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
33 KiB
JSON
61 lines
33 KiB
JSON
{"id": 16361, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Agentic causal disentanglement framework for long-tailed domain generalization in clinical imaging"}], "pattern": "vary-scaffold", "fixes": ["benchmark"], "varies": ["DG/LT method"], "benchmarks": [], "results": [], "workflow": ["Encode clinical knowledge", "Disentangle causal features", "Automated pipeline construction"], "summary": "Modular framework integrating clinical expert knowledge as causal intervention to disentangle domain-invariant and class-discriminative features."}
|
|
{"id": 4574, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Turn-level information-gain reward for RL of multi-turn search agents"}, {"area": "search-research", "role": "train-agent", "why": "Trains search agents with dense intrinsic rewards"}], "pattern": "fix-env-vary-agent", "fixes": ["search environment", "benchmarks"], "varies": ["reward design", "RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Sample rollouts", "Compute per-turn information gain", "Policy optimization"], "summary": "IGPO gives dense turn-level rewards from information gain on the ground-truth answer to train multi-turn search agents."}
|
|
{"id": 5811, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "CaP-Bench evaluates coding agents across abstraction and perception levels for robot manipulation"}, {"area": "embodied-nav", "role": "environment", "why": "CaP-Gym interactive environment for code-as-policy agents"}], "pattern": "vary-both", "fixes": ["frontier models (benchmark)"], "varies": ["abstraction level", "interaction", "perceptual grounding"], "benchmarks": [{"name": "CaP-Bench", "purpose": "evaluate 12 models on manipulation", "is_new": true}, {"name": "CaP-Gym", "purpose": "interactive environment", "is_new": true}], "results": [], "workflow": ["CaP-Gym environment", "CaP-Bench evaluation", "Agent improvement and RL"], "summary": "CaP-X provides gym, benchmark and improvement methods for code-as-policy coding agents in robot manipulation."}
|
|
{"id": 15123, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Script-generation framework for LLM multi-agent social simulation experiment design"}, {"area": "social-sim", "role": "application", "why": "Automates experimental design for social simulation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["experiment design pipeline"], "benchmarks": [], "results": [], "workflow": ["Script composition", "Script finalization", "Actor generation"], "summary": "FSTS automates multi-agent experiment design via script composition, finalization and actor generation."}
|
|
{"id": 9954, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Jointly optimizes role prompts across interacting agents without labels"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["system prompts"], "benchmarks": [], "results": [], "workflow": ["Joint prompt evaluation", "Data-driven refinement", "Iterate prompts"], "summary": "MASPO iteratively refines prompts across a multi-agent system using joint evaluation of downstream success."}
|
|
{"id": 7491, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "defense", "why": "Attention-guided test-time prompt tuning for adversarial robustness of CLIP"}], "pattern": "vary-scaffold", "fixes": ["adversarial attacks"], "varies": ["test-time adaptation method"], "benchmarks": [], "results": [], "workflow": ["Attention rollout", "Identify semantic regions", "Tune prompts"], "summary": "A-TPT uses refined attention rollout to preserve semantic regions during test-time prompt tuning against adversarial attacks on VLMs."}
|
|
{"id": 14226, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Multi-round RL training of small agentic models for tool use"}, {"area": "tool-use", "role": "data", "why": "Dual data flywheels synthesize increasingly hard tasks"}], "pattern": "vary-data", "fixes": ["RL recipe"], "varies": ["synthetic data", "behavior-tree tasks"], "benchmarks": [], "results": [], "workflow": ["Reasoning flywheel", "Agentic flywheel", "Multi-round RL"], "summary": "AgenticQwen small models trained with reasoning and agentic RL plus data flywheels generating harder tool-use tasks."}
|
|
{"id": 15878, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "DPO-trained memory copilot learns memory abstraction for frozen task model"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Hierarchical memory organization with frozen task model"}], "pattern": "vary-scaffold", "fixes": ["frozen task model"], "varies": ["memory abstraction policy"], "benchmarks": [], "results": [], "workflow": ["Frozen task model", "Copilot structures memory", "Hierarchical reuse"], "summary": "MCMA pairs a frozen task model with a DPO-trained memory copilot that structures and abstracts experience."}
|
|
{"id": 14879, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "method", "why": "Dual hierarchical RL dialogue policy for legal inquisitive agents"}], "pattern": "vary-scaffold", "fixes": ["Supreme Court dataset"], "varies": ["dialogue policy"], "benchmarks": [], "results": [], "workflow": ["Strategic dialogue policy", "Utterance generation policy"], "summary": "Dual hierarchical RL with two cooperating policies for proactive questioning in US Supreme Court oral arguments."}
|
|
{"id": 15046, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "RealMem benchmark of project-oriented cross-session memory interaction"}], "pattern": "fix-agent-vary-env", "fixes": ["memory systems"], "varies": ["project scenarios"], "benchmarks": [{"name": "RealMem", "purpose": "long-term project memory evaluation", "is_new": true}], "results": [], "workflow": ["Project foundation", "Multi-agent dialogue generation", "Memory and schedule management"], "summary": "RealMem is a benchmark of over 2,000 cross-session project dialogues evaluating long-term memory of LLM agents."}
|
|
{"id": 3596, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RLVR-based internalization of multimodal policies into model weights"}], "pattern": "vary-data", "fixes": ["policy prompts"], "varies": ["training method"], "benchmarks": [], "results": [], "workflow": ["Prepare policies", "Train with verifiable-reward RL"], "summary": "Internalizes complex multimodal conversational agent policies into the model, avoiding long in-context prompts."}
|
|
{"id": 18347, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "method", "why": "Graph RAG with linguistic knowledge graph for Latin word sense prediction"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["graph metadata"], "benchmarks": [], "results": [], "workflow": ["Retrieve KG context", "Prompt LLM"], "summary": "Combines a Latin linguistic knowledge graph with LLM prompting via graph RAG for sense prediction."}
|
|
{"id": 17328, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "application", "why": "Open-source toolkit for building agents, user simulation, dialog generation and evaluation"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "SDialog is a Python toolkit unifying persona-driven agent simulation, dialog generation, evaluation, interpretability and audio rendering."}
|
|
{"id": 1548, "in_agent_scope": false, "types": [{"area": "train/data", "role": "theory", "why": "Theoretical model of self-improvement dynamics via solver-verifier gap"}], "pattern": "vary-conditions", "fixes": [], "varies": ["LLMs", "datasets"], "benchmarks": [], "results": [], "workflow": [], "summary": "Models LLM self-improvement training dynamics through the solver-verifier gap and quantifies capability limits."}
|
|
{"id": 13855, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "attack", "why": "Bandit-guided LoRA experts adapt red-teaming to target models at inference"}], "pattern": "vary-conditions", "fixes": ["target LLMs"], "varies": ["attack style", "attack generator"], "benchmarks": [{"name": "AdvBench", "purpose": "attack success", "is_new": false}, {"name": "HarmBench", "purpose": "attack success", "is_new": false}], "results": [], "workflow": ["RL-train LoRA experts", "Bandit selects expert", "Probe target"], "summary": "Red-Bandit trains attack-style LoRA experts with RL and selects among them with a bandit at inference."}
|
|
{"id": 12233, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Lightweight memory construction with chain-of-memory utilization for agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory construction and use"], "benchmarks": [], "results": [], "workflow": ["Lightweight construction", "Retrieve fragments", "Organize into inference chains"], "summary": "CoM pairs lightweight memory construction with chained retrieval fragments into inference paths for LLM agents."}
|
|
{"id": 14790, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "SFT+RL teaches agents to decide when and what to summarize"}, {"area": "planning-decision", "role": "train-agent", "why": "Summarization as internal action in long-horizon decision making"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["context management policy"], "benchmarks": [], "results": [], "workflow": ["Cold-start SFT", "Summarization-aware RL"], "summary": "Self-Sum trains agents via SFT cold-start and RL to autonomously summarize context during long-horizon tasks."}
|
|
{"id": 11639, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "On-policy DPO with implicit rewards to reduce MLLM hallucination"}], "pattern": "vary-data", "fixes": ["DPO framework"], "varies": ["preference signal"], "benchmarks": [], "results": [], "workflow": ["SFT warm-up", "Sample responses", "Sift via implicit reward"], "summary": "IRIS uses implicit log-probability rewards to sift self-generated responses for on-policy preference alignment against MLLM hallucination."}
|
|
{"id": 15982, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "FinVQA multilingual Indic financial multimodal reasoning benchmark"}, {"area": "multiling/lowres", "role": "method", "why": "FIND SFT with constraint-aware decoding"}], "pattern": "vary-both", "fixes": [], "varies": ["benchmark", "method"], "benchmarks": [{"name": "FinVQA", "purpose": "financial multimodal reasoning", "is_new": true}], "results": [], "workflow": ["SFT", "Constraint-aware decoding"], "summary": "FinVQA benchmark of 18,900 Indic financial samples plus FIND, SFT with constraint-aware decoding."}
|
|
{"id": 1412, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Mode policy optimization for adaptive reasoning depth in social interaction"}], "pattern": "fix-env-vary-agent", "fixes": ["social tasks"], "varies": ["reasoning modes", "RL"], "benchmarks": [], "results": [], "workflow": ["Define reasoning modes", "Mode policy optimization"], "summary": "ASL trains language agents with adaptive mode policy optimization to adjust reasoning depth in negotiation and collaboration."}
|
|
{"id": 5435, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Self-play RL where agent injects and repairs bugs in repositories"}], "pattern": "fix-env-vary-agent", "fixes": ["sandboxed repos"], "varies": ["self-play training"], "benchmarks": [{"name": "SWE-bench Verified", "purpose": "self-improvement eval", "is_new": false}, {"name": "SWE-Bench Pro", "purpose": "self-improvement eval", "is_new": false}], "results": [], "workflow": ["Inject bug with tests", "Repair bug", "RL self-play"], "summary": "SSR trains a single LLM agent via self-play to inject and repair bugs, improving on SWE-bench."}
|
|
{"id": 18542, "in_agent_scope": false, "types": [{"area": "arch/longctx", "role": "method", "why": "Dual-channel decay KV-cache eviction for dialog inference"}], "pattern": "vary-scaffold", "fixes": ["transformer"], "varies": ["cache policy"], "benchmarks": [], "results": [], "workflow": ["Cumulative attention score", "Recency relevance score", "Evict by score"], "summary": "FD-KVC scores cached KV pairs with attention and recency channels to adapt to topic shifts in dialogs."}
|
|
{"id": 2140, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "OrchestrationBench for workflow planning and constraint-aware tool execution"}, {"area": "planning-decision", "role": "benchmark", "why": "Separates planning from tool execution evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["domains", "tools"], "benchmarks": [{"name": "OrchestrationBench", "purpose": "planning and tool use", "is_new": true}], "results": [], "workflow": [], "summary": "Multilingual benchmark across 17 domains evaluating workflow planning and constraint-aware tool use."}
|
|
{"id": 12803, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Identifies actor-observer asymmetry bias in role-playing agents via new benchmark"}, {"area": "multi-agent", "role": "train-agent", "why": "ReTAS dialectical alignment reduces the bias"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["perspective swap", "alignment training"], "benchmarks": [{"name": "Ambiguous Failure Benchmark", "purpose": "quantify AOA", "is_new": true}], "results": [], "workflow": ["Quantify AOA", "Dialectical alignment training"], "summary": "Shows perspective swapping induces actor-observer bias in agents and mitigates it with ReTAS alignment."}
|
|
{"id": 16744, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent conversational game recommender with risk control"}, {"area": "multi-agent", "role": "scaffold", "why": "Specialized agents for intent, retrieval, ranking, explanation"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["agent decomposition"], "benchmarks": [], "results": [], "workflow": ["Intent parsing", "Tool-augmented retrieval", "Ranking with reflection", "Risk control"], "summary": "MATCHA decomposes game conversational recommendation into specialized agents with risk control for safer, personalized results."}
|
|
{"id": 14751, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Iterative KG evidence subgraph refinement for medical QA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["KG retrieval and pruning"], "benchmarks": [], "results": [], "workflow": ["Retrieve focused KG", "Prune subgraph", "Answer"], "summary": "Iterative framework builds and prunes a concise evidence subgraph for graph-RAG medical QA."}
|
|
{"id": 17413, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "analysis", "why": "Evaluates classifiers mapping LLM free-text responses to questionnaire options"}], "pattern": "vary-conditions", "fixes": ["datasets"], "varies": ["classifier types"], "benchmarks": [], "results": [], "workflow": [], "summary": "Compares fine-tuned encoders and rule-based approaches for mapping generated responses to multiple-choice options."}
|
|
{"id": 14023, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Tree-structured memory module accessed via semantic XPath"}, {"area": "domain-agents", "role": "application", "why": "Demo conversational AI system with memory visualization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory structure"], "benchmarks": [], "results": [], "workflow": ["Tree-structured memory", "Query and update"], "summary": "Semantic XPath is a tree-structured conversational memory module improving over flat RAG with far fewer tokens."}
|
|
{"id": 1110, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Pipeline synthesizing computer-use tasks and trajectories"}, {"area": "web-gui", "role": "benchmark", "why": "Generated tasks measure agents across difficulty levels"}], "pattern": "vary-data", "fixes": ["agents"], "varies": ["task complexity"], "benchmarks": [{"name": "AgentSynth", "purpose": "agent difficulty scaling", "is_new": true}], "results": [{"benchmark": "AgentSynth", "system": "SOTA LLM agents", "base_model": null, "setting": "baseline", "metric": "success rate level 1", "score": 18, "unit": "%"}, {"benchmark": "AgentSynth", "system": "SOTA LLM agents", "base_model": null, "setting": "baseline", "metric": "success rate level 6", "score": 4, "unit": "%"}], "workflow": ["Generate subtasks", "Compose long-horizon tasks"], "summary": "AgentSynth synthesizes over 6,000 long-horizon computer-use tasks and trajectories at low cost with controllable difficulty."}
|
|
{"id": 12058, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Trajectory preference alignment for multi-agent tool-using QA"}, {"area": "search-research", "role": "scaffold", "why": "Six specialized agents integrate external knowledge"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["agent architecture", "preference learning"], "benchmarks": [], "results": [], "workflow": ["Specialized agents", "Preference alignment"], "summary": "AMATA aligns multi-agent tool-using trajectories via intra-trajectory preference learning and inter-agent harmonization for factual QA."}
|
|
{"id": 12907, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Asynchronous LLM-time series fusion for forecasting"}], "pattern": "vary-scaffold", "fixes": ["benchmark"], "varies": ["fusion strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "TimeSAF decouples unimodal feature learning from cross-modal fusion for LLM-guided time series forecasting."}
|
|
{"id": 15971, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "EvoDM dynamic incomplete-information data marketplace with equilibrium benchmarks"}, {"area": "social-sim", "role": "scaffold", "why": "Datamart-Agent with game-tree memory and reflection"}], "pattern": "vary-both", "fixes": ["LLM weights"], "varies": ["environment", "agent scaffold"], "benchmarks": [], "results": [], "workflow": ["Game tree memory", "Mechanism-guided reflection"], "summary": "EvoDM benchmarks LLM agent equilibrium decisions; Datamart-Agent improves them via game tree memory and reflection."}
|
|
{"id": 15430, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "LLM agent loop for personalized meal-level glucose regulation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["intervention pipeline"], "benchmarks": [], "results": [], "workflow": ["Predict glucose", "Adjust meal", "Iterate with feedback"], "summary": "A physio-feedback agentic loop combines a glucose predictor with dietary intervention to regulate postprandial glucose."}
|
|
{"id": 3842, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "theory", "why": "Risk-sensitive selection of agent compositions minimizing VaR and CVaR"}, {"area": "planning-decision", "role": "theory", "why": "Formal analysis over agent graphs"}], "pattern": "n/a", "fixes": [], "varies": ["assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Formalizes agent graphs and optimizes VaR/CVaR of safety-violation loss over feasible agent compositions."}
|
|
{"id": 2234, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "ENACT tests embodied cognition via egocentric forward and inverse world modeling"}], "pattern": "fix-agent-vary-env", "fixes": ["VLMs"], "varies": ["tasks"], "benchmarks": [{"name": "ENACT", "purpose": "embodied world modeling", "is_new": true}], "results": [], "workflow": [], "summary": "ENACT is a benchmark of 8,972 QA pairs probing VLM world modeling from egocentric interaction."}
|
|
{"id": 2634, "in_agent_scope": false, "types": [{"area": "arch/tokenization", "role": "method", "why": "Modality-specific tokenization for text-guided CAD generation and editing"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["tokenizer"], "benchmarks": [], "results": [], "workflow": [], "summary": "Multimodal tokenization aligned with CAD primitives improves LLM text-guided CAD prototyping."}
|
|
{"id": 17912, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "eval-method", "why": "SynTabQA synthetic QA generation from unannotated tables"}, {"area": "apps/classic", "role": "data", "why": "Synthetic QA for few-shot tabular training"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["QA generation"], "benchmarks": [{"name": "SynTabQA", "purpose": "synthetic tabular QA", "is_new": true}], "results": [], "workflow": [], "summary": "SynTabQA automatically generates typed question-answer pairs from any table for evaluation and few-shot enhancement."}
|
|
{"id": 4584, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Test-time in-context prompting and score refinement for 3D point cloud LLM"}], "pattern": "vary-scaffold", "fixes": ["pretrained encoder"], "varies": ["test-time guidance"], "benchmarks": [], "results": [], "workflow": ["Build similarity graph", "Retrieve neighbors", "LLM prompting"], "summary": "PGLLM retrieves graph neighbors' captions as demonstrations and refines scores at test time for 3D understanding."}
|
|
{"id": 12566, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "data", "why": "Citation data construction pipeline, benchmark and training for known/unknown-aware citations"}], "pattern": "vary-data", "fixes": [], "varies": ["training strategy", "data"], "benchmarks": [{"name": "KFC", "purpose": "citation QA", "is_new": true}], "results": [], "workflow": [], "summary": "KFC provides ~42K samples, benchmark and training strategy for fine-grained citations including unknown information."}
|
|
{"id": 14312, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Compares small models as base, tool agent, and routed multi-agent systems"}, {"area": "multi-agent", "role": "analysis", "why": "Deployment tradeoffs of routing multi-agent SLMs"}], "pattern": "vary-conditions", "fixes": ["SLMs <10B"], "varies": ["agent paradigm"], "benchmarks": [], "results": [], "workflow": [], "summary": "Large-scale study of sub-10B models under base, single tool agent, and routing multi-agent paradigms."}
|
|
{"id": 14808, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "data", "why": "Synthetic grammar-violation negative samples for low-resource MT training"}], "pattern": "vary-data", "fixes": ["baselines"], "varies": ["negative samples", "training loss"], "benchmarks": [], "results": [], "workflow": [], "summary": "NSL-MT penalizes grammatical violations to train MT for African low-resource languages with better data efficiency."}
|
|
{"id": 16479, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "FineState-Bench of fine-grained state-conditioned GUI control"}, {"area": "web-gui", "role": "eval-method", "why": "FineState-Metrics stage-wise diagnostic"}], "pattern": "fix-agent-vary-env", "fixes": ["LVLM agents"], "varies": ["UI components", "platforms"], "benchmarks": [{"name": "FineState-Bench", "purpose": "GUI state setting", "is_new": true}], "results": [], "workflow": [], "summary": "FineState-Bench has 2,209 instances across desktop, web and mobile with stage-wise diagnostics."}
|
|
{"id": 3470, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Agentic RL interleaving reasoning with code execution for math"}, {"area": "tool-use", "role": "data", "why": "Converts CoT into tool-augmented SFT trajectories"}], "pattern": "fix-env-vary-agent", "fixes": ["math benchmarks"], "varies": ["SFT data", "agentic RL"], "benchmarks": [], "results": [], "workflow": ["Synthesize trajectories", "SFT", "Agentic RL"], "summary": "AgentMath trains LRMs with synthesized tool trajectories and agentic RL interleaving code execution."}
|
|
{"id": 10338, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "survey", "why": "Interview and survey study of production agents"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies production agent deployments via 20 case studies and 86 surveys on methods, evaluation and challenges."}
|
|
{"id": 10760, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "UltraHorizon ultra long-horizon exploration benchmark"}, {"area": "memory-selfevolve", "role": "analysis", "why": "Probes sustained reasoning and memory"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["environments"], "benchmarks": [{"name": "UltraHorizon", "purpose": "long-horizon exploration", "is_new": true}], "results": [], "workflow": [], "summary": "UltraHorizon evaluates agents on long-horizon partially observable discovery tasks across three environments."}
|
|
{"id": 15271, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Hypothesis-validation multi-agent vulnerability detection"}, {"area": "agent-safety", "role": "application", "why": "Detects code vulnerabilities"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-agent pipeline"], "benchmarks": [], "results": [], "workflow": ["Multi-view analysis", "Form hypothesis", "Validate with context"], "summary": "VulAgent localizes sensitive operations, forms vulnerability hypotheses and validates them against project context."}
|
|
{"id": 11913, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "Answer-divergence-based instruction data selection"}], "pattern": "vary-data", "fixes": ["SFT recipe"], "varies": ["selected data"], "benchmarks": [], "results": [], "workflow": ["Sample generations", "Embed responses", "Score divergence"], "summary": "ADG selects instruction data by divergence of multi-sample outputs; 10K selected examples beat strong selectors."}
|
|
{"id": 9516, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Distills privileged-information teacher into unconditioned student for agentic RL"}, {"area": "tool-use", "role": "train-agent", "why": "Multi-turn agentic environments"}], "pattern": "fix-env-vary-agent", "fixes": ["agentic environments"], "varies": ["distillation method"], "benchmarks": [], "results": [], "workflow": ["Joint teacher-student", "Train with PI", "Transfer without PI"], "summary": "\u03c0-Distill jointly trains PI-conditioned teacher and student in one model to transfer privileged-information capability."}
|
|
{"id": 6284, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "SafeSearch automated red-teaming of search agents"}, {"area": "search-research", "role": "benchmark", "why": "300 test cases across five risk categories"}], "pattern": "fix-agent-vary-env", "fixes": ["search agent scaffolds"], "varies": ["unreliable search results"], "benchmarks": [{"name": "SafeSearch", "purpose": "search agent red-teaming", "is_new": true}], "results": [{"benchmark": "SafeSearch", "system": "GPT-4.1-mini", "base_model": null, "setting": "baseline", "metric": "ASR", "score": 90.5, "unit": "%"}], "workflow": ["Generate test cases", "Sandboxed evaluation"], "summary": "SafeSearch generates 300 test cases and finds high attack success rates on LLM search agents."}
|
|
{"id": 14892, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Graph-based hybrid self-evolving memory for computer-use agents"}, {"area": "web-gui", "role": "scaffold", "why": "Multi-hop retrieval and working-memory refresh"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["memory design"], "benchmarks": [], "results": [], "workflow": ["Graph memory", "Multi-hop retrieval", "Node updates"], "summary": "HyMEM couples symbolic nodes with trajectory embeddings in an evolving graph memory for computer-use agents."}
|
|
{"id": 13709, "in_agent_scope": true, "types": [{"area": "search-research", "role": "defense", "why": "Four-stage guardrails for deep research"}, {"area": "search-research", "role": "benchmark", "why": "DRSafeBench stage-wise safety benchmark"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["guardrails"], "benchmarks": [{"name": "DRSafeBench", "purpose": "stage-wise safety", "is_new": true}], "results": [], "workflow": [], "summary": "DeepResearchGuard adds multi-stage safeguards and open-domain evaluation; DRSafeBench measures stage-wise safety."}
|
|
{"id": 12257, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Dynamic routing for mixture-of-agents without pre-inference"}], "pattern": "vary-scaffold", "fixes": ["model pool"], "varies": ["routing"], "benchmarks": [], "results": [], "workflow": ["Scorer screens candidates", "Judges refine scores"], "summary": "RouteMoA uses a lightweight scorer and judge mixture to select models before inference, cutting MoA cost."}
|
|
{"id": 17476, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "application", "why": "Metadata-first multimodal image-text linking for e-commerce troubleshooting"}], "pattern": "vary-scaffold", "fixes": [], "varies": ["grounding method"], "benchmarks": [], "results": [], "workflow": ["Schema linking", "Metadata-conditioned generation", "Consistency evaluation"], "summary": "MIRAGE uses structured metadata for schema-guided image linking, generation and consistency evaluation."}
|
|
{"id": 13784, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "PolitNuggets long-tail political fact discovery benchmark"}, {"area": "search-research", "role": "eval-method", "why": "FactNet evidence-conditional scoring protocol"}], "pattern": "fix-agent-vary-env", "fixes": ["LRMs"], "varies": ["models", "settings"], "benchmarks": [{"name": "PolitNuggets", "purpose": "agentic fact discovery", "is_new": true}], "results": [], "workflow": [], "summary": "PolitNuggets benchmarks agentic discovery of political facts for 400 elites with FactNet scoring."}
|
|
{"id": 13053, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "SILO-BENCH role-free coordination under information silos"}], "pattern": "fix-agent-vary-env", "fixes": ["frontier LLMs"], "varies": ["protocols", "agent scales"], "benchmarks": [{"name": "SILO-BENCH", "purpose": "multi-agent coordination", "is_new": true}], "results": [], "workflow": [], "summary": "SILO-BENCH has 30 algorithmic tasks for distributed coordination, with 1,620 experiments over protocols and scales."}
|
|
{"id": 11661, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Textual gradient optimization of agent experience library"}], "pattern": "vary-scaffold", "fixes": ["frozen LLM"], "varies": ["memory"], "benchmarks": [], "results": [], "workflow": ["Reflect on failures", "Textual gradients", "Add/Edit/Delete", "Verify"], "summary": "TSGD refines an external experience library via failure-driven add, edit and delete operations under a capacity budget."}
|
|
{"id": 12720, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Orchestrated computer-use agent with reflection memory and tool agents"}, {"area": "multi-agent", "role": "scaffold", "why": "Orchestrator coordinating specialized agents"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["framework"], "benchmarks": [], "results": [], "workflow": ["Orchestrator", "Reflection memory", "Multimodal searcher"], "summary": "OS-Symphony coordinates a reflection-memory agent and tool agents with tutorial search for robust computer use."}
|
|
{"id": 13822, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "survey", "why": "Survey of LLM multi-agent systems for clinical workflows"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Surveys clinical multi-agent systems with a workflow-level taxonomy and evaluation stack."}
|
|
{"id": 13404, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "Bright-Pro for retrievers in agentic search"}, {"area": "search-research", "role": "eval-method", "why": "Evaluation of retrievers inside agent workflows"}], "pattern": "fix-agent-vary-env", "fixes": ["agentic search system"], "varies": ["retrievers"], "benchmarks": [{"name": "Bright-Pro", "purpose": "retriever evaluation", "is_new": true}, {"name": "Bright", "purpose": "prior benchmark", "is_new": false}], "results": [], "workflow": [], "summary": "Bright-Pro evaluates reasoning-intensive retrievers statically and within agentic search workflows."}
|
|
{"id": 14078, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent academic advising system with negotiation"}, {"area": "multi-agent", "role": "scaffold", "why": "Proposal-critique negotiation protocol"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent configuration"], "benchmarks": [{"name": "TartanBench", "purpose": "advising queries", "is_new": true}], "results": [], "workflow": ["Planning agent proposes", "Specialists critique", "Stream reasoning"], "summary": "TartanMaroon is a deployable multi-agent advising system with negotiation, transparency interface and TartanBench."}
|