Files
ai-taxonomy-literature-anal…/data/labels/o31.jsonl
T
zzyang d6fc7c1dfa Add types-of-work classification, benchmarks and score pages
- 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
2026-10-21 11:55:46 +00:00

61 lines
34 KiB
JSON

{"id": 1018, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Transformer with spatio-temporal attention and interleaved recursion for offline multi-task MARL"}], "pattern": "n/a", "fixes": ["MARL benchmarks"], "varies": ["architecture"], "benchmarks": [], "results": [], "workflow": ["Spatio-temporal attention", "Interleaved recursive history modeling"], "summary": "Transformer architecture for offline multi-task multi-agent RL, classical MARL."}
{"id": 1634, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "AT-GRPO: agent- and turn-wise grouped RL for training multi-agent LLM systems"}, {"area": "planning-decision", "role": "train-agent", "why": "On-policy RL training system improves long-horizon planning accuracy"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks (game, plan, code, math)"], "varies": ["RL algorithm", "single vs multi-policy training"], "benchmarks": [{"name": "long-horizon planning tasks", "purpose": "show RL gains", "is_new": false}], "results": [{"benchmark": "long-horizon planning", "system": "AT-GRPO", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 99.5, "unit": "%"}], "workflow": ["Agent/turn-wise grouping", "On-policy MAS rollouts", "Update single or multiple policies"], "summary": "AT-GRPO adapts GRPO to multi-agent LLM systems with role- and turn-wise grouping and a training system, improving planning, coding and math."}
{"id": 16002, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Chinese real-speech benchmark for audio language models across instruction following, knowledge and robustness"}], "pattern": "fix-agent-vary-env", "fixes": ["audio LLMs under test"], "varies": ["tasks", "speech data"], "benchmarks": [{"name": "VCB Bench", "purpose": "evaluate LALMs on Chinese real speech", "is_new": true}], "results": [], "workflow": [], "summary": "VCB Bench is a Chinese benchmark of real human speech evaluating large audio language models on instruction following and knowledge."}
{"id": 17394, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "analysis", "why": "Computational linguistic profiling of Italian police interrogation transcripts"}], "pattern": "vary-conditions", "fixes": ["feature extraction"], "varies": ["transcript format"], "benchmarks": [], "results": [], "workflow": [], "summary": "Forensic linguistic analysis of narrative versus question-answer Italian police transcripts using 147 linguistic features."}
{"id": 2205, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Control-flow hijacking attacks that evade alignment-check defenses in multi-agent systems"}, {"area": "agent-safety", "role": "defense", "why": "Redesigned defense for multi-agent control-flow hijacking"}], "pattern": "vary-conditions", "fixes": ["multi-agent system"], "varies": ["attack", "defense mechanism"], "benchmarks": [{"name": "LlamaFirewall", "purpose": "defense being broken", "is_new": false}], "results": [], "workflow": ["Craft evading attacks", "Redesign defense"], "summary": "Shows control-flow hijacking attacks evade alignment-check defenses in multi-agent systems and proposes fixes."}
{"id": 11860, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "data", "why": "Bandit-guided chunk sampling to synthesize preference data for long-context optimization"}, {"area": "arch/longctx", "role": "method", "why": "Long-context preference optimization"}], "pattern": "vary-data", "fixes": ["preference optimization recipe"], "varies": ["data sampling strategy"], "benchmarks": [], "results": [], "workflow": ["MAB selects informative chunks", "Synthesize preference data", "Preference optimization"], "summary": "LongMab uses multi-armed bandit rollout to pick informative chunks for synthetic long-context preference data."}
{"id": 15509, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "MemPO: RL so agent autonomously summarizes and manages own memory"}, {"area": "planning-decision", "role": "train-agent", "why": "Long-horizon context management via policy optimization"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["memory policy", "RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Agent summarizes memory", "Policy optimization over memory"], "summary": "MemPO trains agents to self-manage memory during long-horizon interaction via policy optimization."}
{"id": 15696, "in_agent_scope": false, "types": [{"area": "multiling/multi", "role": "method", "why": "Debiased language-preference-guided query fusion for multilingual RAG"}, {"area": "knowledge/rag", "role": "analysis", "why": "Finds evaluation benchmark priors distort perceived English preference"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["query fusion"], "benchmarks": [], "results": [], "workflow": ["Debias language preference", "Fuse multilingual queries"], "summary": "Shows benchmark biases inflate English preference in multilingual RAG and proposes debiased query fusion."}
{"id": 5361, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Diagnoses template collapse in multi-turn agentic RL via mutual-information proxy"}, {"area": "planning-decision", "role": "train-agent", "why": "SNR-aware filtering prioritizes high-variance prompts"}], "pattern": "vary-conditions", "fixes": ["agentic RL tasks"], "varies": ["metric", "prompt filtering"], "benchmarks": [], "results": [], "workflow": ["MI proxy diagnoses collapse", "SNR-aware prompt filtering"], "summary": "Identifies template collapse in agentic RL, explains via signal-to-noise, and mitigates by filtering high-variance prompts."}
{"id": 12805, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Turn-level tree search with turn-wise credit assignment and turn-based policy optimization"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL objective", "exploration"], "benchmarks": [], "results": [], "workflow": ["Entropy-guided tree expansion", "Turn-wise credit assignment", "Turn-based policy update"], "summary": "AT2PO combines turn-level tree exploration, credit assignment, and turn-based policy optimization for multi-turn agentic RL."}
{"id": 9233, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "KV cache compression guided by beacon queries for reasoning models"}], "pattern": "n/a", "fixes": ["reasoning model"], "varies": ["KV compression"], "benchmarks": [], "results": [], "workflow": ["Detect thought-revisiting tokens", "Use beacon queries to score KV"], "summary": "BeaconKV compresses KV cache using beacon queries to better predict future attention in long reasoning."}
{"id": 4370, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Generalist GUI and embodied agent with layer-heterogeneous MoE"}, {"area": "embodied-nav", "role": "data", "why": "Unified action space and large-scale mixed GUI/embodied data"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["architecture", "training data"], "benchmarks": [], "results": [], "workflow": ["Layer-heterogeneous MoE", "Unify action spaces", "Train on mixed data"], "summary": "OmniActor handles GUI/embodied data conflict with layer-heterogeneous MoE and unified actions."}
{"id": 7926, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Scenario benchmark of 50 deception cases for human-AI interaction"}, {"area": "social-sim", "role": "environment", "why": "Multi-agent simulation with IntentNet and TrustNet for deception and trust"}], "pattern": "vary-both", "fixes": ["LLM agents"], "varies": ["scenarios", "detectors"], "benchmarks": [{"name": "OpenDeception", "purpose": "deception scenarios", "is_new": true}], "results": [], "workflow": ["Simulate deception dialogues", "IntentNet infers intent", "TrustNet estimates susceptibility"], "summary": "OpenDeception evaluates deception risk in human-AI dialogue with a 50-case benchmark, intent inference and user trust estimation."}
{"id": 11720, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Step-level filtering of noisy CUA rollouts into reliable trajectories"}, {"area": "web-gui", "role": "train-agent", "why": "SFT of Qwen2.5-VL on WebSTAR; step-level reward model"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["data filtering"], "benchmarks": [{"name": "WebVoyager", "purpose": "evaluate trained CUA", "is_new": false}], "results": [], "workflow": ["Synthesize rollouts from CUA", "Step-level filtering", "Reasoning augmentation", "SFT"], "summary": "WebSTAR synthesizes 13.3K computer-use trajectories with step-level filtering, enabling a 7B SFT model beating UI-TARS-1.5-7B."}
{"id": 11149, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "GRPO-based RL for egocentric interaction reasoning and pixel grounding in MLLMs"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["RL method"], "benchmarks": [], "results": [], "workflow": ["Analysis-guided reasoning", "GRPO training"], "summary": "EARL uses analysis-guided GRPO to improve MLLM egocentric interaction understanding and grounding."}
{"id": 9440, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Multi-step online RL of GUI-Cursor with cursor visual feedback"}, {"area": "web-gui", "role": "method", "why": "Reframes GUI grounding as interactive cursor search"}], "pattern": "fix-env-vary-agent", "fixes": ["base models", "benchmarks"], "varies": ["grounding formulation", "RL reward"], "benchmarks": [], "results": [], "workflow": ["Move cursor iteratively", "Evaluate spatial relation", "Online RL dense reward"], "summary": "GUI-Cursor reframes GUI grounding as interactive cursor movement trained with multi-step RL."}
{"id": 15994, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "LLM agent builds knowledge graph memory incrementally"}, {"area": "planning-decision", "role": "scaffold", "why": "Exploration in partially observable environments"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["Build knowledge graph from interaction", "Use graph for decisions"], "summary": "GraphMind gives LLM agents a dynamic knowledge-graph memory for sequential decisions in partially observable environments."}
{"id": 13738, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "MLCL diagnostic benchmark for multilingual tool calling"}, {"area": "tool-use", "role": "analysis", "why": "Error analysis of multilingual tool-calling failures"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["language"], "benchmarks": [{"name": "MLCL", "purpose": "multilingual tool calling diagnosis", "is_new": true}], "results": [], "workflow": [], "summary": "MLCL benchmarks tool calling across Chinese, Hindi and Igbo and analyzes errors despite correct intent."}
{"id": 18646, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "data", "why": "ToxicTags meme dataset with two-stage annotation"}, {"area": "arch/multimodal", "role": "method", "why": "Entropy-guided multi-task learning for toxic meme detection"}], "pattern": "n/a", "fixes": ["-"], "varies": ["dataset", "model"], "benchmarks": [{"name": "ToxicTags", "purpose": "toxic meme detection", "is_new": true}], "results": [], "workflow": ["Collaborative tags", "Entropy-guided multi-task learning"], "summary": "Introduces ToxicTags meme dataset and an entropy-guided multi-task model for fine-grained toxic meme detection."}
{"id": 4805, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "analysis", "why": "Shows MLLM generation lags understanding"}, {"area": "arch/multimodal", "role": "method", "why": "Internal-gap self-improvement framework"}], "pattern": "n/a", "fixes": ["MLLMs"], "varies": ["self-improvement"], "benchmarks": [], "results": [], "workflow": ["Evaluate gap", "Understanding guides generation"], "summary": "Reveals understanding-generation gap in unified MLLMs and proposes self-improvement using stronger understanding."}
{"id": 15031, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "attack", "why": "Shapley value-guided prompt stealing for text-to-music models"}], "pattern": "fix-agent-vary-env", "fixes": ["text-to-music model"], "varies": ["attack queries"], "benchmarks": [], "results": [], "workflow": ["Shapley attribution", "Query search to reconstruct prompt"], "summary": "AudioStealer reconstructs hidden prompts of text-to-music models via Shapley-guided query search."}
{"id": 13098, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "MHGPO: critic-free heterogeneous-group RL for multi-agent search systems"}, {"area": "search-research", "role": "train-agent", "why": "End-to-end training of LLM search agents"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["RL algorithm", "rollout sampling"], "benchmarks": [], "results": [], "workflow": ["Heterogeneous group rollouts", "Relative advantage estimation", "Policy update"], "summary": "MHGPO trains multi-agent LLM search systems end-to-end with group-relative advantages across heterogeneous rollouts, without critics."}
{"id": 7737, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "theory", "why": "Markov potential game alignment guarantee for agent-human oversight"}, {"area": "agent-safety", "role": "defense", "why": "Oversight interface balancing autonomy and deference"}], "pattern": "n/a", "fixes": [], "varies": ["assumptions", "oversight policy"], "benchmarks": [], "results": [], "workflow": ["Model as Markov game", "Learn ask/play, trust/oversee"], "summary": "Models agent-human oversight as a Markov potential game with alignment guarantee for balancing safety and autonomy."}
{"id": 15779, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "FedGUI benchmark for federated GUI agents across platforms, devices, OS"}, {"area": "web-gui", "role": "environment", "why": "Six curated datasets for heterogeneity"}], "pattern": "fix-agent-vary-env", "fixes": ["FL methods"], "varies": ["platform heterogeneity"], "benchmarks": [{"name": "FedGUI", "purpose": "federated GUI agent training", "is_new": true}], "results": [], "workflow": [], "summary": "FedGUI benchmarks federated training of GUI agents under cross-platform, device and OS heterogeneity."}
{"id": 10108, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Zero-shot VLM active mapping with 360-BEV and waypoint generation"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["representation", "waypoint strategy"], "benchmarks": [], "results": [], "workflow": ["Build 360-BEV", "Generate candidate waypoints", "VLM selects"], "summary": "Zero-shot VLM-based active mapping using fused 360-BEV representation and waypoint generation."}
{"id": 4831, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Entropy-guided flexible rank allocation for LoRA"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["rank allocation"], "benchmarks": [], "results": [], "workflow": ["Entropy metric per matrix", "Allocate ranks"], "summary": "FlexLoRA allocates LoRA ranks with matrix-wise entropy-guided metrics."}
{"id": 17557, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "benchmark", "why": "Annotated Arabic hope speech dataset from YouTube comments on Gaza"}], "pattern": "n/a", "fixes": [], "varies": ["dataset"], "benchmarks": [{"name": "AraHopeCorpus", "purpose": "Arabic hope speech annotation", "is_new": true}], "results": [], "workflow": [], "summary": "AraHopeCorpus is an annotated dataset of Arabic hope speech from ten thousand YouTube comments."}
{"id": 4600, "in_agent_scope": true, "types": [{"area": "search-research", "role": "eval-method", "why": "SeekBench process-level evaluation of epistemic competence"}, {"area": "search-research", "role": "analysis", "why": "Reveals behavioral gaps in RL-tuned search agents"}], "pattern": "vary-evaluation", "fixes": ["search agents"], "varies": ["metrics"], "benchmarks": [{"name": "SeekBench", "purpose": "process-level search agent evaluation", "is_new": true}], "results": [], "workflow": ["Annotate traces", "LLM-as-judge pipeline"], "summary": "SeekBench evaluates search agents' groundedness, recovery and calibration beyond answer accuracy."}
{"id": 1928, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Safiron pre-execution guardrail for agent plans"}, {"area": "agent-safety", "role": "data", "why": "AuraGen synthetic risky trajectories"}, {"area": "agent-safety", "role": "benchmark", "why": "Pre-Exec Bench for plan-level guardrails"}], "pattern": "vary-both", "fixes": ["agent planners"], "varies": ["guardrail", "data", "benchmark"], "benchmarks": [{"name": "Pre-Exec Bench", "purpose": "pre-execution guardrail evaluation", "is_new": true}], "results": [], "workflow": ["Synthesize trajectories", "Inject risks", "Train guardian"], "summary": "Builds data engine, guardian model and benchmark for pre-execution safety of agent plans."}
{"id": 3584, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "eval-method", "why": "User-aware agent evaluation with automated error analysis"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["evaluation framework"], "benchmarks": [], "results": [], "workflow": ["Simulate users", "Evaluate", "Diagnose errors"], "summary": "Unified user-aware agent evaluation framework with automated error diagnosis."}
{"id": 12623, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "method", "why": "Theory-of-mind neuro-symbolic logical reasoning via multi-agent collaboration"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["reasoning framework"], "benchmarks": [], "results": [], "workflow": ["Bidirectional reasoning", "Multi-agent collaboration"], "summary": "Leibniz uses theory-of-mind multi-agent collaboration for neuro-symbolic logical reasoning."}
{"id": 14663, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "PerMemSafe benchmark of implicit personalized safety for self-evolving agents"}, {"area": "memory-selfevolve", "role": "benchmark", "why": "Long-horizon memory-based personalization risks"}], "pattern": "fix-agent-vary-env", "fixes": ["self-evolving agents"], "varies": ["conversation contexts"], "benchmarks": [{"name": "PerMemSafe", "purpose": "implicit personalized safety", "is_new": true}], "results": [], "workflow": [], "summary": "PerMemSafe is the first benchmark of implicit personalized safety for memory-accumulating agents."}
{"id": 2342, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Feed-forward guided hierarchical Gaussian on-the-fly 3D reconstruction"}], "pattern": "n/a", "fixes": [], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["3D foundation model pose", "Hierarchical Gaussians"], "summary": "ARTDECO combines feed-forward models with SLAM-style pipeline for on-the-fly 3D reconstruction."}
{"id": 5634, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL with structural-proximity dense reward and context management for multimodal search agents"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["reward", "context"], "benchmarks": [], "results": [], "workflow": ["Dense advantage from structure", "Reduce redundant context"], "summary": "DR-MMSearchAgent addresses interaction collapse in multimodal search agents."}
{"id": 13901, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "Persona-conditioned LLMs in power-asymmetric dialogues"}, {"area": "agent-safety", "role": "analysis", "why": "Compliance with unsafe requests by status"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["persona status"], "benchmarks": [], "results": [], "workflow": [], "summary": "Examines whether LLMs with high or low status personas show human socio-cognitive effects."}
{"id": 12406, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "attack", "why": "FCV-Attack induces functionally correct yet vulnerable patches"}, {"area": "agent-safety", "role": "attack", "why": "Vulnerability across agent-model combinations"}], "pattern": "vary-conditions", "fixes": ["code agents"], "varies": ["adversarial input"], "benchmarks": [{"name": "SWE-Bench", "purpose": "attack evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "FCV-Attack makes code agents produce patches passing tests yet vulnerable across 12 agent-model combinations."}
{"id": 7023, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "data", "why": "Soft Verified Generation synthesizes 200K+ trajectories"}, {"area": "code-swe", "role": "train-agent", "why": "SFT of open coding agents specialized to repositories"}], "pattern": "vary-data", "fixes": ["SFT"], "varies": ["synthetic data"], "benchmarks": [], "results": [], "workflow": ["Generate trajectories from repo", "Soft verify", "SFT"], "summary": "SERA trains repository-specialized coding agents cheaply with SFT on synthetic soft-verified trajectories."}
{"id": 9042, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "analysis", "why": "Systematic study of replay buffers for LLM post-training"}, {"area": "align/rlvr", "role": "method", "why": "Experience replay reduces compute"}], "pattern": "vary-conditions", "fixes": ["RL algorithm"], "varies": ["replay buffer design"], "benchmarks": [], "results": [], "workflow": ["Store rollouts", "Reuse in training"], "summary": "Shows replay buffers cut inference compute in LLM RL post-training without hurting performance."}
{"id": 12021, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "RL post-training with GNN-based structural rewards for social graph simulation"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["rewards"], "benchmarks": [], "results": [], "workflow": ["Graph supervision", "GNN rewards", "RL"], "summary": "Graphia trains LLM agents via RL on social graph data to simulate interaction behavior."}
{"id": 14310, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "analysis", "why": "Compares sentiment approaches for B2B support CSAT prediction"}], "pattern": "vary-conditions", "fixes": ["CSAT data"], "varies": ["approaches"], "benchmarks": [], "results": [], "workflow": [], "summary": "Five-year case study evaluating 17 sentiment approaches against CSAT ground truth."}
{"id": 12869, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Factorized memory with working context and knowledge graph for deep research"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "RFT crystallizes context into snapshots"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["Fluid working context", "Crystallize snapshots", "Knowledge graph"], "summary": "Cognitive Scaffold factorizes memory for long-horizon deep research agents."}
{"id": 13701, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "analysis", "why": "Effect of Kohlberg moral stage prompts on LLM problem solving"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["moral prompts"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies how moral-condition prompts influence LLM general problem solving."}
{"id": 10722, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Agentic VLM with dynamic cognitive map and RL for spatial reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["VLM"], "varies": ["pipeline", "RL reward"], "benchmarks": [], "results": [], "workflow": ["Build cognitive map", "Active exploration", "RL"], "summary": "Agentic VLM actively explores with dynamic cognitive map for spatial reasoning, reinforced with RL."}
{"id": 12722, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "BrowseComp-Plus fixed-corpus deep search benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["corpus", "retriever"], "benchmarks": [{"name": "BrowseComp-Plus", "purpose": "fair deep search eval", "is_new": true}, {"name": "BrowseComp", "purpose": "source of queries", "is_new": false}], "results": [{"benchmark": "BrowseComp-Plus", "system": "Search-R1 (BM25)", "base_model": null, "setting": "baseline", "metric": "accuracy", "score": 3.86, "unit": "%"}, {"benchmark": "BrowseComp-Plus", "system": "GPT-5 (BM25)", "base_model": null, "setting": "baseline", "metric": "accuracy", "score": 55.9, "unit": "%"}, {"benchmark": "BrowseComp-Plus", "system": "GPT-5 + Qwen3-Embedding-8B", "base_model": null, "setting": "baseline", "metric": "accuracy", "score": 70.1, "unit": "%"}], "workflow": [], "summary": "BrowseComp-Plus provides a fixed human-verified corpus to evaluate deep search agents and retrievers separately."}
{"id": 16785, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Stepwise proficiency-controlled sentence simplification with path planning"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["prompting"], "benchmarks": [], "results": [], "workflow": ["Plan path", "Select exemplars", "CoT generation"], "summary": "Multilingual unsupervised proficiency-controlled simplification via stepwise decomposition."}
{"id": 718, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Offline latent world model imagination with uncertainty-aware branch scoring"}, {"area": "embodied-nav", "role": "scaffold", "why": "World-model use to reduce real interactions"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["planning"], "benchmarks": [{"name": "WebShop", "purpose": "reduced API calls", "is_new": false}], "results": [], "workflow": ["Imagine futures", "Score with uncertainty", "Distill reflection"], "summary": "DreamPhase plans via latent world model imagination without modifying the LLM, reducing real interactions."}
{"id": 12537, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Tree-structured subtask exploration with page action memory"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["scaffold"], "benchmarks": [{"name": "WebArena", "purpose": "web agent evaluation", "is_new": false}], "results": [{"benchmark": "WebArena", "system": "Branch-and-Browse", "base_model": null, "setting": "proposed", "metric": "task success rate", "score": 35.8, "unit": "%"}], "workflow": ["Subtask tree", "State replay", "Action memory"], "summary": "Branch-and-Browse web agent with tree exploration and action memory."}
{"id": 4501, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "SupervisorAgent runtime supervision reducing tokens"}], "pattern": "vary-scaffold", "fixes": ["base agents"], "varies": ["supervision"], "benchmarks": [{"name": "GAIA", "purpose": "token efficiency", "is_new": false}], "results": [], "workflow": ["Context filter triggers", "Correct errors", "Purify observations"], "summary": "SupervisorAgent supervises multi-agent systems at runtime, cutting token use without lowering success."}
{"id": 13119, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Structured multi-step LLM agent for entity alignment"}, {"area": "domain-agents", "role": "application", "why": "Knowledge graph fusion"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["reasoning structure"], "benchmarks": [], "results": [], "workflow": ["Multi-step reasoning", "Align entities"], "summary": "EA-Agent applies structured multi-step reasoning to knowledge graph entity alignment."}
{"id": 14334, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "application", "why": "Multi-agent voice phishing simulation"}, {"area": "agent-safety", "role": "environment", "why": "Simulated phishing with tactic mining"}], "pattern": "n/a", "fixes": [], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": ["Main agent orchestrates", "Dialogue generation", "Tactic search"], "summary": "VishBox v2 multi-agent system generating adaptive voice phishing simulations."}
{"id": 463, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Chain-of-LoRA video-language agent with role-specific LoRAs"}, {"area": "planning-decision", "role": "scaffold", "why": "Planner, grounder, verifier roles"}], "pattern": "fix-env-vary-agent", "fixes": ["backbone"], "varies": ["agent design"], "benchmarks": [], "results": [], "workflow": ["Planner", "Grounder", "Verifier", "Answerer via LoRA"], "summary": "VideoMind is a chain-of-LoRA agent for temporally grounded video reasoning."}
{"id": 15715, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "eval-method", "why": "Interactive evaluation with simulated patient and Information Coverage Rate"}, {"area": "domain-agents", "role": "analysis", "why": "Evidence elicitation of agents"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["metric"], "benchmarks": [], "results": [], "workflow": ["Simulated patient", "Atomic evidence measure"], "summary": "Interactive evaluation of evidence elicitation in diagnosis consultation using ICR."}
{"id": 13650, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Characterizes Tool-Induced Myopia in code-interpreter use"}, {"area": "tool-use", "role": "benchmark", "why": "PyMath of 1,679 competition problems"}], "pattern": "vary-conditions", "fixes": ["TaLMs"], "varies": ["tool access"], "benchmarks": [{"name": "PyMath", "purpose": "tool-induced reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "Shows tool-augmented LMs treat tool outputs as substitutes for reasoning, studied on PyMath."}
{"id": 1223, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "MARTI framework for multi-agent reinforced training and inference"}, {"area": "multi-agent", "role": "environment", "why": "Asynchronous multi-turn rollout infrastructure"}], "pattern": "fix-env-vary-agent", "fixes": ["math tasks"], "varies": ["multi-agent training"], "benchmarks": [], "results": [], "workflow": ["Multi-agent interaction", "Distributed training", "Rewards"], "summary": "MARTI open-source framework for scalable RL of multi-agent LLM systems."}
{"id": 985, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Agentic LLM framework automating vehicle routing problems end to end"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["framework"], "benchmarks": [], "results": [], "workflow": ["Interpret intent", "Design algorithm", "Execute"], "summary": "AFL automates solving complex vehicle routing problems using LLM agents."}
{"id": 3894, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "MCP Security Bench evaluating attacks across tool-use pipeline"}, {"area": "tool-use", "role": "attack", "why": "12 MCP attack types"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["attacks"], "benchmarks": [{"name": "MCP Security Bench", "purpose": "MCP attack resilience", "is_new": true}], "results": [], "workflow": [], "summary": "MSB evaluates nine LLM agents against 12 MCP attack types with a robustness metric."}
{"id": 1990, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "MCTS multi-agent mining of mobile GUI trajectories"}, {"area": "web-gui", "role": "train-agent", "why": "Fine-tuning GUI agents on mined data"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["data mining"], "benchmarks": [], "results": [], "workflow": ["MCTS exploration", "Multi-agent guidance", "Intent recycling"], "summary": "M2-Miner automatically mines mobile GUI trajectories via multi-agent MCTS."}
{"id": 14671, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "data", "why": "CFlowPsyD counseling dialogue dataset"}, {"area": "apps/dialogue", "role": "method", "why": "Self-optimizing multi-agent dialogue generation"}], "pattern": "vary-data", "fixes": [], "varies": ["data generation"], "benchmarks": [{"name": "CFlowPsyD", "purpose": "counseling dialogues", "is_new": true}], "results": [], "workflow": ["Seed data", "Multi-agent generation", "Self-optimize"], "summary": "CFlowPsy generates asynchronous counseling dialogues, producing CFlowPsyD."}
{"id": 10043, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "ALIGN gossip network for reputation among LLM agents"}, {"area": "social-sim", "role": "environment", "why": "Indirect reciprocity setting"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["gossip mechanism"], "benchmarks": [], "results": [], "workflow": ["Share gossip", "Form reputations", "Coordinate norms"], "summary": "ALIGN lets self-interested LLM agents sustain indirect reciprocity through gossip."}
{"id": 6555, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "environment", "why": "CVE-Factory converts CVE metadata to executable tasks"}, {"area": "code-swe", "role": "benchmark", "why": "LiveCVEBench with 190 tasks"}, {"area": "agent-safety", "role": "train-agent", "why": "Fine-tuned Qwen3-32B"}], "pattern": "vary-both", "fixes": ["pipeline"], "varies": ["tasks", "environments"], "benchmarks": [{"name": "LiveCVEBench", "purpose": "vulnerability tasks", "is_new": true}, {"name": "Terminal Bench", "purpose": "generalization", "is_new": false}], "results": [{"benchmark": "LiveCVEBench", "system": "Qwen3-32B finetuned", "base_model": "Qwen3-32B", "setting": "finetuned", "metric": "success", "score": 35.8, "unit": "%"}, {"benchmark": "LiveCVEBench", "system": "Qwen3-32B", "base_model": "Qwen3-32B", "setting": "baseline", "metric": "success", "score": 5.3, "unit": "%"}, {"benchmark": "Terminal Bench", "system": "Qwen3-32B finetuned", "base_model": "Qwen3-32B", "setting": "finetuned", "metric": "success", "score": 31.3, "unit": "%"}, {"benchmark": "Terminal Bench", "system": "Qwen3-32B", "base_model": "Qwen3-32B", "setting": "baseline", "metric": "success", "score": 12.5, "unit": "%"}], "workflow": ["Multi-agent CVE reproduction", "Validate environment"], "summary": "CVE-Factory automatically builds executable vulnerability tasks, enabling a benchmark and training environments."}