Files
ai-taxonomy-literature-anal…/data/labels/o00.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
28 KiB
JSON

{"id": 12300, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Query-guided visual token pruning for long-video LLMs"}], "pattern": "vary-scaffold", "fixes": ["model", "benchmarks"], "varies": ["token pruning method"], "benchmarks": [{"name": "Video-MME", "purpose": "evaluate pruning Pareto", "is_new": false}, {"name": "MLVU", "purpose": "evaluate pruning Pareto", "is_new": false}], "results": [], "workflow": ["budget allocation", "semantic scout selection", "compensation"], "summary": "Decoupled query-guided visual token pruning for efficient long-video LLMs compatible with Flash Attention."}
{"id": 21, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "DeepSynth: 120 tasks requiring synthesis across sources"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "DeepSynth", "purpose": "deep information synthesis agents", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 120 deep information synthesis tasks across 7 domains for evaluating tool-using agents."}
{"id": 16412, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "survey", "why": "Survey of agent memory evolving from storage to experience"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Survey of LLM agent memory mechanisms framed as Storage, Reflection and Experience stages."}
{"id": 18482, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "method", "why": "Iteratively refines essay scoring rubrics by LLM reflection"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["rubric"], "benchmarks": [{"name": "ASAP", "purpose": "essay scoring", "is_new": false}, {"name": "ASAP 2.0", "purpose": "essay scoring", "is_new": false}, {"name": "TOEFL11", "purpose": "essay scoring", "is_new": false}], "results": [], "workflow": ["score essays", "reflect on mismatches", "revise rubric"], "summary": "Reflect-and-Revise iteratively refines essay scoring rubrics for LLMs using their own rationales and human-label discrepancies."}
{"id": 16677, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Red-teaming framework generating adversarial tests for black-box agents"}], "pattern": "fix-env-vary-agent", "fixes": ["target agent"], "varies": ["adversarial test cases"], "benchmarks": [], "results": [], "workflow": ["generate seed cases", "refine attacks", "distill red-teamer"], "summary": "SIRAJ red-teams black-box LLM agents with diverse seed tests, iterative attack refinement and distilled reasoning."}
{"id": 9336, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "theory", "why": "Stability analysis and protection for MARL with Byzantine nodes"}], "pattern": "vary-conditions", "fixes": [], "varies": ["Byzantine assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Stability analysis and protection design for decentralized MARL in service systems with Byzantine nodes."}
{"id": 15134, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Stage-aware framework for backdoor attacks across agent workflow"}], "pattern": "fix-env-vary-agent", "fixes": ["agent workflow"], "varies": ["backdoor trigger stage"], "benchmarks": [], "results": [], "workflow": [], "summary": "BackdoorAgent is a modular framework analyzing backdoor trigger injection and propagation across planning, memory and tool-use stages."}
{"id": 1929, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Multi-agent system for data visualization generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent workflow"], "benchmarks": [], "results": [], "workflow": ["metadata analysis", "task planning", "code generation", "reflection"], "summary": "CoDA multi-agent system for collaborative data visualization with metadata analysis, planning, coding and reflection."}
{"id": 15600, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Backdoor fine-tuned agents to exfiltrate data via tool calls"}], "pattern": "fix-env-vary-agent", "fixes": ["tool-use agent"], "varies": ["backdoor trigger"], "benchmarks": [], "results": [], "workflow": ["embed trigger", "memory-access call", "exfiltrate"], "summary": "Backdoor attack embedding semantic triggers in fine-tuned agents that exfiltrate user data through disguised tool calls."}
{"id": 2162, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "application", "why": "Tool-calling LLM for audio effect chains"}, {"area": "tool-use", "role": "data", "why": "LP-Fx instruction dataset with CoT and tool calls"}], "pattern": "vary-both", "fixes": ["general LLM"], "varies": ["audio fx pipeline"], "benchmarks": [{"name": "LP-Fx", "purpose": "audio effects tool-calling data", "is_new": true}], "results": [], "workflow": ["understand audio", "select effects", "estimate parameters"], "summary": "LLM2Fx-Tools uses tool calling and CoT to infer audio effect chains; introduces LP-Fx dataset."}
{"id": 8360, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Benchmark for code agents in data-intensive Kaggle sandbox"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "CoDA-Bench", "purpose": "code and data intelligence of agents", "is_new": true}], "results": [], "workflow": [], "summary": "CoDA-Bench jointly evaluates code and data intelligence of agents in a Kaggle-based data-intensive Linux sandbox."}
{"id": 12636, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Benchmark of HS code assignment with hierarchical rules"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "HSCodeComp", "purpose": "hierarchical rule application", "is_new": true}], "results": [], "workflow": [], "summary": "HSCodeComp benchmark of 632 products requiring agents to assign 10-digit HS codes via hierarchical tariff rules."}
{"id": 9414, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Regret-minimization iterative fine-tuning for decision-making LLMs"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["training procedure"], "benchmarks": [], "results": [], "workflow": ["sample trajectories", "select low-regret", "fine-tune"], "summary": "Iterative RMFT distills low-regret decision trajectories into the base LLM to improve online decision-making."}
{"id": 12422, "in_agent_scope": true, "types": [{"area": "search-research", "role": "application", "why": "Personalized deep research tool MyScholarQA"}, {"area": "search-research", "role": "eval-method", "why": "User-study evaluation showing synthetic benchmarks mislead"}], "pattern": "vary-evaluation", "fixes": ["tool"], "varies": ["evaluation protocol"], "benchmarks": [], "results": [], "workflow": ["infer user profile", "propose actions", "write report"], "summary": "MySQA personalized deep research tool; real-user evaluation contrasts with synthetic-user LLM-judge benchmarks."}
{"id": 12946, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "Multicultural simulated-town benchmark for LLM agents"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["cultural contexts"], "benchmarks": [{"name": "LiveCultureBench", "purpose": "cultural appropriateness of agents", "is_new": true}], "results": [], "workflow": [], "summary": "LiveCultureBench embeds LLM agents in a simulated town to evaluate task success and socio-cultural norm adherence."}
{"id": 9865, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Test-driven agent for parametric CAD assembly generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent design"], "benchmarks": [], "results": [], "workflow": ["synthesize code", "self-test", "refine"], "summary": "SPADA, a test-driven agent generating constrained parametric CAD assemblies as code."}
{"id": 6297, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Multiscale VLM features for reference-guided video editing"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["feature fusion"], "benchmarks": [], "results": [], "workflow": [], "summary": "MiVE uses multiscale hierarchical VLM-layer features for reference-guided video editing."}
{"id": 13335, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Fix credit assignment in multi-turn self-reflection RL"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL credit assignment"], "benchmarks": [], "results": [], "workflow": [], "summary": "Addresses Echo Trap in multi-turn self-reflection by improving credit assignment beyond trajectory-level GRPO rewards."}
{"id": 15831, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Multi-turn RL with similarity-based policy optimization for document QA agents"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["iterative retrieval", "vision-aware synthesis", "SPO training"], "summary": "MM-Doc-R1 trains agents for long-document VQA with SPO, a multi-turn RL baseline estimator."}
{"id": 3800, "in_agent_scope": true, "types": [{"area": "search-research", "role": "data", "why": "Formalization-driven synthesis of information-seeking training data"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": ["formalize task", "expand agentically"], "summary": "WebShaper synthesizes information-seeking agent data via formalization to align reasoning structure and questions."}
{"id": 14752, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RL-trained self-emergent linguistic toolchains for visual RAG"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training for tool use"], "benchmarks": [], "results": [], "workflow": ["emerge language actions", "reason with them"], "summary": "Lang2Act lets VLMs self-discover linguistic tool actions for fine-grained visual perception in visual RAG."}
{"id": 17372, "in_agent_scope": false, "types": [{"area": "safety/privacy", "role": "method", "why": "Context-aware PII detection for QA queries"}], "pattern": "vary-scaffold", "fixes": [], "varies": ["PII detector"], "benchmarks": [], "results": [], "workflow": [], "summary": "CAPID is a privacy-preserving approach to detect only contextually irrelevant PII in user questions."}
{"id": 12032, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Behavior calibration training for tool-integrated reasoning agents"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training framework"], "benchmarks": [], "results": [], "workflow": ["explore trajectories", "calibrate behavior", "train"], "summary": "ET-Agent calibrates redundant or insufficient tool-call behavior in tool-integrated reasoning agents via training."}
{"id": 18490, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "Role-playing agents' goal alignment under user distraction"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["distraction conditions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Evaluates whether role-playing LLM agents keep goals under goal-competing distraction using a user-simulator framework."}
{"id": 7221, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Trust-region fine-tuning for LLM team coordination"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["team training algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "TeamTR controls occupancy shift when fine-tuning heterogeneous LLM teams to stabilize multi-agent coordination."}
{"id": 6466, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Self-evolving computer use agent via experiential learning"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "Autonomous exploration on unfamiliar software"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["self-evolution training"], "benchmarks": [], "results": [], "workflow": ["explore software", "generate tasks", "learn from experience"], "summary": "SEAgent lets computer-use agents self-evolve on novel software through autonomous exploration and experiential learning."}
{"id": 305, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Perplexity-guided inference-time self-correction"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["inference intervention"], "benchmarks": [], "results": [], "workflow": ["monitor perplexity", "intervene"], "summary": "Once-More applies perplexity-guided intervention for continuous self-correction during generation."}
{"id": 7, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "method", "why": "Lifelong multi-turn jailbreak generation framework"}], "pattern": "vary-conditions", "fixes": ["target LLM"], "varies": ["attack strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "PLAGUE is a plug-and-play framework for adaptively generating multi-turn jailbreak exploits."}
{"id": 1607, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Label-free self-evolving VLM for image quality assessment"}], "pattern": "fix-env-vary-agent", "fixes": [], "varies": ["self-training via voting and ranking"], "benchmarks": [], "results": [], "workflow": ["generate pseudo-labels by voting", "rank", "train"], "summary": "EvoQuality self-improves VLM image quality perception without labels using voting and ranking."}
{"id": 14734, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Reasoning-augmented multimodal agent for recommendation"}], "pattern": "vary-scaffold", "fixes": [], "varies": ["reasoning pipeline"], "benchmarks": [], "results": [], "workflow": ["observe", "deliberate", "act"], "summary": "ReasonRec is a multimodal recommendation agent with Observe-Deliberate-Act explicit reasoning and uncertainty."}
{"id": 485, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "RL-trained Collaborator agent in fixed topology"}], "pattern": "fix-env-vary-agent", "fixes": ["multi-agent topology"], "varies": ["collaborator policy"], "benchmarks": [], "results": [], "workflow": [], "summary": "AgentPO trains a Collaborator agent with RL to improve collaboration in fixed multi-agent systems on math tasks."}
{"id": 18484, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "method", "why": "Probabilistic evaluation of turn-taking under ambiguity"}], "pattern": "vary-evaluation", "fixes": ["predictors"], "varies": ["metric"], "benchmarks": [], "results": [], "workflow": [], "summary": "Argues against binary turn-taking evaluation and evaluates predicted probabilities under inherent ambiguity."}
{"id": 11987, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "File-system-based dual-agent test-time scaling for deep research"}], "pattern": "vary-scaffold", "fixes": ["model"], "varies": ["workspace scaffold"], "benchmarks": [], "results": [], "workflow": ["context builder notes", "report writer"], "summary": "FS-Researcher uses persistent file-system workspace with dual agents to scale deep research beyond context limits."}
{"id": 15476, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Self-evolving healthcare agent for role-norm compliance"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Self-evolving healthcare LLM agent learns from role-based social experience with modeled communicator traits."}
{"id": 6834, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Importance-diversity trajectory step selection for web agents"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["selected data"], "benchmarks": [], "results": [], "workflow": ["score importance", "select diverse subset"], "summary": "Weasel selects fixed-budget trajectory steps balancing importance and diversity for better out-of-domain web agents."}
{"id": 11776, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Explicit warmth/competence trait inference for coordination"}], "pattern": "vary-scaffold", "fixes": ["model"], "varies": ["trait inference"], "benchmarks": [], "results": [], "workflow": ["infer partner traits", "guide decisions"], "summary": "ETI has LLM agents infer partner warmth and competence to improve multi-agent coordination."}
{"id": 5325, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Text-free speech-to-speech LLM"}], "pattern": "vary-scaffold", "fixes": [], "varies": ["architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "Speech-to-speech LLM without text guidance using layer-splitting architecture and frozen pretraining."}
{"id": 13552, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "Economic and trade competition benchmark for LLMs"}], "pattern": "fix-env-vary-agent", "fixes": ["LLMs under test"], "varies": ["economic tasks"], "benchmarks": [{"name": "Market-Bench", "purpose": "economic competition of LLM agents", "is_new": true}], "results": [], "workflow": [], "summary": "Market-Bench evaluates LLM agents in a configurable multi-agent supply chain with auctions and retail pricing."}
{"id": 2818, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Monitor red-teaming for detecting covert agent misbehavior"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["monitor design and awareness"], "benchmarks": [{"name": "SHADE-Arena", "purpose": "tool-calling monitoring", "is_new": false}, {"name": "CUA-SHADE-Arena", "purpose": "computer-use monitoring", "is_new": false}], "results": [], "workflow": ["red-team monitors", "hybrid monitor"], "summary": "Stress tests agent monitors via monitor red teaming and proposes a hybrid hierarchical-sequential monitor."}
{"id": 11556, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "method", "why": "RL fine-tuning steering generation away from classifier boundary"}], "pattern": "vary-scaffold", "fixes": ["classifier"], "varies": ["fine-tuning method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Boundary Guidance RL fine-tuning avoids classifier margins to cut false positives and negatives in filtered generation."}
{"id": 17462, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Entity extraction benchmark on phone-call transcripts"}, {"area": "train/data", "role": "method", "why": "Synthetic data pipeline with linguistic verbalization patterns"}], "pattern": "vary-data", "fixes": ["extractors"], "varies": ["synthetic training data"], "benchmarks": [{"name": "LingVarBench", "purpose": "phone-call entity extraction", "is_new": true}], "results": [], "workflow": ["sample entities", "apply verbalization patterns", "generate transcripts"], "summary": "LingVarBench: benchmark and synthetic data pipeline for entity extraction from phone-call transcripts."}
{"id": 12757, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Momentum uncertainty-guided training-free test-time reasoning budgeting"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["thinking budget allocation"], "benchmarks": [], "results": [], "workflow": ["track step uncertainty", "allocate budget"], "summary": "MUR allocates thinking budget adaptively using momentum of step uncertainty, reducing overthinking without training."}
{"id": 17247, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Sparsely grounded visual navigation task for MLLMs in cities"}], "pattern": "fix-env-vary-agent", "fixes": ["MLLMs under test"], "varies": ["tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "Introduces Sparsely Grounded Visual Navigation task to evaluate MLLM city navigation using web-scale knowledge."}
{"id": 7869, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "environment", "why": "Synthetic code-driven environment generation for agentic RL"}], "pattern": "fix-agent-vary-env", "fixes": ["RL algorithm"], "varies": ["environments"], "benchmarks": [], "results": [], "workflow": ["generate environments", "interact with tools", "RL train"], "summary": "Agent World Model synthesizes 1,000 code-driven environments for scalable agentic reinforcement learning."}
{"id": 6338, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Coordinating planning and execution when fine-tuning agents"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training objective"], "benchmarks": [], "results": [], "workflow": ["plan", "execute", "joint optimize"], "summary": "CoPE jointly optimizes long-term planning and multi-step execution when fine-tuning LLM agents."}
{"id": 16423, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "Benchmark for emotional-need-aware proactive memory retrieval"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "ENPMR-Bench", "purpose": "proactive emotional memory retrieval", "is_new": true}], "results": [], "workflow": [], "summary": "ENPMR-Bench evaluates emotional-support agents on inferring latent needs and proactively retrieving memory."}
{"id": 14690, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Open-domain tabular QA dataset for forecasting"}], "pattern": "fix-env-vary-agent", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "ODTQA-FoRe", "purpose": "tabular forecasting and reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "ODTQA-FoRe is a real-estate dataset for open-domain tabular QA with forecasting and forecast-based reasoning."}
{"id": 15503, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "RL for tool-using text-to-CAD agents"}, {"area": "domain-agents", "role": "environment", "why": "Agentic CAD framework with CAD engine tools"}], "pattern": "fix-env-vary-agent", "fixes": ["CAD engine"], "varies": ["agent training"], "benchmarks": [], "results": [], "workflow": ["tool-call CAD engine", "RL train"], "summary": "ToolCAD deploys LLMs as tool-using agents for text-to-CAD and trains them with reinforcement learning."}
{"id": 16747, "in_agent_scope": false, "types": [{"area": "train/pretrain", "role": "method", "why": "Efficient pretraining on historical corpora for semantic change"}], "pattern": "vary-data", "fixes": [], "varies": ["pretraining corpora"], "benchmarks": [], "results": [], "workflow": [], "summary": "Efficient pretraining on modest historical corpora yields models useful for diachronic semantic change discovery."}
{"id": 1709, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Training-free error-notebook retrieval of CAD assembly parts"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["inference-time adaptation"], "benchmarks": [], "results": [], "workflow": ["collect errors", "build notebook", "retrieve parts"], "summary": "Training-free inference-time framework using Error Notebooks for part retrieval in CAD assemblies with VLMs."}
{"id": 14215, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "E-commerce deep research benchmark with internal tools"}], "pattern": "fix-env-vary-agent", "fixes": ["DR systems under test"], "varies": ["tasks"], "benchmarks": [{"name": "DeepResearch Retail", "purpose": "tool-augmented deep research in e-commerce", "is_new": true}], "results": [], "workflow": [], "summary": "DeepResearch Retail evaluates deep research systems combining web search and internal tools on e-commerce data."}
{"id": 18521, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Streaming cough detection for conversational health agents"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": ["rolling audio buffer", "detect events"], "summary": "HealthCUES streams respiratory cough detection within spoken dialogue for healthcare conversational agents."}
{"id": 18366, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "analysis", "why": "Reddit corpus on food transition debate"}], "pattern": "vary-conditions", "fixes": [], "varies": ["communities"], "benchmarks": [{"name": "DRiFT", "purpose": "food transition Reddit corpus", "is_new": true}], "results": [], "workflow": [], "summary": "DRiFT corpus of 17.5M Reddit comments analyzed for language change in the protein transition debate."}
{"id": 10548, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Information self-locking failure in RL for active reasoning"}], "pattern": "vary-conditions", "fixes": ["RL algorithm"], "varies": ["feedback conditions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Identifies information self-locking where outcome-based RL agents fail to elicit and internalize evidence in active reasoning."}
{"id": 12044, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RL agentic MT with Wiktionary search for neologisms"}, {"area": "tool-use", "role": "data", "why": "Neologism translation dataset across 16 languages"}], "pattern": "fix-env-vary-agent", "fixes": ["search toolkit"], "varies": ["agent training"], "benchmarks": [], "results": [], "workflow": ["search Wiktionary", "translate", "RL train"], "summary": "NeoAMT agentic translation with Wiktionary search toolkit trained by RL on a 75-direction neologism dataset."}
{"id": 10504, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Attention-head-guided data selection and PEFT"}], "pattern": "vary-data", "fixes": ["benchmarks"], "varies": ["data and parameters selected"], "benchmarks": [], "results": [], "workflow": ["identify key heads", "select data", "tune"], "summary": "P2D links sparse key attention heads to data selection for efficient LLM adaptation."}
{"id": 10670, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Credit assignment from trajectories to prompt rules"}], "pattern": "vary-scaffold", "fixes": ["roles and topology", "model weights"], "varies": ["agent prompts"], "benchmarks": [], "results": [], "workflow": ["analyze trajectories", "assign credit", "revise prompt rules"], "summary": "Optimizes multi-agent prompts by assigning credit from interaction trajectories to individual prompt components."}
{"id": 4817, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Performance-aware tool selection for visual generation agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["tool performance awareness"], "benchmarks": [], "results": [], "workflow": [], "summary": "PerfGuard models tool performance boundaries to improve planning and execution in AIGC agents."}
{"id": 5746, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Process reward agents steering retrieval-grounded reasoning online"}], "pattern": "vary-scaffold", "fixes": ["reasoner"], "varies": ["step-level reward guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "Process reward agents provide retrieval-grounded step feedback during inference for knowledge-intensive reasoning."}
{"id": 17949, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Transcription guidelines for Australian child speech corpus"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": ["ASR draft", "manual correction"], "summary": "Workflow and guidelines for transcribing spontaneous Australian English child speech in AusKidTalk."}