feedback for LLM-judge evaluation"}], "pattern": "vary-evaluation", "fixes": ["agents under test"], "varies": ["evaluation metrics"], "benchmarks": [], "results": [], "workflow": ["Ground feedback to behaviors", "Cluster behaviors", "Create metrics for LLM judge"], "summary": "AutoLibra turns open-ended human feedback into concrete behavior metrics, with coverage and redundancy meta-metrics."}
|
|
{"id": 15833, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Generate-judge-refine agentic pipeline for treatment planning with a treatment-specific judge"}, {"area": "domain-agents", "role": "scaffold", "why": "Iterative inference-time refinement with frozen LLMs"}], "pattern": "vary-scaffold", "fixes": ["base LLMs"], "varies": ["generate-judge-refine loop"], "benchmarks": [], "results": [], "workflow": ["Generate draft plan", "Judge with TheraJudge", "Refine iteratively"], "summary": "TheraAgent iteratively generates, judges and refines treatment plans for more precise and safer regimens."}
|
|
{"id": 1136, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Multi-agent debate relevance assessment with human escalation"}, {"area": "eval/bench", "role": "benchmark", "why": "BRIDGE refined IR benchmark with missing relevant chunks"}], "pattern": "vary-evaluation", "fixes": ["retrievers"], "varies": ["relevance labeling protocol"], "benchmarks": [{"name": "BRIDGE", "purpose": "refined IR benchmark", "is_new": true}], "results": [{"benchmark": "BRIDGE", "system": "DREAM", "base_model": null, "setting": "proposed", "metric": "labeling accuracy", "score": 95.2, "unit": "%"}], "workflow": ["Debate with opposing stances", "Escalate uncertain cases to humans", "Rebuild benchmark"], "summary": "DREAM uses LLM debate to complete missing IR annotations, yielding the BRIDGE benchmark."}
|
|
{"id": 983, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Perceiver, planning and robustness modules for long-horizon GUI tasks"}, {"area": "web-gui", "role": "benchmark", "why": "LongGUIBench long-horizon game and app GUI tasks"}], "pattern": "vary-scaffold", "fixes": ["MLLM backbones"], "varies": ["GUI agent framework"], "benchmarks": [{"name": "LongGUIBench", "purpose": "long-horizon GUI evaluation", "is_new": true}], "results": [], "workflow": ["Detect elements and index them", "Plan long-horizon steps", "Execute and verify"], "summary": "LongHorizonUI improves long-horizon GUI agents and introduces LongGUIBench."}
|
|
{"id": 1304, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Selective KV-pair sharing as communication between LLMs"}], "pattern": "vary-scaffold", "fixes": ["models", "tasks"], "varies": ["communication protocol"], "benchmarks": [], "results": [], "workflow": ["Score KV layers by attention", "Select layers with Gaussian prior", "Share KV between LLMs"], "summary": "KVComm lets LLMs communicate efficiently by sharing selected KV pairs instead of text or hidden states."}
|
|
{"id": 12177, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "eval-method", "why": "RPS and AGS metrics quantify distillation-induced similarity in agent behavior"}, {"area": "tool-use", "role": "analysis", "why": "Compares 18 models across families on tau-Bench"}], "pattern": "vary-evaluation", "fixes": ["tau-Bench tasks"], "varies": ["similarity metrics", "models"], "benchmarks": [{"name": "tau-Bench", "purpose": "measure behavior similarity", "is_new": false}, {"name": "tau^2-Bench", "purpose": "measure behavior similarity", "is_new": false}], "results": [{"benchmark": "tau-Bench", "system": "within-family vs cross-family pairs", "base_model": null, "setting": "baseline", "metric": "AGS difference", "score": 5.9, "unit": "pp"}], "workflow": ["Model response patterns", "Build action graphs", "Compare against reference model"], "summary": "Introduces metrics for non-mandatory behavioral similarity among tool-use agents, finding within-family models are more similar."}
|
|
{"id": 4264, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "Intervention-driven debugging validated by whether failed tasks are repaired"}, {"area": "multi-agent", "role": "scaffold", "why": "Active verification via message and plan edits"}], "pattern": "vary-evaluation", "fixes": ["multi-agent systems"], "varies": ["debugging protocol"], "benchmarks": [], "results": [], "workflow": ["Generate failure hypotheses", "Intervene on messages or plans", "Verify task repair"], "summary": "DoVer debugs LLM multi-agent systems through targeted interventions and measures whether failures are actually fixed."}
|
|
{"id": 16824, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Single general-tool agent covering SWE, research and web browsing"}], "pattern": "vary-scaffold", "fixes": ["tasks"], "varies": ["tool set"], "benchmarks": [{"name": "SWE-Bench Multimodal", "purpose": "SWE evaluation", "is_new": false}, {"name": "GAIA", "purpose": "deep research evaluation", "is_new": false}, {"name": "The Agent Company", "purpose": "web task evaluation", "is_new": false}], "results": [], "workflow": ["Provide code, search, browser tools", "Add multimodal file viewer"], "summary": "OpenHands-Versa uses a small general toolset to match or beat specialized agents across SWE, research and web tasks."}
|
|
{"id": 9371, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Hardware-aware multi-armed bandit steers LLM kernel optimization search"}], "pattern": "vary-scaffold", "fixes": ["code LLMs"], "varies": ["search strategy"], "benchmarks": [], "results": [], "workflow": ["Prune with profiling bounds", "Cluster runtime behavior", "Bandit selects strategies"], "summary": "KernelBand frames GPU kernel optimization as a hierarchical bandit to guide LLM exploration."}
|
|
{"id": 10764, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Budget-aware MCTS, modular construction and reflective memory for ML engineering agents"}, {"area": "science-data", "role": "scaffold", "why": "Autonomous AI research agent"}], "pattern": "vary-scaffold", "fixes": ["LLM backbone"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["Plan with cost-aware MCTS", "Design-decompose-implement", "Reflect on solution differences"], "summary": "MARS is a modular agent with cost-aware search and comparative reflective memory for automated ML engineering."}
|
|
{"id": 14354, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Planning navigator, executor and Perceive tool for long-horizon multimodal tool use"}], "pattern": "vary-scaffold", "fixes": ["MLLMs"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["Global Navigator plans", "Agentic Executor uses tools", "Response Synthesizer answers"], "summary": "ToolScope unifies global planning with a Perceive tool to reduce visual context degradation in long-horizon VQA."}
|
|
{"id": 70, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "AnomReason benchmark for semantic anomalies in AI-generated images"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["anomaly annotations"], "benchmarks": [{"name": "AnomReason", "purpose": "semantic anomaly benchmark", "is_new": true}], "results": [], "workflow": ["Annotate via multi-agent pipeline", "Structure quadruples", "Evaluate models"], "summary": "AnomReason is a large benchmark of structured semantic anomaly annotations for AI-generated images."}
|
|
{"id": 17277, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Compares cheap-talk communication with curriculum learning for cooperation"}], "pattern": "vary-conditions", "fixes": ["games"], "varies": ["communication", "curriculum"], "benchmarks": [{"name": "Stag Hunt", "purpose": "cooperation test", "is_new": false}, {"name": "Iterated Public Goods Game with Punishment", "purpose": "curriculum test", "is_new": false}], "results": [{"benchmark": "4-player Stag Hunt", "system": "cheap talk", "base_model": null, "setting": "proposed", "metric": "cooperation", "score": 48.3, "unit": "%"}], "workflow": [], "summary": "Cheap-talk communication raises cooperation in LLM agents while curriculum learning proves design-sensitive."}
|
|
{"id": 14152, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Compute-aware Pareto study of ReAct variants and test-time strategies"}], "pattern": "vary-scaffold", "fixes": ["Gemini 2.5 Flash"], "varies": ["scaffold strategy", "token budget"], "benchmarks": [{"name": "HotPotQA", "purpose": "accuracy vs tokens", "is_new": false}, {"name": "FEVER", "purpose": "accuracy vs tokens", "is_new": false}, {"name": "GSM8K", "purpose": "accuracy vs tokens", "is_new": false}], "results": [], "workflow": [], "summary": "Budgeted evaluation of ReAct-style strategies finds Reflexion strongest on tool-heavy tasks but with optimistic costs."}
|
|
{"id": 13350, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Single-model internal debate to resolve knowledge conflicts in RAG"}], "pattern": "n/a", "fixes": ["model"], "varies": ["conflict-resolution prompting"], "benchmarks": [], "results": [], "workflow": ["Generate competing hypotheses", "Internal debate", "Resolve conflict"], "summary": "Debate-of-Thoughts resolves RAG knowledge conflicts through hypothesis generation and internal deliberation."}
|
|
{"id": 1251, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Collaborative Battleship decision-oriented dialogue task vs humans"}, {"area": "planning-decision", "role": "scaffold", "why": "Monte Carlo inference strategies improving information seeking"}], "pattern": "vary-scaffold", "fixes": ["LM agents"], "varies": ["inference strategy"], "benchmarks": [{"name": "Collaborative Battleship", "purpose": "explore-act dialogue task", "is_new": true}], "results": [], "workflow": ["Sample hypotheses", "Estimate information gain", "Choose questions and shots"], "summary": "Introduces Collaborative Battleship and Bayesian Monte Carlo inference that improves LM agents' information seeking."}
|
|
{"id": 15017, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "eval-method", "why": "AIC metric measures answer stability across stochastic passes for hallucination detection"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["detection metric"], "benchmarks": [], "results": [], "workflow": ["Sample stochastic passes", "Measure answer stability", "Combine with output uncertainty"], "summary": "AIC is an epistemic-aware answer-level metric combined with output uncertainty to detect hallucinations."}
|
|
{"id": 3928, "in_agent_scope": true, "types": [{"area": "science-data", "role": "environment", "why": "MedAgentGym executable sandbox for biomedical coding agents"}, {"area": "science-data", "role": "benchmark", "why": "Benchmarks 29 LLMs on biomedical data science"}, {"area": "science-data", "role": "train-agent", "why": "Med-Copilot trained with offline and online RL"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks and sandbox", "RL training"], "benchmarks": [{"name": "MedAgentGym", "purpose": "biomedical coding tasks", "is_new": true}], "results": [], "workflow": ["Build sandbox tasks", "Sample trajectories", "Train with SFT and RL"], "summary": "MedAgentGym provides scalable training environment and benchmark, enabling Med-Copilot gains via offline and online RL."}
|
|
{"id": 8582, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "RL-trained model iteratively zooms with tools for GUI grounding"}], "pattern": "vary-both", "fixes": ["benchmark"], "varies": ["grounding model", "tool-invoked focus refinement"], "benchmarks": [{"name": "ScreenSpot-Pro", "purpose": "GUI grounding accuracy", "is_new": false}], "results": [{"benchmark": "ScreenSpot-Pro", "system": "GUI-Spotlight", "base_model": null, "setting": "finetuned", "metric": "accuracy", "score": 52.8, "unit": "%"}], "workflow": ["Invoke specialized tools", "Narrow focus iteratively", "Predict target element"], "summary": "GUI-Spotlight trains a model to iteratively focus on screen regions, improving GUI visual grounding."}
|
|
{"id": 13178, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Action generator and reflection module for verifying and refining KBQA logical forms"}], "pattern": "vary-both", "fixes": ["KBQA benchmarks"], "varies": ["generate-verify-refine framework"], "benchmarks": [], "results": [], "workflow": ["Generate query paths hop-by-hop", "Reflect on executability", "Refine logical form"], "summary": "ARI-KBQA generates, verifies and refines logical forms for reliable knowledge base question answering."}
|
|
{"id": 18610, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "CUS-QA regional open-ended QA dataset in text and vision"}, {"area": "eval/judge", "role": "eval-method", "why": "Analyzes reliability of automatic metrics against human judgments"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["regional QA data"], "benchmarks": [{"name": "CUS-QA", "purpose": "regional QA benchmark", "is_new": true}], "results": [], "workflow": [], "summary": "CUS-QA is a local-knowledge QA benchmark for Czech, Slovak and Ukrainian contexts, with analysis of automatic metrics."}
|
|
{"id": 14771, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "train-agent", "why": "Plan-refine agent framework used to build SFT data and hierarchical DPO with MCTS for long-form writing"}], "pattern": "vary-data", "fixes": ["base model"], "varies": ["training data and DPO"], "benchmarks": [], "results": [], "workflow": ["Plan and refine text", "Build SFT dataset", "Hierarchical DPO via MCTS"], "summary": "SuperWriter introduces structured plan-refine generation and trains a 7B model with SFT and hierarchical DPO."}
|
|
{"id": 14438, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Reasoning reward model gives structured feedback for agentic RL"}], "pattern": "vary-both", "fixes": ["agent benchmarks"], "varies": ["reward signal", "integration strategy"], "benchmarks": [], "results": [], "workflow": ["Score trajectories with reasoning trace", "Critique flaws", "Integrate feedback in text or reward"], "summary": "Agent-RRM provides structured reasoning feedback for agentic trajectories, with three integration strategies."}
|
|
{"id": 15111, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Diversity-scaling trajectory synthesis for MCP code agents"}], "pattern": "vary-data", "fixes": ["training budget"], "varies": ["trajectory diversity"], "benchmarks": [], "results": [], "workflow": ["Cluster business services", "Blueprint-driven multi-agent generation", "Scale diversity not quantity"], "summary": "TDScaling synthesizes diverse tool-use trajectories, showing diversity beats quantity for code agent training."}
|
|
{"id": 4912, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "MCPMark 127 realistic CRUD-heavy MCP tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["minimal agent framework", "LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "MCPMark", "purpose": "stress-test MCP use", "is_new": true}], "results": [], "workflow": [], "summary": "MCPMark benchmarks LLMs on 127 realistic, programmatically verified MCP tasks."}
|
|
{"id": 10389, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "environment", "why": "CTF-Dojo executable CTF environments with automated CTF-Forge pipeline"}, {"area": "agent-safety", "role": "train-agent", "why": "Trains agents to find vulnerabilities on verified trajectories"}], "pattern": "fix-agent-vary-env", "fixes": ["training recipe"], "varies": ["environments"], "benchmarks": [{"name": "CTF-Dojo", "purpose": "CTF training runtime", "is_new": true}], "results": [], "workflow": ["Convert artifacts to environments", "Sample verified trajectories", "Train agents"], "summary": "CTF-Dojo supplies 658 containerized CTF environments to train cybersecurity agents."}
|
|
{"id": 2517, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Training-free LLM planning with vision-guided layout refinement for 3D scenes"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["scene generation framework"], "benchmarks": [], "results": [], "workflow": ["LLM drafts layout", "Vision refines layout", "Check physical plausibility"], "summary": "Scenethesis couples LLM scene planning with vision-guided refinement for plausible interactive 3D scenes."}
|
|
{"id": 16346, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "analysis", "why": "Controlled study of how LLMs aggregate conflicting retrieved evidence"}, {"area": "knowledge/rag", "role": "benchmark", "why": "GroupQA controversial questions with stance-annotated evidence"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["evidence presentation", "evidence strength"], "benchmarks": [{"name": "GroupQA", "purpose": "evidence aggregation study", "is_new": true}], "results": [], "workflow": [], "summary": "GroupQA dataset reveals LLMs' evidence aggregation biases such as order and repetition effects in RAG."}
|
|
{"id": 5546, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Design and control agents debate with LLM judges for robot co-design"}, {"area": "embodied-nav", "role": "application", "why": "Morphology-reward co-optimization on MuJoCo"}], "pattern": "vary-scaffold", "fixes": ["simulator"], "varies": ["debate-based design loop"], "benchmarks": [{"name": "MuJoCo locomotion benchmarks", "purpose": "co-design evaluation", "is_new": false}], "results": [], "workflow": ["Design agent proposes", "Control agent responds", "Judges give feedback"], "summary": "Debate2Create uses multi-agent LLM debate grounded in physics simulation for robot morphology and reward co-design."}
|