Files
ai-taxonomy-literature-anal…/data/labels/o39.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
31 KiB
JSON

{"id": 16667, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Compares DSPy instruction optimizers across prompting techniques incl. ReAct/CodeAct for tabular fact verification"}], "pattern": "vary-conditions", "fixes": ["models", "benchmarks"], "varies": ["optimizer", "prompting technique", "model scale"], "benchmarks": [], "results": [], "workflow": [], "summary": "Systematic comparison of DSPy instruction optimizers over CoT, ReAct and CodeAct for tabular fact verification."}
{"id": 11790, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "data", "why": "Builds Arabic VQA dataset synthetically from Wikipedia templates, usable for post-training VLMs"}], "pattern": "vary-data", "fixes": ["VLMs", "training recipe"], "varies": ["training data"], "benchmarks": [{"name": "AraVQA", "purpose": "Arabic factoid VQA dataset for post-training", "is_new": true}], "results": [], "workflow": ["Extract Wikipedia template info", "LLM generates questions", "Build dataset"], "summary": "Pipeline using Wikipedia templates and an LLM to synthesize AraVQA, a 50K-question Arabic factoid VQA dataset."}
{"id": 15843, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "LLM copilot agent and text interface on DCS World for aviation flight tasks"}, {"area": "embodied-nav", "role": "environment", "why": "Textual DCS interface unifying cockpit data for LLMs"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["domain pipeline"], "benchmarks": [{"name": "Digital Combat Simulator", "purpose": "simulator environment for copilot", "is_new": false}], "results": [], "workflow": ["Textual DCS interface", "FalconAgent copilot"], "summary": "FalconCopilot: an LLM copilot agent plus semantic interface on DCS World for aviation autonomy."}
{"id": 13161, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL with judge-derived per-round contribution weights rescaling outcome advantages for search agents"}, {"area": "planning-decision", "role": "train-agent", "why": "Credit assignment via contribution-weighted GRPO"}], "pattern": "fix-env-vary-agent", "fixes": ["search environment"], "varies": ["RL reward/advantage weighting"], "benchmarks": [], "results": [], "workflow": ["LLM judge scores each search round", "Rescale outcome advantages", "GRPO update"], "summary": "CW-GRPO uses an LLM judge to weight per-round contributions, improving credit assignment when RL-training search agents."}
{"id": 14218, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent framework automating e-commerce SOPs, with web agent and document verification"}, {"area": "web-gui", "role": "application", "why": "WebAgent via demonstration-based learning for dynamic web interfaces"}], "pattern": "n/a", "fixes": ["general LLMs"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": ["SOP Groomer refines docs", "WebAgent executes", "Document verification agent"], "summary": "Agent-Ops: end-to-end multi-agent system automating e-commerce SOPs with a web agent and document verification."}
{"id": 4287, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Mobile GUI benchmark of 20K human demonstrations for proactive and personalized agents"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks", "user context"], "benchmarks": [{"name": "FingerTip 20K", "purpose": "proactive personalized mobile agent benchmark", "is_new": true}], "results": [], "workflow": [], "summary": "FingerTip 20K benchmark of Android demonstrations to evaluate proactive task suggestion and personalized execution by mobile agents."}
{"id": 6472, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "method", "why": "Introduces agentic confidence calibration with trajectory-level feature diagnostic"}, {"area": "planning-decision", "role": "analysis", "why": "Studies overconfidence in agent failures across trajectories"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["calibration method"], "benchmarks": [], "results": [], "workflow": ["Extract trajectory process features", "Interpretable model predicts confidence"], "summary": "Holistic Trajectory Calibration extracts process-level trajectory features to calibrate agent confidence."}
{"id": 9881, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Test-time MPC-style questioning control for visual design clarification"}], "pattern": "vary-scaffold", "fixes": ["LLM", "design tasks"], "varies": ["questioning strategy"], "benchmarks": [], "results": [], "workflow": ["Plan lookahead question sequence", "Surrogate reward for low burden", "Receding-horizon replanning"], "summary": "A-MPQC treats multi-round clarification as trajectory optimization to reduce user burden in visual design generation."}
{"id": 17916, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Large human preference dataset and rubrics for evaluating long-form QA"}], "pattern": "vary-evaluation", "fixes": ["LLM evaluators"], "varies": ["rubrics", "metrics"], "benchmarks": [{"name": "LFQA-HP-1M", "purpose": "human preference LFQA data", "is_new": true}], "results": [], "workflow": [], "summary": "LFQA-HP-1M: 1.3M pairwise preference annotations and nine rubrics for long-form QA evaluation, plus LLM-evaluator bias analysis."}
{"id": 17311, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Toolkit measuring linguistic diversity over dependency-parsed corpora"}], "pattern": "n/a", "fixes": [], "varies": ["diversity measures"], "benchmarks": [{"name": "Parallel Universal Dependencies", "purpose": "validate toolkit across languages", "is_new": false}], "results": [], "workflow": ["Query dependency trees", "Compute diversity"], "summary": "DELTA: open-source framework computing variety, balance and disparity of linguistic phenomena on CoNLL-U corpora."}
{"id": 4802, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "benchmark", "why": "Expert-annotated mental health decision dataset with fairness and ambiguity focus"}], "pattern": "fix-env-vary-agent", "fixes": ["LMs under test"], "varies": ["tasks"], "benchmarks": [{"name": "Mental health clinical dataset", "purpose": "fairness and ambiguity of LMs in psychiatry", "is_new": true}], "results": [], "workflow": [], "summary": "Clinician-annotated dataset spanning treatment, diagnosis, documentation, monitoring and triage to test LM fairness in mental healthcare."}
{"id": 6596, "in_agent_scope": false, "types": [{"area": "train/continual", "role": "method", "why": "Test-time learning selecting tokens by future uncertainty reduction"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["TTL update rule"], "benchmarks": [], "results": [], "workflow": ["Score tokens by future gain", "Update only selected tokens"], "summary": "FG-TTL updates LLMs at test time only on tokens that reduce subsequent uncertainty, avoiding collapse."}
{"id": 17339, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Agentic RAG system for classifying and allocating climate early warning investments from reports"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["domain pipeline"], "benchmarks": [{"name": "CREWS Fund corpus", "purpose": "evaluate EWS investment extraction", "is_new": false}], "results": [], "workflow": ["Hybrid retrieval", "CoT reasoning", "Classify and allocate budgets"], "summary": "Agent-based RAG system extracting and allocating early-warning-system climate finance from MDB reports."}
{"id": 9845, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "Dynamic benchmark of open Lean sorry tasks from real GitHub projects"}, {"area": "code-swe", "role": "analysis", "why": "Compares LLMs, agentic and symbolic provers on tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["provers under test"], "varies": ["tasks"], "benchmarks": [{"name": "SorryDB", "purpose": "dynamic Lean theorem completion benchmark", "is_new": true}], "results": [], "workflow": [], "summary": "SorryDB: continuously updating benchmark of real-world Lean tasks, evaluating LLMs, agents and symbolic provers."}
{"id": 7673, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "application", "why": "Multi-agent system validating and repairing repository-level code translation"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["agent architecture"], "benchmarks": [], "results": [], "workflow": ["Split equivalence validation into subtasks", "Agents analyze", "Repair translation"], "summary": "MatchFixAgent: language-agnostic multi-agent LLM framework for validating and repairing repository-level code translations."}
{"id": 1642, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Plug-and-play visual flow relay to reduce hallucination snowballing in VLM multi-agent systems"}, {"area": "multi-agent", "role": "analysis", "why": "Attention analysis of visual hallucination snowballing"}], "pattern": "vary-scaffold", "fixes": ["VLMs"], "varies": ["message relay"], "benchmarks": [], "results": [], "workflow": ["Select key vision tokens", "Relay visual flow between agents"], "summary": "ViF relays inter-agent messages with selected visual tokens to mitigate hallucination snowballing in VLM multi-agent systems."}
{"id": 4786, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Unified protocol converting heterogeneous agent datasets into one SFT format"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["data format and mixture"], "benchmarks": [], "results": [], "workflow": ["Convert datasets to ADP", "Unify for training", "SFT agents"], "summary": "Agent Data Protocol unifies 13 fragmented agent datasets for SFT across tool use, browsing, coding and SWE."}
{"id": 17627, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Large multilingual fact-checking dataset over high-volume OECD tables with synthetic claims"}], "pattern": "fix-agent-vary-env", "fixes": ["baseline system"], "varies": ["claims", "tables"], "benchmarks": [{"name": "OECD tabular claims dataset", "purpose": "fact-checking over large tables", "is_new": true}], "results": [], "workflow": ["Select data points via semantic frames", "Generate claims"], "summary": "Frame-guided synthetic claims dataset of 78,503 multilingual claims over large OECD tables."}
{"id": 17474, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "application", "why": "Legal QA system with vocabulary deconstruction, layout parsing and long-form answers"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": ["Deconstruct vocabulary", "Parse layouts", "Generate comprehensive answer"], "summary": "Long-context long-form QA system for legal documents handling vocabulary, layout and footnotes."}
{"id": 17453, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "application", "why": "MAPE data flywheel to improve deployed enterprise RAG assistant via fine-tuning"}], "pattern": "n/a", "fixes": ["deployed assistant"], "varies": ["fine-tuned router/rephraser"], "benchmarks": [], "results": [], "workflow": ["Monitor feedback", "Analyze failures", "Fine-tune components"], "summary": "Data flywheel for NVIDIA's enterprise assistant fine-tunes routing and rephrasal from user feedback."}
{"id": 17886, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free video QA via relevant event segmentation and background frame"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["video prompting"], "benchmarks": [], "results": [], "workflow": ["Segment relevant events", "Add background frame", "Prompt LLM"], "summary": "VideoEvent: lightweight training-free video QA selecting relevant temporal events for LLM prompts."}
{"id": 11847, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Iterative value refinement for value-guided decoding"}], "pattern": "n/a", "fixes": ["base policy"], "varies": ["value function training"], "benchmarks": [], "results": [], "workflow": ["Value exploration", "Self-refine value iteratively"], "summary": "Iterative Value Refinement improves value-guided decoding by training value functions on broader, iteratively generated data."}
{"id": 1868, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Prompt-guided pooling for unified short and long video LLM"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["pooling method"], "benchmarks": [], "results": [], "workflow": ["CLIP prompt alignment", "Prompt-guided pooling", "Compress visual tokens"], "summary": "PPLLaVA uses instruction-aware pooling to compress tokens for unified short and long video understanding."}
{"id": 4446, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Hierarchical MoE for continual vision-and-language navigation"}], "pattern": "fix-env-vary-agent", "fixes": ["navigation tasks"], "varies": ["model architecture"], "benchmarks": [], "results": [], "workflow": ["Macro router picks strategy", "Micro router aligns perception"], "summary": "M3E: dual-router mixture of experts to reduce forgetting in continual VLN."}
{"id": 11329, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Diagnostic benchmark of multi-turn clarification under input faults"}, {"area": "tool-use", "role": "analysis", "why": "Probes agent cooperative breakdowns in execution environments"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["input faults"], "benchmarks": [{"name": "DRIFT-BENCH", "purpose": "multi-turn clarification under input faults", "is_new": true}], "results": [], "workflow": [], "summary": "DRIFT-BENCH evaluates LLM agents on clarification under faulty user inputs using a persona user simulator."}
{"id": 14552, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Hierarchical quantized structure-aware retriever for long-document QA"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["retriever"], "benchmarks": [{"name": "QASPER", "purpose": "long-document QA evaluation", "is_new": false}, {"name": "Natural Questions", "purpose": "QA evaluation", "is_new": false}], "results": [], "workflow": ["Align hierarchical graph with vocabulary", "Hybrid scoring"], "summary": "HQDR integrates document structure into retrieval for long-document QA."}
{"id": 4761, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Centralized single-agent RL for ride-sharing order dispatch"}], "pattern": "n/a", "fixes": [], "varies": ["RL method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Triple-BERT: centralized single-agent RL for large-scale order dispatching, replacing MARL."}
{"id": 16774, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Quantifies and human-analyzes problem drift in multi-agent debate"}], "pattern": "vary-conditions", "fixes": ["debate method"], "varies": ["tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analysis of how multi-agent debate drifts from the initial problem across ten tasks, with expert-identified causes."}
{"id": 16267, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Model-agnostic agent for numerical data via code generation and multimodal reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent design"], "benchmarks": [], "results": [], "workflow": ["Generate code", "Multimodal reasoning"], "summary": "Agent analyzing sensor and time-series data through code generation and multimodal reasoning."}
{"id": 11838, "in_agent_scope": false, "types": [{"area": "reason/math", "role": "train-agent", "why": "Critic-guided RL for Lean formalization with CriticLeanGPT"}, {"area": "reason/math", "role": "benchmark", "why": "CriticLeanBench for judging formalization fidelity"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["critic training"], "benchmarks": [{"name": "CriticLeanBench", "purpose": "distinguish correct formalizations", "is_new": true}], "results": [], "workflow": ["Train critic via SFT and RL", "Guide formalization"], "summary": "CriticLean trains a critic with SFT and RL to guide math formalization, plus CriticLeanBench."}
{"id": 12935, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent scenario-based assessment of critical thinking"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["assessment method"], "benchmarks": [], "results": [], "workflow": ["Multi-agent dialogue", "Assessment CoT"], "summary": "MASA: interpretable multi-agent workflow assessing critical thinking from behavior, aligning with expert ratings."}
{"id": 10793, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Meta-evolves memory architecture jointly with experience"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["Evolve experience", "Evolve memory design"], "summary": "MemEvolve jointly evolves agent experience and the memory architecture itself."}
{"id": 3355, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "environment", "why": "Reasoning-based experience model synthesizes transitions for scalable agent RL"}, {"area": "planning-decision", "role": "train-agent", "why": "Online RL on synthetic experience"}], "pattern": "fix-agent-vary-env", "fixes": ["RL algorithm"], "varies": ["environment"], "benchmarks": [], "results": [], "workflow": ["Distill dynamics into experience model", "Synthesize rollouts", "RL train agent"], "summary": "DreamGym synthesizes diverse experiences via a reasoning model to enable scalable RL for agents."}
{"id": 16301, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Co-training for temporal entity alignment with richness-guided attention"}], "pattern": "n/a", "fixes": [], "varies": ["alignment model"], "benchmarks": [], "results": [], "workflow": ["Richness-guided attention", "Neighborhood consensus"], "summary": "RCTEA aligns entities across temporal knowledge graphs using richness-guided co-training."}
{"id": 9144, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Benchmark for DRC script coding agents with execution-based scoring"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "Rule2DRC", "purpose": "DRC script synthesis benchmark", "is_new": true}], "results": [], "workflow": [], "summary": "Rule2DRC: 1,000-task benchmark with 13,921 layouts for evaluating DRC coding agents by execution."}
{"id": 15992, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Training-free modality-aware agents with hybrid fusion for sensing"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent decomposition"], "benchmarks": [], "results": [], "workflow": ["Modality-specific agents", "Hybrid fusion"], "summary": "ConSensus decomposes multimodal sensing into specialized agents with semantic and statistical consensus."}
{"id": 16362, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Confidence-driven adaptive model selection in evolutionary agents"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["routing"], "benchmarks": [], "results": [], "workflow": ["Estimate solvability from confidence", "Route to LLM"], "summary": "AdaptEvolve selects LLMs per step using generation confidence to trade off efficiency and capability."}
{"id": 17094, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "Multi-agent LLM simulation of misinformation persuasion across demographics"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["demographics"], "benchmarks": [], "results": [], "workflow": [], "summary": "PANDORA studies misinformation persuasion among demographic LLM agents, finding susceptibility differences and echo chambers."}
{"id": 16951, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "analysis", "why": "Evaluates value of triple extraction for web QA with LLMs"}], "pattern": "vary-conditions", "fixes": ["benchmark"], "varies": ["LLMs", "augmentation"], "benchmarks": [], "results": [], "workflow": [], "summary": "Extends a benchmark with extraction annotations and tests whether triple extraction still helps LLM QA."}
{"id": 4773, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Scales Decision Pre-Trained Transformer for in-context RL with flow matching"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["training method"], "benchmarks": [], "results": [], "workflow": ["Train DPT with flow matching", "Multi-domain tasks"], "summary": "Extends DPT to hundreds of tasks using flow matching for scalable in-context RL."}
{"id": 12531, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "GRPO RL for proactive task scheduling with timing-aware rewards"}, {"area": "domain-agents", "role": "data", "why": "Automated annotation of opportunity time windows"}], "pattern": "vary-data", "fixes": ["environment"], "varies": ["reward design", "labels"], "benchmarks": [], "results": [], "workflow": ["Annotate time windows", "Define proactiveness metrics", "GRPO training"], "summary": "ProActor trains proactive scheduling agents with RL and window-based annotations."}
{"id": 5194, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "First investigation of opponent shaping among LLM agents via ShapeLLM"}, {"area": "social-sim", "role": "method", "why": "ShapeLLM adapts model-free opponent shaping to transformers"}], "pattern": "vary-conditions", "fixes": ["games"], "varies": ["agent interactions"], "benchmarks": [], "results": [], "workflow": [], "summary": "ShapeLLM tests whether LLM agents can shape other agents' behavior via interaction."}
{"id": 6025, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Simulated gig economy for LLM agents"}, {"area": "social-sim", "role": "analysis", "why": "Examines strategic capabilities in AI labor markets"}], "pattern": "vary-conditions", "fixes": ["simulator"], "varies": ["agent strategies"], "benchmarks": [{"name": "AI-Work", "purpose": "simulated gig economy", "is_new": true}], "results": [], "workflow": [], "summary": "AI-Work simulates a gig economy where LLM agents compete for jobs, studying strategic and economic dynamics."}
{"id": 13186, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "InquireBench for mobile agent safe inquiry"}, {"area": "web-gui", "role": "train-agent", "why": "Two-stage RFT teaching mobile agents to request human help"}], "pattern": "vary-both", "fixes": ["VLM base"], "varies": ["training", "benchmark"], "benchmarks": [{"name": "InquireBench", "purpose": "mobile agent proactive inquiry", "is_new": true}], "results": [], "workflow": ["Two-stage training", "RL fine-tuning"], "summary": "InquireMobile trains VLM mobile agents via RL to ask for human assistance, with InquireBench."}
{"id": 2989, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Object-centric automated evaluation framework for multi-turn image editing"}], "pattern": "vary-evaluation", "fixes": ["editing models"], "varies": ["metric"], "benchmarks": [], "results": [], "workflow": ["Decompose image into objects", "Synthesize instructions", "Evaluate"], "summary": "EdiVal-Agent: automated object-centric framework for fine-grained multi-turn image editing evaluation."}
{"id": 13691, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "LLM multi-agent simulation of exemplar-driven moral learning"}], "pattern": "vary-conditions", "fixes": ["agent loop"], "varies": ["games", "motivations"], "benchmarks": [], "results": [], "workflow": ["Plan-act-observe-reflect loop"], "summary": "Simulates how role models reshape collective morality among LLM agents across four games."}
{"id": 1624, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Autoregressive token-based traffic scenario generation for driving simulation"}], "pattern": "n/a", "fixes": [], "varies": ["generation model"], "benchmarks": [], "results": [], "workflow": ["Tokenize scene", "Autoregressive transformer"], "summary": "InfGen generates infinite traffic scenarios autoregressively for realistic closed-loop simulation."}
{"id": 12238, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Appraisal-guided reasoning with graph RL for emotion-cause triplets"}], "pattern": "n/a", "fixes": ["task"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["Appraisal-guided generation", "Graph-structured RL"], "summary": "ECFlow combines appraisal-guided traces and structure-aware RL for emotion-cause triplet extraction."}
{"id": 16905, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Two-stage SFT for GUI understanding and native reasoning with reflection"}, {"area": "web-gui", "role": "data", "why": "Synthesized data for hierarchical and reflection reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["GUI benchmarks"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["Stage 1 grounding SFT", "Stage 2 reasoning SFT"], "summary": "InfiGUIAgent trained via two-stage SFT with native reasoning and expectation-reflection."}
{"id": 17473, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent debate system for startup evaluation"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": ["Build question tree", "Debate for and against", "Score"], "summary": "DIALECTIC: multi-agent debate system producing startup investment arguments and scores."}
{"id": 16568, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Dual-agent adaptive reasoning over hierarchical knowledge graph for multi-hop QA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval and reasoning"], "benchmarks": [], "results": [], "workflow": ["Build HiGra", "Hybrid retrieval", "Seeker and Librarian"], "summary": "HiGraAgent unifies graph retrieval with dual-agent reasoning for multi-hop QA."}
{"id": 5591, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Full-duplex voice agent benchmark extending tau2-bench"}, {"area": "social-sim", "role": "environment", "why": "Realistic voice user simulator"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["voice vs text"], "benchmarks": [{"name": "tau-Voice", "purpose": "voice agent grounded tasks", "is_new": true}, {"name": "tau2-bench", "purpose": "extended base", "is_new": false}], "results": [], "workflow": [], "summary": "\u03c4-Voice benchmarks full-duplex voice agents with a realistic voice user simulator."}
{"id": 15026, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Agentic adversarial refinement for property-based test generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["framework"], "benchmarks": [], "results": [], "workflow": ["Generate property", "Validator makes counter-implementations", "Refine"], "summary": "PROBE hardens LLM-generated properties via adversarial counter-implementations."}
{"id": 9617, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Planner and Reflector with bidirectional experience pool for heuristic evolution"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["evolution framework"], "benchmarks": [], "results": [], "workflow": ["Planner schedules operators", "Reflector distills lessons"], "summary": "RefineEvo makes LLM heuristic design planning-guided and experience-driven."}
{"id": 11738, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "method", "why": "Range-coding provably secure linguistic steganography"}], "pattern": "n/a", "fixes": [], "varies": ["steganography method"], "benchmarks": [], "results": [], "workflow": ["Range coding", "Rotation mechanism"], "summary": "Efficient provably secure steganography using range coding with near-full entropy utilization."}
{"id": 4744, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "Interactive memory benchmark with simulated users"}, {"area": "memory-selfevolve", "role": "environment", "why": "Environment for on-policy memory evaluation and optimization"}], "pattern": "fix-agent-vary-env", "fixes": ["assistants"], "varies": ["environment"], "benchmarks": [{"name": "AMemGym", "purpose": "interactive memory benchmark", "is_new": true}], "results": [], "workflow": [], "summary": "AMemGym: interactive environment for on-policy evaluation of assistant memory."}
{"id": 12526, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Frozen LLM imaginative planning with world model"}], "pattern": "vary-scaffold", "fixes": ["frozen LLM"], "varies": ["world model planning"], "benchmarks": [], "results": [], "workflow": ["World model to symbolic states", "Predict outcomes", "Monte Carlo state estimation"], "summary": "IMPLEMENT lets frozen LLMs plan with a lightweight object-centric world model."}
{"id": 12488, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Adaptive pacing curriculum for DPO based on input complexity and output ambiguity"}, {"area": "align/pref", "role": "analysis", "why": "Analyzes asymmetric learning dynamics"}], "pattern": "n/a", "fixes": ["DPO"], "varies": ["training schedule"], "benchmarks": [{"name": "UltraFeedback", "purpose": "training data", "is_new": false}, {"name": "AlpacaEval 2.0", "purpose": "evaluate", "is_new": false}, {"name": "MT-Bench", "purpose": "evaluate", "is_new": false}], "results": [], "workflow": [], "summary": "DECOPO paces DPO training separately for input complexity and output ambiguity."}
{"id": 12975, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL aligning retrieval decisions with knowledge boundary"}], "pattern": "fix-env-vary-agent", "fixes": ["search tools"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": ["Stability metric", "Quadrant taxonomy", "Quadrant reward"], "summary": "KbPO calibrates retrieval decisions via knowledge boundaries in agentic search RL."}
{"id": 17423, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Agentic workflow converting pharmaceutical batch records to JSON"}], "pattern": "n/a", "fixes": ["models"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": ["Hybrid OCR", "VLM and LLM schema-guided", "Validate JSON"], "summary": "Agentic AI workflow converting paper batch records to validated JSON."}