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
28 KiB
JSON
61 lines
28 KiB
JSON
{"id": 17419, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "method", "why": "Active learning with transformer classifier to refine keyword-collected corpora for text analysis"}], "pattern": "vary-data", "fixes": ["text classifier"], "varies": ["labeling strategy"], "benchmarks": [{"name": "German newspaper corpus on depression and schizophrenia", "purpose": "evaluate corpus refinement", "is_new": false}], "results": [], "workflow": ["train classifier", "query uncertain snippets", "exclude irrelevant documents"], "summary": "Active learning iteratively refines text corpora, reaching F1 0.8 with 100-150 labels on German newspaper articles."}
|
|
{"id": 15683, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "Compares AI and human attributes in simulated negotiations and human-subject study"}], "pattern": "vary-conditions", "fixes": ["scenarios"], "varies": ["AI design traits", "human personality"], "benchmarks": [], "results": [], "workflow": [], "summary": "Compares effects of AI design and human personality in simulated and human-subject negotiation and transaction interactions."}
|
|
{"id": 16209, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Training-free budget guidance steering reasoning length with lightweight predictor"}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["decoding guidance"], "benchmarks": [], "results": [], "workflow": ["predict remaining length", "guide decoding to budget"], "summary": "Budget guidance steers reasoning LLMs toward a target thinking budget without fine-tuning using a lightweight length predictor."}
|
|
{"id": 7887, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Fisher-information-guided initialization of LoRA subspace"}], "pattern": "vary-both", "fixes": ["LoRA architecture"], "varies": ["initialization"], "benchmarks": [], "results": [], "workflow": ["estimate Fisher information", "select subspace", "initialize LoRA"], "summary": "Initializes LoRA in a task-relevant Fisher subspace rather than weight geometry to improve fine-tuning."}
|
|
{"id": 3071, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Open-source multi-turn RL framework for long-horizon agent training"}, {"area": "planning-decision", "role": "environment", "why": "Unified framework spanning diverse realistic environments"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["RL algorithm", "training"], "benchmarks": [], "results": [], "workflow": ["decoupled environment", "multi-turn rollout", "RL update"], "summary": "AgentGym-RL is a modular framework for training LLM agents from scratch via multi-turn RL across diverse environments."}
|
|
{"id": 2097, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Test-time adaptation of agents via environment interaction"}, {"area": "web-gui", "role": "train-agent", "why": "Adapts to unseen websites and function sets at test time"}], "pattern": "fix-env-vary-agent", "fixes": ["base LLM"], "varies": ["adaptation strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "Two strategies adapt LLM agents at test time using environment-specific information to fix syntactic and semantic mismatches."}
|
|
{"id": 16426, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Preference optimization for low-resource KGQA logical form generation"}], "pattern": "vary-both", "fixes": ["KGQA benchmarks"], "varies": ["training objective"], "benchmarks": [], "results": [], "workflow": ["generate logical forms", "preference optimize"], "summary": "Applies preference optimization to LLMs for generating logical forms in low-resource complex KGQA."}
|
|
{"id": 14805, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Test-time self-improvement via training on uncertain samples"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["test-time training"], "benchmarks": [], "results": [], "workflow": ["detect uncertain samples", "generate similar data", "test-time fine-tune"], "summary": "Test-time self-improvement lets agents identify uncertain samples, synthesize data and fine-tune themselves at test time."}
|
|
{"id": 14759, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "VeilEval privacy-preserving benchmark for psychological dialogues"}, {"area": "domain-agents", "role": "scaffold", "why": "EmoRes multi-agent framework with topic mining and self-reflection"}], "pattern": "vary-both", "fixes": ["LLMs"], "varies": ["benchmark", "framework"], "benchmarks": [{"name": "VeilEval", "purpose": "evaluate psychological dialogues", "is_new": true}], "results": [], "workflow": ["mine topics", "emotional agent responds", "self-reflect"], "summary": "Introduces VeilEval benchmark for psychological support dialogues and EmoRes multi-agent framework improving reasoning."}
|
|
{"id": 8001, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Minimal agentic baseline for theorem proving with refinement and library search"}], "pattern": "vary-scaffold", "fixes": ["frontier LLMs"], "varies": ["agent design choices"], "benchmarks": [], "results": [], "workflow": ["refine proofs iteratively", "search library", "manage context"], "summary": "Proposes a minimal agentic theorem-proving baseline with iterative refinement, library search and context management, competitive at lower cost."}
|
|
{"id": 10145, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Synthesizes GUI trajectories from unlabeled internet videos"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": ["filter videos", "extract trajectories", "pretrain GUI agent"], "summary": "Video2GUI extracts grounded GUI interaction trajectories from internet videos for large-scale GUI agent pretraining."}
|
|
{"id": 10342, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RL-aligned SLLM agent probing acoustic evidence tools"}, {"area": "domain-agents", "role": "application", "why": "Speech emotion recognition as multi-turn inquiry"}], "pattern": "fix-env-vary-agent", "fixes": ["tools"], "varies": ["agent training"], "benchmarks": [], "results": [], "workflow": ["maintain candidate set", "probe acoustic tools", "RL alignment"], "summary": "ADEPT turns a speech LLM into a tool-using agent for emotion recognition with RL and evidence probing."}
|
|
{"id": 6013, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Holistic optimization of LLM multi-agent collaboration"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["agent prompts", "collaboration structure"], "benchmarks": [], "results": [], "workflow": [], "summary": "OMAC is a framework to optimize LLM-based multi-agent collaboration holistically across agents and their interactions."}
|
|
{"id": 15493, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Training framework for goal-oriented API-call sequences"}, {"area": "tool-use", "role": "data", "why": "Synthesizes training data for goal-oriented tool use"}], "pattern": "fix-env-vary-agent", "fixes": ["tool benchmarks"], "varies": ["training data", "model"], "benchmarks": [], "results": [], "workflow": ["synthesize data", "fine-tune model"], "summary": "GOAT trains smaller open-source agents for goal-oriented tool use with synthesized data and fine-tuning."}
|
|
{"id": 16174, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "data", "why": "Chain-of-agents generation of counseling dialogues from static corpora"}], "pattern": "vary-data", "fixes": ["LLMs"], "varies": ["generation pipeline"], "benchmarks": [], "results": [], "workflow": ["profile client", "simulate stages", "safety monitor", "expert supervise"], "summary": "PsyChain uses collaborative agents modeling client personality and supervision to generate personalized professional counseling dialogues."}
|
|
{"id": 2471, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "Design-logic-guided multidisciplinary reasoning question synthesis"}], "pattern": "vary-data", "fixes": ["SFT recipe"], "varies": ["synthetic data"], "benchmarks": [], "results": [], "workflow": ["extract design logic", "generate questions", "SFT"], "summary": "DESIGNER synthesizes multidisciplinary reasoning questions from raw documents using reusable design logic for SFT."}
|
|
{"id": 14269, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "method", "why": "Aligns synthetic LLM choices with human data for econometric inference"}], "pattern": "vary-both", "fixes": ["estimators"], "varies": ["bias correction"], "benchmarks": [], "results": [], "workflow": ["generate synthetic choices", "learn correction", "econometric inference"], "summary": "AEM generates synthetic LLM choices, learns bias correction from small human samples, and recovers econometric estimates."}
|
|
{"id": 17979, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Prompt-based LLM agents in RL tasks with non-linguistic reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["prompt design"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies prompt-based LLM agents in RL tasks requiring structured non-linguistic reasoning."}
|
|
{"id": 12268, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Graph-accelerated hybrid LLM-agent and ABM social simulator"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["simulator design"], "benchmarks": [], "results": [], "workflow": ["graph memory", "hybrid agents", "parallel ABM"], "summary": "GASim speeds large-scale social simulation by combining LLM agents with ABM using graph-optimized memory."}
|
|
{"id": 12485, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Agent speculatively retrieves repo context during indexing for code completion"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval strategy"], "benchmarks": [], "results": [], "workflow": ["explore repo at indexing", "build speculative context", "complete code"], "summary": "SpecAgent explores repository files in advance to build speculative context, improving latency and completion quality."}
|
|
{"id": 4407, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "data", "why": "Human-guided AI curation of preference data for reward models"}], "pattern": "vary-data", "fixes": ["RM training"], "varies": ["preference data"], "benchmarks": [], "results": [], "workflow": ["AI annotate", "human guide", "curate"], "summary": "Curates preference datasets via human-guided AI feedback to train stronger reward models."}
|
|
{"id": 1049, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Online agentic RL with difficulty-adaptive GRPO for mobile GUI agents"}], "pattern": "fix-env-vary-agent", "fixes": ["mobile environments"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["sample tasks", "adaptive GRPO", "update agent"], "summary": "MobileRL trains mobile GUI agents with online RL using Difficulty-Adaptive GRPO."}
|
|
{"id": 14205, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "method", "why": "Personalized conversation starter generation for conversational agents"}], "pattern": "vary-both", "fixes": ["chat assistants"], "varies": ["starter generation"], "benchmarks": [], "results": [], "workflow": [], "summary": "Introduces conversation starter generation to break the first-message barrier with personalized suggestions."}
|
|
{"id": 16237, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Controlled study comparing single-agent and multi-agent semantic diversity"}], "pattern": "vary-conditions", "fixes": ["prompt conditioning"], "varies": ["single vs multi-agent"], "benchmarks": [], "results": [], "workflow": [], "summary": "Under matched prompts, single-agent generation beats multi-agent systems in semantic diversity."}
|
|
{"id": 5732, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Native omni-modal agent trained with RL for active perception in video"}], "pattern": "fix-env-vary-agent", "fixes": ["video benchmarks"], "varies": ["agent design", "RL"], "benchmarks": [], "results": [], "workflow": ["observe", "think", "act selectively"], "summary": "OmniAgent frames video understanding as a POMDP Observation-Thought-Action loop trained with RLVR."}
|
|
{"id": 12545, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "application", "why": "Cognitive-theory-driven student writing simulation"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["simulation framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "CPT-Agent simulates student writing development with bounded cognition and selective feedback comprehension."}
|
|
{"id": 5512, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Tournament-based relative ranking RL for open-ended agents"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["reward scheme"], "benchmarks": [], "results": [], "workflow": ["sample trajectories", "tournament ranking", "RL update"], "summary": "ArenaRL replaces pointwise reward scoring with tournament-based relative ranking for RL on open-ended agent tasks."}
|
|
{"id": 2349, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Curvature-guided policy optimization for multi-domain RL"}], "pattern": "vary-both", "fixes": ["domains"], "varies": ["optimizer"], "benchmarks": [], "results": [], "workflow": [], "summary": "CGPO mitigates multi-domain conflicts in RL for LLM reasoning using curvature-guided policy optimization."}
|
|
{"id": 12446, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "World-model-driven user agents learned from human interactions"}, {"area": "domain-agents", "role": "application", "why": "Recommender system evaluation with simulated users"}], "pattern": "fix-env-vary-agent", "fixes": ["recommenders"], "varies": ["user agent"], "benchmarks": [], "results": [], "workflow": ["world model next-state", "align agent", "simulate users"], "summary": "AlignUSER learns world-model-driven LLM user agents from human interactions to evaluate recommender systems."}
|
|
{"id": 18605, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "ResearchQA of survey-mined queries and rubrics across 75 fields"}], "pattern": "fix-agent-vary-env", "fixes": ["systems"], "varies": ["tasks"], "benchmarks": [{"name": "ResearchQA", "purpose": "evaluate scholarly QA", "is_new": true}], "results": [], "workflow": [], "summary": "ResearchQA offers 21K queries and 160K rubric items derived from surveys to evaluate scholarly long-form answers."}
|
|
{"id": 15311, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Curriculum learning for imitative novel generation"}], "pattern": "vary-both", "fixes": ["base LLM"], "varies": ["training curriculum"], "benchmarks": [], "results": [], "workflow": [], "summary": "Introduces imitative novel generation and a curriculum learning approach from style to story."}
|
|
{"id": 11298, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "MCP-Persona benchmark for personal applications via simulated environments"}, {"area": "tool-use", "role": "environment", "why": "Environment simulation of personal apps"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks", "environments"], "benchmarks": [{"name": "MCP-Persona", "purpose": "evaluate MCP agents", "is_new": true}], "results": [], "workflow": [], "summary": "MCP-Persona is the first benchmark for MCP agents on personal social applications using environment simulation."}
|
|
{"id": 12697, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Inference framework for high-arousal emotional TTS with prior and guidance rectification"}], "pattern": "vary-both", "fixes": ["TTS models"], "varies": ["inference procedure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Rectifies initialization prior and guidance for stable high-arousal emotional flow-matching text-to-speech."}
|
|
{"id": 12307, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "method", "why": "Reasoning trace inversion to improve abstention"}], "pattern": "vary-both", "fixes": ["reasoning models"], "varies": ["abstention method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Query Misalignment Framework and trace inversion methods improve abstention in reasoning LLMs."}
|
|
{"id": 17353, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Compares single vs multi-agent pipelines on visual geometry math"}], "pattern": "vary-scaffold", "fixes": ["MLLMs"], "varies": ["single vs multi-agent"], "benchmarks": [{"name": "Geometry3K", "purpose": "compare agent designs", "is_new": false}, {"name": "MathVerse", "purpose": "compare agent designs", "is_new": false}, {"name": "OlympiadBench", "purpose": "compare agent designs", "is_new": false}, {"name": "We-Math", "purpose": "compare agent designs", "is_new": false}], "results": [], "workflow": [], "summary": "Systematically compares single and multi-agent pipelines on four visual math benchmarks across models."}
|
|
{"id": 7064, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "AMA-Bench for long-horizon agent memory over trajectories"}], "pattern": "fix-agent-vary-env", "fixes": ["memory systems"], "varies": ["tasks"], "benchmarks": [{"name": "AMA-Bench", "purpose": "evaluate agent memory", "is_new": true}], "results": [], "workflow": [], "summary": "AMA-Bench evaluates long-horizon memory on real agent-environment trajectories rather than dialogue."}
|
|
{"id": 8640, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "WorldTravel multimodal travel planning with coupled constraints"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "WorldTravel", "purpose": "evaluate planning", "is_new": true}], "results": [], "workflow": [], "summary": "WorldTravel has 150 real travel scenarios across 5 cities with 15+ interdependent constraints."}
|
|
{"id": 10807, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "LLM-based decision model for offline multi-agent RL"}], "pattern": "vary-both", "fixes": ["offline datasets"], "varies": ["model"], "benchmarks": [], "results": [], "workflow": [], "summary": "DLM formulates offline MARL as dialogue-style sequence prediction with an LLM."}
|
|
{"id": 4536, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "method", "why": "Self-reflective creative ideation for system algorithm generation"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["ideation framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "MetaMuse uses self-reflection to make LLMs generate more creative, diverse system algorithms."}
|
|
{"id": 11193, "in_agent_scope": false, "types": [{"area": "arch/tokenization", "role": "method", "why": "Concept-guided image tokenization"}], "pattern": "vary-both", "fixes": ["generation models"], "varies": ["tokenizer"], "benchmarks": [], "results": [], "workflow": [], "summary": "Concept-guided tokenization adds semantic guidance to image tokenizers to bridge reconstruction and generation."}
|
|
{"id": 18535, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Agentic multi-turn dialogue for hazard identification"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["monolithic vs agentic"], "benchmarks": [], "results": [], "workflow": [], "summary": "HAZDIAL tests whether multi-agent multi-turn dialogue improves LLM hazard identification for operational safety."}
|
|
{"id": 12015, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "MAS-Bench for GUI-shortcut hybrid mobile agents"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "MAS-Bench", "purpose": "evaluate hybrid agents", "is_new": true}], "results": [], "workflow": [], "summary": "MAS-Bench evaluates GUI-shortcut hybrid mobile agents including autonomous shortcut generation."}
|
|
{"id": 8748, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "tau-Knowledge extends tau-Bench to unstructured knowledge"}, {"area": "search-research", "role": "benchmark", "why": "Retrieval plus tool use in conversations"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["environments"], "benchmarks": [{"name": "tau-Knowledge", "purpose": "evaluate agents", "is_new": true}, {"name": "tau-Bench", "purpose": "extended", "is_new": false}], "results": [], "workflow": [], "summary": "tau-Knowledge evaluates conversational agents that must retrieve unstructured knowledge and use tools."}
|
|
{"id": 1821, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "ManagerBench safety-pragmatism tradeoff in managerial scenarios"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["scenarios"], "benchmarks": [{"name": "ManagerBench", "purpose": "evaluate safety trade-off", "is_new": true}], "results": [], "workflow": [], "summary": "ManagerBench evaluates whether agents choose harmful actions when operational goals conflict with human safety."}
|
|
{"id": 17571, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Bilingual Latin-English QA benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "RespondeoQA", "purpose": "evaluate Latin QA", "is_new": true}], "results": [], "workflow": [], "summary": "RespondeoQA provides about 7,800 Latin-English QA and translation pairs."}
|
|
{"id": 3895, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "theory", "why": "Information-bottleneck analysis of compressor-predictor LM systems"}], "pattern": "vary-conditions", "fixes": ["predictor"], "varies": ["compressor size"], "benchmarks": [], "results": [], "workflow": [], "summary": "Frames compressor-predictor LM collaboration as information bottleneck and studies scaling with rate-distortion theory."}
|
|
{"id": 12299, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Agentic comparative reasoning with retrieval for legal judgment prediction"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["reasoning framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "GLARE lets models retrieve external knowledge and compare charges for legal judgment prediction."}
|
|
{"id": 17885, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "application", "why": "Framework for building accurate RAG QA applications"}], "pattern": "vary-both", "fixes": ["benchmarks"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "AccurateRAG provides a pipeline for data processing, fine-tuning, embedding and evaluation to build RAG QA systems."}
|
|
{"id": 1972, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "AgentFlow on-policy RL of modular planner in-the-flow"}, {"area": "planning-decision", "role": "scaffold", "why": "Specialized modules coordinating planning and tools"}], "pattern": "fix-env-vary-agent", "fixes": ["tool environments"], "varies": ["training method"], "benchmarks": [], "results": [], "workflow": ["planner", "executor", "verifier", "generator"], "summary": "AgentFlow trains a modular agentic system in-the-flow with on-policy RL for planning and tool use."}
|
|
{"id": 16319, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "SWE-Mutation evaluating LLM test suite generation"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "SWE-Mutation", "purpose": "evaluate test generation", "is_new": true}], "results": [], "workflow": [], "summary": "SWE-Mutation evaluates whether LLMs generate reliable software test suites."}
|
|
{"id": 3219, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Monitor-based implicit RAG with multi-agent refinement for scientific reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Eigen-1 combines token-level implicit retrieval with structured multi-agent refinement for scientific reasoning."}
|
|
{"id": 5256, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Entropy-guided dynamic tokens for graph-LLM molecular alignment"}], "pattern": "vary-both", "fixes": ["LLM backbone"], "varies": ["connector"], "benchmarks": [], "results": [], "workflow": [], "summary": "EDT-Former generates dynamic tokens aligning molecular graphs with frozen LLMs."}
|
|
{"id": 2036, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "GneissWeb 10T-token pretraining dataset with dedup and quality filters"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["data"], "benchmarks": [{"name": "GneissWeb", "purpose": "pretraining dataset", "is_new": true}], "results": [], "workflow": [], "summary": "GneissWeb is a roughly 10 trillion token dataset built via dedup and ensemble quality filters."}
|
|
{"id": 493, "in_agent_scope": true, "types": [{"area": "science-data", "role": "data", "why": "DataMind synthesizes data for data-analytic agents"}, {"area": "science-data", "role": "train-agent", "why": "Trains generalist data-analytic agents"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": [], "summary": "DataMind is a scalable data synthesis and training recipe for open-source generalist data-analytic agents."}
|
|
{"id": 16611, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "RECAP benchmark for intent rewriting in agentic planning"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "RECAP", "purpose": "evaluate intent rewriting", "is_new": true}], "results": [], "workflow": [], "summary": "RECAP is a benchmark for conversation intent rewriting to support agent planning."}
|
|
{"id": 3077, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Agent system for respiratory sound generation and diagnosis"}], "pattern": "vary-both", "fixes": ["models"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "Resp-Agent uses an adversarial curriculum agent to schedule synthetic data and improve respiratory disease diagnosis."}
|
|
{"id": 14754, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "LLM multi-agent Chinese court simulation with evaluation"}], "pattern": "vary-both", "fixes": ["LLMs"], "varies": ["simulation system"], "benchmarks": [], "results": [], "workflow": [], "summary": "Proposes an LLM-agent court simulation system with systematic procedure evaluation for legal training."}
|
|
{"id": 13760, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Collaborative memory for agentic recommender"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory design"], "benchmarks": [], "results": [], "workflow": [], "summary": "MemRec connects isolated user-item memories into collaborative memory for LLM agentic recommenders."}
|
|
{"id": 1348, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "analysis", "why": "Study of emergent risks in self-evolving agents"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["evolution pathways"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies misevolution, harmful deviations in self-evolving agents across four evolutionary pathways."}
|
|
{"id": 16492, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent generative engine optimization with reusable strategy skills"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Distills editing patterns into reusable skills"}], "pattern": "vary-scaffold", "fixes": ["generative engines"], "varies": ["strategy learning"], "benchmarks": [], "results": [], "workflow": [], "summary": "MAGEO coordinates planning, editing and evaluation agents and distills reusable engine-specific optimization skills."}
|