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
32 KiB
JSON
61 lines
32 KiB
JSON
{"id": 14765, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Studies human delegation and adoption decisions with AI agents in a QA game"}], "pattern": "vary-conditions", "fixes": ["AI agents", "game setup"], "varies": ["human reliance decisions", "human-AI pairings"], "benchmarks": [], "results": [], "workflow": [], "summary": "Human-subject study of when humans delegate to or adopt suggestions from AI teammates in cooperative question answering."}
|
|
{"id": 804, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Self-evolving Challenger-Solver co-training from a single base LLM without data"}], "pattern": "vary-data", "fixes": ["base LLM"], "varies": ["self-generated tasks", "training algorithm"], "benchmarks": [], "results": [], "workflow": ["Init Challenger and Solver", "Challenger proposes frontier tasks", "Solver trains on tasks", "Co-evolve iteratively"], "summary": "Fully autonomous framework where Challenger and Solver co-evolve to improve reasoning with zero external data."}
|
|
{"id": 9583, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "RPA-SFT and alignment training to resolve knowledge conflicts in KI-VQA"}, {"area": "arch/multimodal", "role": "benchmark", "why": "Constructs REAL-VQA dataset"}], "pattern": "vary-data", "fixes": ["MLLM backbone"], "varies": ["training method", "conflict supervision"], "benchmarks": [{"name": "REAL-VQA", "purpose": "training and evaluating conflict handling", "is_new": true}], "results": [], "workflow": ["Build REAL-VQA", "Reasoning-pivot aware SFT", "Pivot alignment training"], "summary": "Reasoning-pivot framework with dataset and training to handle knowledge conflicts in knowledge-intensive VQA."}
|
|
{"id": 14022, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Structured four-network agent memory with retain, recall, reflect operations"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory architecture"], "benchmarks": [{"name": "LongMemEval", "purpose": "evaluating long-term memory", "is_new": false}, {"name": "LoCoMo", "purpose": "evaluating long-term memory", "is_new": false}], "results": [], "workflow": ["Retain into four networks", "Hybrid parallel recall", "Reflect to update opinions"], "summary": "Memory system separating facts from beliefs with temporal entity graphs, evaluated on LongMemEval and LoCoMo."}
|
|
{"id": 12040, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Adaptive program reasoning for long video understanding with VideoLLMs"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["reasoning method"], "benchmarks": [], "results": [], "workflow": ["Adaptive reasoning decision", "Generate executable programs", "Refine low-confidence predictions"], "summary": "Framework letting VideoLLMs adaptively reason with executable programs and self-refine for long video understanding."}
|
|
{"id": 1957, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Introduces 200-task continuous control benchmark"}, {"area": "embodied-nav", "role": "train-agent", "why": "Newt multitask world model pretrained then online RL"}], "pattern": "vary-both", "fixes": ["RL setting"], "varies": ["tasks", "world model training"], "benchmarks": [{"name": "Newt 200-task benchmark", "purpose": "multitask continuous control", "is_new": true}], "results": [], "workflow": ["Pretrain on demonstrations", "Joint online RL across tasks"], "summary": "Benchmark of 200 control tasks and Newt, a language-conditioned multitask world model trained with online RL."}
|
|
{"id": 15012, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Compresses KV cache for latent communication between LLM agents"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["communication compression"], "benchmarks": [], "results": [], "workflow": ["Learnable probes compress KV", "Fixed-size latent transfer", "Bound error theoretically"], "summary": "CondenseFlow compresses KV caches for O(1) latent communication in multi-agent systems, cutting memory over 99%."}
|
|
{"id": 4087, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "data", "why": "Counterfactual replay and fault injection annotate failed trajectories into TracerTraj"}, {"area": "multi-agent", "role": "train-agent", "why": "Trains AgenTracer-8B failure tracer with RL"}], "pattern": "vary-data", "fixes": ["failure attribution task"], "varies": ["training data", "tracer model"], "benchmarks": [{"name": "TracerTraj", "purpose": "failure attribution training", "is_new": true}], "results": [], "workflow": ["Annotate via counterfactual replay", "Inject programmed faults", "Train tracer with RL"], "summary": "Automated annotation of failed multi-agent trajectories and an 8B tracer for failure attribution."}
|
|
{"id": 16458, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multiple LLMs collaborate within MCTS to propose and aggregate reasoning steps"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-LLM search"], "benchmarks": [], "results": [], "workflow": ["Agents propose steps", "MCTS backbone search", "Aggregate and refine"], "summary": "Mixture-of-Search-Agents uses multiple LLMs in MCTS for better reasoning on math and commonsense tasks."}
|
|
{"id": 15839, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Temporal semantic memory with semantic timeline and durative memory for personalization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory design"], "benchmarks": [], "results": [], "workflow": ["Build semantic timeline", "Consolidate durative memories", "Temporal retrieval"], "summary": "Memory framework modeling actual occurrence time and durative information for personalized dialogue agents."}
|
|
{"id": 15140, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "SFT then RL with order-sensitive rewards for temporal tool use in TKGQA"}], "pattern": "fix-env-vary-agent", "fixes": ["TKGQA task"], "varies": ["training recipe", "tool API"], "benchmarks": [], "results": [], "workflow": ["Design temporal tool API", "SFT tool-interleaved CoT", "RL with fine-grained rewards"], "summary": "Tool-integrated reasoning model trained with SFT and RL to call temporal tools for knowledge graph QA."}
|
|
{"id": 16775, "in_agent_scope": false, "types": [{"area": "eval/robust", "role": "eval-method", "why": "Linguistically-driven minimal variations of test data for robustness evaluation"}, {"area": "eval/robust", "role": "analysis", "why": "Finds task-dependent brittleness across models"}], "pattern": "vary-evaluation", "fixes": ["models"], "varies": ["linguistic variations"], "benchmarks": [], "results": [], "workflow": [], "summary": "FLUKE framework generates controlled linguistic perturbations to assess model robustness across six NLP tasks."}
|
|
{"id": 13020, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Jointly trains solve, critique and guided re-exploration in one policy via RL"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL objective"], "benchmarks": [], "results": [], "workflow": ["Solve", "Critique own answer", "Hint-guided re-exploration"], "summary": "RL framework jointly optimizing solving, critiquing and re-exploration for test-time self-improvement."}
|
|
{"id": 8901, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "ProAct-75 proactive agent benchmark with step annotations and task graphs"}, {"area": "embodied-nav", "role": "method", "why": "ProAct-Helper MLLM baseline grounded in task graphs"}], "pattern": "fix-env-vary-agent", "fixes": ["MLLM baseline"], "varies": ["tasks"], "benchmarks": [{"name": "ProAct-75", "purpose": "training and evaluating proactive agents", "is_new": true}], "results": [], "workflow": ["Ground decisions in task graphs"], "summary": "Benchmark of 75 tasks with 91,581 annotated steps and a task-graph MLLM baseline for proactive agents."}
|
|
{"id": 7252, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Planner-Coder-Verifier multi-agent end-to-end GPU program generation"}, {"area": "code-swe", "role": "train-agent", "why": "Rubric-based agentic RL for the Coder"}], "pattern": "vary-both", "fixes": ["GPU tasks"], "varies": ["agent design", "coder training"], "benchmarks": [], "results": [], "workflow": ["Planner designs system", "Coder implements", "Verifier profiles with Nsys/NCU"], "summary": "Multi-agent framework with rubric-based agentic RL for end-to-end GPU program generation."}
|
|
{"id": 12665, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "RL with summarization-based context compression optimized end-to-end for long-horizon tool use"}], "pattern": "fix-env-vary-agent", "fixes": ["tool-use tasks"], "varies": ["context management", "RL pipeline"], "benchmarks": [], "results": [], "workflow": ["Periodically summarize history", "Policy gradient over summaries", "Train end-to-end"], "summary": "Agentic RL that learns context summarization jointly with tool use to scale beyond context limits."}
|
|
{"id": 13748, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "WiserUI-Bench of UI A/B test pairs for MLLM UI/UX understanding"}], "pattern": "fix-agent-vary-env", "fixes": ["MLLMs"], "varies": ["UI design pairs"], "benchmarks": [{"name": "WiserUI-Bench", "purpose": "predicting winning UI and interpreting why", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 300 real A/B-tested UI pairs evaluating whether MLLMs understand how design affects user behavior."}
|
|
{"id": 14384, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Trait, distortion, vision and speech agents fused for emotion recognition"}], "pattern": "vary-scaffold", "fixes": ["datasets"], "varies": ["agent framework"], "benchmarks": [{"name": "MELD", "purpose": "emotion recognition", "is_new": false}, {"name": "IEMOCAP", "purpose": "emotion recognition", "is_new": false}], "results": [], "workflow": ["Trait and distortion agents", "Vision and speech agents", "Fusion agent predicts"], "summary": "Multi-agent cognitive-modeling framework for multimodal emotion recognition in conversations."}
|
|
{"id": 6299, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Multi-agent educational social simulator with value-driven agents and measurement protocol"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["simulation settings"], "benchmarks": [], "results": [], "workflow": ["Configure value-driven agents", "Simulate social dynamics", "Dual-track measurement"], "summary": "LLM multi-agent simulator for studying educational social dynamics with psychologically grounded agents."}
|
|
{"id": 12848, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Newsroom-style editor and reporter agents for timeline summarization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["multi-agent structure"], "benchmarks": [], "results": [], "workflow": ["Editor allocates evidence", "Reporters draft events", "Adjudicate and consolidate"], "summary": "MAS-TLS multi-agent framework producing balanced, less redundant chronological reports."}
|
|
{"id": 5915, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Agentic LLM symbolic regression decoupling generation from search control"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search guidance"], "benchmarks": [{"name": "LLM-SRBench", "purpose": "symbolic regression evaluation", "is_new": false}], "results": [], "workflow": ["Adaptive operators", "Structural diagnosis tools", "Reflective memory"], "summary": "Deliberate Evolution agentic framework for sample-efficient LLM symbolic regression using 40% of standard budget."}
|
|
{"id": 14762, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Builds ToolPref-Pairwise-30K preference data"}, {"area": "tool-use", "role": "train-agent", "why": "Trains ToolRM reward models"}, {"area": "tool-use", "role": "benchmark", "why": "TRBench-BFCL for tool-use reward models"}], "pattern": "vary-data", "fixes": ["Qwen3 base models"], "varies": ["preference data", "reward model"], "benchmarks": [{"name": "BFCL", "purpose": "tool calling basis", "is_new": false}, {"name": "TRBench_BFCL", "purpose": "evaluating tool-use RMs", "is_new": true}], "results": [], "workflow": ["Rule-based pairwise data", "Train RMs", "Evaluate on TRBench"], "summary": "Reward models for tool use with a pairwise preference dataset and benchmark."}
|
|
{"id": 14337, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Localization-aware evaluation framework for Korean in-vehicle assistants"}, {"area": "apps/dialogue", "role": "analysis", "why": "Finds unstable honorific control and weak clarification"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["evaluation framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Evaluation framework for Korean in-vehicle assistants revealing unstable honorifics and weak proactivity."}
|
|
{"id": 1964, "in_agent_scope": true, "types": [{"area": "search-research", "role": "data", "why": "WebLeaper synthesizes entity-rich IS tasks and efficient trajectories"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training tasks", "trajectories"], "benchmarks": [], "results": [], "workflow": ["Tree-structured IS tasks", "Embed many target entities", "Generate efficient trajectories"], "summary": "Constructs high-coverage information-seeking tasks and trajectories to train more search-efficient web agents."}
|
|
{"id": 2255, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Training-free CAD generation with DeepSeek-R1 and visual plus CoT feedback"}], "pattern": "vary-scaffold", "fixes": ["DeepSeek-R1"], "varies": ["refinement feedback"], "benchmarks": [], "results": [], "workflow": ["Generate parametric CAD", "Visual and CoT feedback", "Self-refine"], "summary": "Training-free self-refining parametric CAD generation using locally deployed DeepSeek-R1."}
|
|
{"id": 13214, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "environment", "why": "MANTIS multi-agent terminal simulator robust to prompt injection"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM"], "varies": ["simulator design"], "benchmarks": [], "results": [], "workflow": ["Filter-based command routing", "Agentic file system", "History pruning"], "summary": "Multi-agent terminal simulation framework improving realism, state consistency and prompt-injection robustness."}
|
|
{"id": 14688, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent self-correcting prompt refinement for complex text-to-video"}], "pattern": "vary-scaffold", "fixes": ["T2V model"], "varies": ["prompt refinement"], "benchmarks": [], "results": [], "workflow": ["Route to scenario", "Policy-conditioned rewriting", "Semantic verification and revision"], "summary": "Self-correcting multi-agent prompt refinement framework for complex-scenario text-to-video generation."}
|
|
{"id": 1391, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Dark patterns steer web agents"}, {"area": "agent-safety", "role": "benchmark", "why": "DECEPTICON environment of 850 dark-pattern tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["web agents"], "varies": ["dark pattern tasks"], "benchmarks": [{"name": "DECEPTICON", "purpose": "testing dark pattern effectiveness", "is_new": true}], "results": [{"benchmark": "DECEPTICON", "system": "SOTA agents", "base_model": null, "setting": "baseline", "metric": "dark pattern success rate (over)", "score": 70, "unit": "%"}], "workflow": [], "summary": "Shows dark patterns steer web agents to malicious outcomes in over 70% of tasks via a new test environment."}
|
|
{"id": 16730, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "QFrBLiMP Quebec-French minimal pairs benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["minimal pairs"], "benchmarks": [{"name": "QFrBLiMP", "purpose": "Quebec-French grammaticality", "is_new": true}, {"name": "MultiBLiMP-Fr", "purpose": "comparison", "is_new": false}], "results": [], "workflow": [], "summary": "1,761 human-annotated Quebec-French minimal pairs evaluating LLM grammatical competence."}
|
|
{"id": 15096, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Studies generalization and coordination of adaptive multi-agent systems"}], "pattern": "vary-conditions", "fixes": ["MAS methods"], "varies": ["domains"], "benchmarks": [], "results": [], "workflow": [], "summary": "Empirical study finding adaptive multi-agent systems overfit topology and show illusory coordination."}
|
|
{"id": 13051, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Hierarchical VLM planner plus low-level executor for image-goal navigation"}], "pattern": "fix-env-vary-agent", "fixes": ["navigation task"], "varies": ["policy architecture"], "benchmarks": [], "results": [], "workflow": ["VLM short-horizon plan", "Low-level execution"], "summary": "HRNav decomposes image-goal navigation into VLM high-level planning and low-level execution."}
|
|
{"id": 4852, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Dual-stream decoupled attention for faster view synthesis"}], "pattern": "n/a", "fixes": ["RealEstate10K"], "varies": ["architecture"], "benchmarks": [{"name": "RealEstate10K", "purpose": "novel view synthesis", "is_new": false}], "results": [{"benchmark": "RealEstate10K", "system": "Efficient-LVSM", "base_model": null, "setting": "proposed", "metric": "PSNR", "score": 29.86, "unit": "dB"}], "workflow": [], "summary": "Efficient-LVSM decouples attention for novel view synthesis with faster training and inference."}
|
|
{"id": 16017, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "AutoScale multi-agent scale generation"}, {"area": "science-data", "role": "benchmark", "why": "SCALE-1.2K dataset and evaluation"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["agent workflow"], "benchmarks": [{"name": "SCALE-1.2K", "purpose": "scale development benchmark", "is_new": true}], "results": [], "workflow": ["Collaborative agents draft", "Data-driven refinement", "Multi-LLM judging"], "summary": "Multi-agent framework automating psychological scale generation with a dataset and evaluation system."}
|
|
{"id": 14525, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Critic-free RL with pre-estimated value baseline and data filtering"}], "pattern": "fix-env-vary-agent", "fixes": ["agentic tasks"], "varies": ["advantage estimation"], "benchmarks": [], "results": [], "workflow": ["Pre-rollout value estimates", "Filter samples", "Train with static baseline"], "summary": "PVPO stabilizes group-based RL for agentic reasoning using pre-estimated value baselines."}
|
|
{"id": 15893, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Meta-cognitive agent inferring persuadee mental states for persuasion"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["Infer latent mental states", "Choose strategy", "Generate response"], "summary": "Meta-cognitive agent framework for complex persuasive dialogue."}
|
|
{"id": 16233, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "analysis", "why": "Counterfactual fairness evaluation of LLM contact-center QA across 18 LLMs"}, {"area": "safety/fairness", "role": "eval-method", "why": "CFR and MASD fairness metrics"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["counterfactual attributes"], "benchmarks": [], "results": [], "workflow": [], "summary": "Measures demographic and style bias in LLM QA scoring via counterfactual flip rate."}
|
|
{"id": 17395, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "application", "why": "Thesis proposal on perspective-aware knowledge extraction with MAS and ontologies"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": ["Ontology patterns", "Multi-agent extraction", "Evaluate on medical data"], "summary": "Proposal for multi-agent LLM extraction preserving conflicting stakeholder perspectives in knowledge graphs."}
|
|
{"id": 17174, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Self-prediction training and UCB-guided re-masking for discrete diffusion"}], "pattern": "n/a", "fixes": ["diffusion LM"], "varies": ["training objective", "decoding"], "benchmarks": [], "results": [], "workflow": ["Deeper self-prediction training", "UCB re-masking decoding"], "summary": "Unified training and bandit-based decoding closing training-inference gap in discrete text diffusion."}
|
|
{"id": 14664, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Tag-guided hierarchical knowledge graph RAG"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["RAG design"], "benchmarks": [], "results": [], "workflow": ["Extract tags", "Build tag chains", "Hierarchical retrieval"], "summary": "TagRAG builds hierarchical tag graphs for efficient global reasoning and incremental updates."}
|
|
{"id": 1530, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Lightweight sequential monitors against decomposition attacks"}, {"area": "agent-safety", "role": "benchmark", "why": "DecomposedHarm dataset of 4,634 tasks"}], "pattern": "vary-data", "fixes": ["threat model"], "varies": ["monitor"], "benchmarks": [{"name": "DecomposedHarm", "purpose": "evaluating decomposition attacks and monitors", "is_new": true}], "results": [{"benchmark": "DecomposedHarm", "system": "frontier models", "base_model": null, "setting": "baseline", "metric": "attack success rate (average)", "score": 87, "unit": "%"}], "workflow": ["Observe conversation", "Detect harmful intent sequence"], "summary": "External sequential monitor and DecomposedHarm dataset for decomposed harmful agent tasks."}
|
|
{"id": 18361, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "method", "why": "Classifier-based extraction of answer candidates for question generation"}], "pattern": "n/a", "fixes": ["FairytaleQA"], "varies": ["candidate selection"], "benchmarks": [{"name": "FairytaleQA", "purpose": "training and evaluation", "is_new": false}], "results": [], "workflow": ["Parse candidates", "Generate questions", "Classify quality with DeBERTa"], "summary": "Extracts and quality-scores answer candidates for educational question generation."}
|
|
{"id": 3489, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "ACE evolving context playbooks via generation, reflection and curation"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["context adaptation"], "benchmarks": [], "results": [], "workflow": ["Generate", "Reflect", "Curate incremental updates"], "summary": "ACE treats contexts as evolving playbooks to enable self-improving agents without weight updates."}
|
|
{"id": 3753, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "EHR-ChatQA interactive database agent benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["interaction flows"], "benchmarks": [{"name": "EHR-ChatQA", "purpose": "clarification, tool use, SQL generation", "is_new": true}], "results": [], "workflow": [], "summary": "Interactive benchmark with simulated users for EHR database agents handling ambiguity and value mismatch."}
|
|
{"id": 7444, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "CompDiag-Bench sequential diagnosis benchmark"}, {"area": "domain-agents", "role": "train-agent", "why": "Salus multi-agent RL for diagnostic testing"}], "pattern": "vary-both", "fixes": ["clinical task"], "varies": ["agent training"], "benchmarks": [{"name": "CompDiag-Bench", "purpose": "sequential diagnostic testing", "is_new": true}], "results": [], "workflow": ["Request tests", "Multi-agent RL", "Reach diagnosis"], "summary": "Benchmark and multi-agent RL method for strategic diagnostic test selection."}
|
|
{"id": 17629, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "benchmark", "why": "Guidelines for Labovian narrative analysis of Japanese"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Annotation guidelines for Labovian structural analysis of Japanese oral narratives with high annotator agreement."}
|
|
{"id": 6347, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Learns reusable skills via non-parametric PPO without parameter updates"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["skill library"], "benchmarks": [], "results": [], "workflow": ["Formalize Skill-MDP", "Generate candidate skills", "PPO gate verification"], "summary": "Skill-Pro lets LLM agents learn and verify reusable procedural skills without training."}
|
|
{"id": 7326, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "RL-trained search controller for object-grounded video reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["reasoning framework"], "benchmarks": [], "results": [], "workflow": ["Search-guided controller", "Ground steps to regions", "RL with format reward"], "summary": "Chain-of-Glimpse anchors video reasoning steps to object regions with RL-trained controller."}
|
|
{"id": 7933, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "AutoElicit perturbs benign instructions to elicit unsafe CUA behaviors"}], "pattern": "fix-agent-vary-env", "fixes": ["CUAs"], "varies": ["instruction perturbations"], "benchmarks": [], "results": [], "workflow": ["Perturb benign instruction", "Use execution feedback", "Elicit harms"], "summary": "Agentic framework surfacing hundreds of severe unintended behaviors in computer-use agents from benign inputs."}
|
|
{"id": 1788, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "SFT on CoT dataset then RL for driving VLA reasoning"}, {"area": "embodied-nav", "role": "data", "why": "Builds nuScenesR2-6K CoT dataset"}], "pattern": "vary-data", "fixes": ["driving task"], "varies": ["training recipe"], "benchmarks": [{"name": "nuScenesR2-6K", "purpose": "SFT dataset", "is_new": true}], "results": [], "workflow": ["CoT SFT", "RL for reasoning", "Self-reflection"], "summary": "VLA driving model with reasoning and self-reflection trained via CoT SFT and RL."}
|
|
{"id": 16274, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "data", "why": "Instruction dataset for Arabic poetry generation"}, {"area": "apps/dialogue", "role": "train-agent", "why": "Fine-tunes LLMs on poetry data"}], "pattern": "vary-data", "fixes": ["LLMs"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": [], "summary": "Large instruction dataset for Arabic and dialect poetry tasks, with fine-tuning experiments."}
|
|
{"id": 12923, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Coverage-guided MCP agent benchmark with metamorphic testing"}, {"area": "tool-use", "role": "eval-method", "why": "Fact-grounded metamorphic evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["MCP agents"], "varies": ["tasks"], "benchmarks": [{"name": "SGVEF benchmark", "purpose": "evaluating MCP agents", "is_new": true}], "results": [], "workflow": ["Adaptive exploration", "Synthesize metamorphic pairs", "Validate", "Iterate"], "summary": "Framework generating high-coverage MCP benchmark and evaluating 8 agents."}
|
|
{"id": 17533, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Few-shot CoT LLM agents for language service selection"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["prompting procedure"], "benchmarks": [], "results": [], "workflow": ["Extract functional requirements", "Induce non-functional needs", "Select services"], "summary": "User-centered LLM agents select and compose language services via structured reasoning."}
|
|
{"id": 12713, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Markovian structure-aware alignment of LLMs for time series forecasting"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["alignment"], "benchmarks": [], "results": [], "workflow": ["Distill Markov graph from language", "Transfer prior to time series"], "summary": "MGSAA aligns linguistic and temporal structure to adapt LLMs for forecasting."}
|
|
{"id": 10532, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "SimpleMem semantic lossless compression memory with intent-aware retrieval"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory framework"], "benchmarks": [], "results": [], "workflow": ["Structured compression", "Online synthesis", "Intent-aware retrieval"], "summary": "Efficient lifelong memory for LLM agents using semantic compression."}
|
|
{"id": 17847, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "JamC-QA Japan-specific knowledge benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["questions"], "benchmarks": [{"name": "JamC-QA", "purpose": "Japan-specific knowledge", "is_new": true}], "results": [], "workflow": [], "summary": "2,309 human-written multiple-choice questions on Japan-specific knowledge across eight categories."}
|
|
{"id": 3723, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Unified medical segmentation integrating priors, in-context and interactive knowledge"}], "pattern": "n/a", "fixes": [], "varies": ["model design"], "benchmarks": [], "results": [], "workflow": ["Semantic priors", "In-context examples", "Interactive feedback"], "summary": "K-Prism unifies three knowledge paradigms in a universal medical image segmentation model."}
|
|
{"id": 11635, "in_agent_scope": false, "types": [{"area": "train/continual", "role": "method", "why": "Selective backward refinement for prompt-based continual learning"}], "pattern": "n/a", "fixes": ["backbones"], "varies": ["continual method"], "benchmarks": [], "results": [], "workflow": ["Check task correlation", "Refine in non-interfering directions"], "summary": "SABER enables positive backward transfer in replay-free prompt continual learning."}
|
|
{"id": 11588, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "method", "why": "LLM-guided loop bound generation with formal validation for termination proofs"}], "pattern": "n/a", "fixes": ["termination benchmarks"], "varies": ["verification tool"], "benchmarks": [], "results": [], "workflow": ["LLM guesses loop bounds", "Formal check", "Violation analysis"], "summary": "LIFT combines LLM bound generation with sound validation for program termination verification."}
|
|
{"id": 10401, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Two-stage red-teaming for reasoning-level denial-of-service on agents"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["adversarial trigger", "payload"], "benchmarks": [{"name": "WebShop", "purpose": "agent testbed", "is_new": false}], "results": [], "workflow": ["Optimize adversarial trigger", "ICL-guided genetic payload"], "summary": "OTora inflates agent reasoning depth and tool use while preserving correctness."}
|
|
{"id": 1992, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Dual-agent planner and writer with dynamic outlines for deep research"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["research pipeline"], "benchmarks": [], "results": [], "workflow": ["Planner interleaves search and outline", "Memory bank of evidence", "Writer composes sections"], "summary": "WebWeaver dual-agent framework for open-ended deep research reports."}
|