Files
ai-taxonomy-literature-anal…/data/labels/o43.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
29 KiB
JSON

{"id": 15817, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Flattens multi-level table headers and uses pseudo-code reasoning for table QA"}], "pattern": "vary-scaffold", "fixes": ["LLM", "table QA data"], "varies": ["header flattening", "reasoning format"], "benchmarks": [], "results": [], "workflow": ["Flatten headers", "Global understanding", "Pseudo-code reasoning"], "summary": "SMART decouples table structure understanding from reasoning for complex table question answering."}
{"id": 8174, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Decouples answer authority to reduce evidence misalignment in long-video agents"}, {"area": "search-research", "role": "analysis", "why": "Diagnoses temporal and semantic groundedness failures of video agents"}], "pattern": "vary-conditions", "fixes": ["video QA task"], "varies": ["answer authority", "training reward"], "benchmarks": [], "results": [], "workflow": ["Diagnose misalignment", "Decouple answer authority", "Train agent"], "summary": "VideoSEAL mitigates evidence misalignment in agentic long video understanding by decoupling answer authority."}
{"id": 482, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Multi-agent RL that counters lazy-agent behavior with deliberation"}, {"area": "multi-agent", "role": "theory", "why": "Theoretical analysis of why lazy agents arise"}], "pattern": "fix-env-vary-agent", "fixes": ["reasoning tasks"], "varies": ["multi-agent RL objective"], "benchmarks": [], "results": [], "workflow": ["Analyze lazy behavior", "Modify RL training", "Encourage deliberation"], "summary": "Explains and fixes lazy agent behavior in multi-agent LLM reasoning trained with RL."}
{"id": 9279, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Zero-shot 4D human-object interaction generation with part affordance graphs"}], "pattern": "n/a", "fixes": [], "varies": ["generation pipeline"], "benchmarks": [], "results": [], "workflow": ["Decompose object parts", "Generate reference videos", "Synthesize 4D HOI"], "summary": "HOI-PAGE generates 4D human-object interactions from text via part affordance reasoning."}
{"id": 15187, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Appraisal-guided CoT reasoning and retrieval for emotional support"}], "pattern": "n/a", "fixes": ["MLLM"], "varies": ["reasoning and retrieval method"], "benchmarks": [], "results": [], "workflow": ["Generate appraisal chains", "Verify with dual signals", "Retrieve"], "summary": "AG-CTR2 improves multimodal emotional support conversations with appraisal-guided reasoning and retrieval."}
{"id": 6932, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "analysis", "why": "Ablates components of ML-engineering multi-agent systems in 4000+ experiments"}, {"area": "code-swe", "role": "benchmark", "why": "Introduces K-live, a benchmark of active competitions"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["architectural components"], "benchmarks": [{"name": "K-live", "purpose": "contamination-free ML engineering evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Controlled ablation of multi-agent ML engineering components plus a new live-competition benchmark."}
{"id": 6985, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Rollout augmentation synthesizes self-correction supervision for VLM RL"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL rollouts", "training stages"], "benchmarks": [], "results": [], "workflow": ["Recombine rollouts", "Two-stage RL"], "summary": "Octopus learns self-correction in VLMs via rollout augmentation and two-stage RL."}
{"id": 13417, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "train-agent", "why": "RL with evidence-aware rewards and self-correcting preference for radiology reports"}], "pattern": "fix-env-vary-agent", "fixes": ["RRG task"], "varies": ["reward", "preference learning"], "benchmarks": [], "results": [], "workflow": ["Evidence-aware reward", "Self-correcting preference learning"], "summary": "ESC-RL improves radiology report generation with evidence-aware rewards and preference learning."}
{"id": 15234, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Privacy-preserving LLM agent selecting analytic functions over glucose data"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": ["LLM selects functions", "Local computation", "Answer"], "summary": "CGM-Agent answers questions over continuous glucose data while keeping raw data private."}
{"id": 15143, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Process critic with explorative demonstrations for multi-turn agentic RL"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["RL exploration", "process critic"], "benchmarks": [], "results": [], "workflow": ["Critic scores steps", "Inject explorative demonstrations", "RL update"], "summary": "ProCeedRL improves exploration in agentic RL with a process critic and demonstrations."}
{"id": 16743, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Benchmark of utility and vulnerability of planner-executor multi-agent systems"}, {"area": "agent-safety", "role": "attack", "why": "Evaluates adversarial manipulation of planner-executor MAS"}], "pattern": "fix-agent-vary-env", "fixes": ["MAS architectures"], "varies": ["attack surfaces", "tasks"], "benchmarks": [{"name": "PEAR", "purpose": "planner-executor MAS robustness", "is_new": true}], "results": [], "workflow": [], "summary": "PEAR benchmarks utility and vulnerability of planner-executor LLM multi-agent systems."}
{"id": 11609, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Compositional pretraining and entropy-guided decoding for TCR-pMHC modeling"}], "pattern": "n/a", "fixes": [], "varies": ["pretraining curriculum", "decoding"], "benchmarks": [], "results": [], "workflow": ["Continual pretraining curriculum", "Entropy-guided iterative decoding"], "summary": "DecoderTCR models TCR-pMHC interactions with compositional pretraining and entropy-guided decoding."}
{"id": 11980, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Trains history-aware routers on synthesized trajectories for MCP tool routing"}, {"area": "tool-use", "role": "data", "why": "Synthesizes multi-turn trajectories from dependency-rich candidate graph"}], "pattern": "fix-env-vary-agent", "fixes": ["MCP benchmarks"], "varies": ["router training data"], "benchmarks": [], "results": [], "workflow": ["Build candidate graph", "Synthesize trajectories", "Train router"], "summary": "ACE-Router trains plug-and-play history-aware routers for large MCP tool ecosystems."}
{"id": 7761, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Agents reflect on failures to learn cooperation rules without weight updates"}, {"area": "multi-agent", "role": "scaffold", "why": "Aligns partners and tasks in embodied multi-agent settings"}], "pattern": "vary-scaffold", "fixes": ["LLM", "environment"], "varies": ["reflection-based cooperation laws"], "benchmarks": [], "results": [], "workflow": ["Reflect on failures", "Extract misaligned patterns", "Learn cooperation laws"], "summary": "LLawCo learns cooperation laws from reflection for embodied multi-agent behavior."}
{"id": 15263, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Two-tier reward with trajectory aggregation for agentic Text-to-SQL RL"}], "pattern": "fix-env-vary-agent", "fixes": ["Text-to-SQL tasks"], "varies": ["reward design"], "benchmarks": [], "results": [], "workflow": ["Aggregated trajectory reward", "Column-set matching step reward", "RL"], "summary": "SQL-ASTRA densifies sparse feedback for multi-turn agentic Text-to-SQL RL."}
{"id": 18517, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Question-tree guidance for interview dialogue topic control"}], "pattern": "n/a", "fixes": [], "varies": ["question selection method"], "benchmarks": [], "results": [], "workflow": ["Generate candidates", "Append to tree", "Select next question"], "summary": "Question Tree steers interview dialogue toward target information with topic continuation or switching."}
{"id": 15166, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Multi-agent sample mining and pseudo-label self-evolution for relevance models"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "Self-evolving relevance model from query streams"}], "pattern": "vary-data", "fixes": ["industrial search"], "varies": ["samples and pseudo-labels"], "benchmarks": [], "results": [], "workflow": ["Mine informative samples", "Reliable pseudo-labeling", "Update model"], "summary": "SERM self-evolves a search relevance model using multi-agent sample mining and labeling."}
{"id": 12417, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "survey", "why": "Taxonomy and survey of LLM agents for legal tasks"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Surveys LLM agents in law, giving taxonomy, applications and challenges."}
{"id": 5032, "in_agent_scope": false, "types": [{"area": "safety/privacy", "role": "attack", "why": "RL-guided reconstruction attack on machine unlearning"}], "pattern": "fix-agent-vary-env", "fixes": ["unlearned model"], "varies": ["attack method"], "benchmarks": [], "results": [], "workflow": ["Use traces as rewards", "Generator explores input space", "Recover forgotten data"], "summary": "ReTrace recovers unlearned data by treating residual traces as RL rewards."}
{"id": 11859, "in_agent_scope": true, "types": [{"area": "search-research", "role": "analysis", "why": "Systematic study of context interference in multi-turn search agents"}, {"area": "search-research", "role": "scaffold", "why": "Mitigates interference for reliability and efficiency"}], "pattern": "vary-conditions", "fixes": ["search agents"], "varies": ["context content", "context management"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies and mitigates context interference in multi-turn search agents."}
{"id": 17833, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "Persian-English medical QA dataset benchmarking 41 models"}, {"area": "multiling/lowres", "role": "benchmark", "why": "Low-resource Persian evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["models under test"], "varies": ["language", "specialty"], "benchmarks": [{"name": "PersianMedQA", "purpose": "Persian medical MCQ evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "PersianMedQA is a 20,785-question bilingual medical QA benchmark on 41 LLMs."}
{"id": 7617, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "analysis", "why": "Cognitive modeling of LLM behavior in sequential jailbreaks"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["elicitation task"], "benchmarks": [{"name": "Contextual Iowa Gambling Task", "purpose": "elicit decision behavior", "is_new": true}], "results": [], "workflow": [], "summary": "Models LLM sequential jailbreak behavior with a Rescorla-Wagner cognitive framework."}
{"id": 4191, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "train-agent", "why": "Demonstration-guided policy optimization for molecular optimization"}, {"area": "align/rlvr", "role": "method", "why": "Combines reference molecules with RLVR"}], "pattern": "fix-env-vary-agent", "fixes": ["molecule datasets"], "varies": ["training algorithm"], "benchmarks": [], "results": [], "workflow": ["Use reference molecules as supervision", "Policy optimization"], "summary": "DePO adapts LLMs to molecular optimization with demonstration-guided RL."}
{"id": 18445, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "analysis", "why": "Geometry analysis of BERT and GPT-2 embeddings by linguistic features"}], "pattern": "vary-conditions", "fixes": ["BERT", "GPT-2"], "varies": ["token features", "layers"], "benchmarks": [{"name": "UD English-EWT", "purpose": "token partitioning", "is_new": false}], "results": [], "workflow": [], "summary": "Analyzes embedding geometry of BERT and GPT-2 across layers and linguistic features."}
{"id": 16107, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "Multi-chart image QA dataset evaluating nine MLMs"}], "pattern": "fix-agent-vary-env", "fixes": ["nine MLMs"], "varies": ["multi-chart questions"], "benchmarks": [{"name": "PolyChartQA", "purpose": "multi-chart QA", "is_new": true}], "results": [], "workflow": [], "summary": "PolyChartQA is a multi-chart QA benchmark of 534 images and 2,694 QA pairs."}
{"id": 13840, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "data", "why": "Off-policy influence-based data selection for RLVR"}], "pattern": "vary-data", "fixes": ["RLVR recipe"], "varies": ["selected data"], "benchmarks": [], "results": [], "workflow": ["Estimate influence off-policy", "Select data", "Run RLVR"], "summary": "Selects RLVR training data efficiently using off-policy influence estimation."}
{"id": 919, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "train-agent", "why": "Trains tool-integrated LLM judges with RL and Python executor"}], "pattern": "vary-evaluation", "fixes": ["judge tasks"], "varies": ["judge training"], "benchmarks": [], "results": [], "workflow": ["Integrate Python executor", "Diverse RL training", "Judge"], "summary": "TIR-Judge trains LLM judges with tool-integrated RL."}
{"id": 11717, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "benchmark", "why": "Multi-hop QA benchmark on new and long-tail knowledge"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["knowledge familiarity"], "benchmarks": [{"name": "MINTQA", "purpose": "multi-hop QA with new/tail knowledge", "is_new": true}], "results": [], "workflow": [], "summary": "MINTQA evaluates multi-hop RAG QA with 10,479 and 17,887 question pairs."}
{"id": 14583, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Layered detection pipeline for securing MCP tool interactions"}], "pattern": "fix-agent-vary-env", "fixes": ["threat model"], "varies": ["defense stages"], "benchmarks": [], "results": [], "workflow": ["Static scanning", "Neural detector", "Final arbitration"], "summary": "MCP-Guard is a multi-stage defense against prompt injection and exfiltration in MCP."}
{"id": 8907, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Dynamic one-shot policy refinement for resource-efficient RLVR"}, {"area": "align/rlvr", "role": "theory", "why": "Sample-complexity lower bound for unlocking reasoning"}], "pattern": "vary-data", "fixes": ["reasoning benchmarks"], "varies": ["training sample budget"], "benchmarks": [], "results": [], "workflow": ["Select one-shot sample", "Dynamic refinement"], "summary": "Shows RLVR can be data- and compute-efficient via dynamic one-shot refinement."}
{"id": 9429, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Multi-agent RL for planner, filter, answerer deep search"}, {"area": "multi-agent", "role": "train-agent", "why": "Role-specialized agents trained jointly"}], "pattern": "fix-env-vary-agent", "fixes": ["search benchmarks"], "varies": ["role decomposition", "MARL training"], "benchmarks": [], "results": [], "workflow": ["Planner navigates", "Filter curates memory", "Answerer synthesizes"], "summary": "DECOR trains role-specialized agents with MARL for deep search."}
{"id": 12603, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Tool-integrated policy optimization for GUI agent with copilot"}, {"area": "web-gui", "role": "scaffold", "why": "Memory decoupling and copilot for retrieval and computation"}], "pattern": "fix-env-vary-agent", "fixes": ["GUI tasks"], "varies": ["copilot tools", "policy training"], "benchmarks": [], "results": [], "workflow": ["Decouple memory", "Copilot assists", "Policy optimization"], "summary": "UI-Copilot aids long-horizon GUI agents with a memory and calculation copilot."}
{"id": 9532, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Auction-based strategy bidding among small agents with shared memory"}, {"area": "search-research", "role": "analysis", "why": "Shows small agents fail to scale on deep search and coding"}], "pattern": "vary-scaffold", "fixes": ["small models"], "varies": ["coordination mechanism"], "benchmarks": [], "results": [], "workflow": ["Agents bid with plans", "Score cost-value", "Refine auction memory"], "summary": "SALE lets small agents bid strategic plans to scale on complex tasks."}
{"id": 16538, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "benchmark", "why": "Verb sense clustering for more robust activity recognition evaluation"}], "pattern": "vary-evaluation", "fixes": ["recognition systems"], "varies": ["evaluation protocol"], "benchmarks": [{"name": "imSitu", "purpose": "analyze verb ambiguity", "is_new": false}], "results": [], "workflow": [], "summary": "Proposes vision-language verb sense clusters for robust visual activity recognition evaluation."}
{"id": 16867, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "Reference-level feedback guides synthesis of instruction data"}], "pattern": "vary-data", "fixes": ["Llama-3.1-8B SFT"], "varies": ["synthetic data"], "benchmarks": [{"name": "REFED", "purpose": "10K synthesized instruction pairs", "is_new": true}], "results": [], "workflow": ["Extract reference traits", "Guide synthesis"], "summary": "Reference-level feedback yields REFED, higher-quality synthetic instruction data."}
{"id": 17426, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "EPAG benchmark of LLM pre-consultation using diagnostic guidelines"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["HPI amount", "language"], "benchmarks": [{"name": "EPAG", "purpose": "pre-consultation evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "EPAG benchmarks LLM pre-consultation ability against diagnostic guidelines."}
{"id": 12038, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "eval-method", "why": "PTE hardware-aware efficiency metric for tool-integrated reasoning"}, {"area": "tool-use", "role": "analysis", "why": "Unveils inefficiency patterns in TIR"}], "pattern": "vary-evaluation", "fixes": ["TIR systems"], "varies": ["efficiency metric"], "benchmarks": [{"name": "PTE", "purpose": "TIR efficiency metric", "is_new": true}], "results": [], "workflow": [], "summary": "Introduces PTE, a hardware-aware efficiency metric for tool-integrated reasoning."}
{"id": 1786, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "KV caching and guided diffusion to accelerate diffusion LM inference"}], "pattern": "n/a", "fixes": ["Dream 7B", "LLaDA 8B"], "varies": ["inference method"], "benchmarks": [], "results": [], "workflow": ["KV caching", "Guided diffusion sampling"], "summary": "FlashDLM speeds up diffusion language model inference with KV caching and guidance."}
{"id": 14739, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "LLM-agent framework simulating legal societies"}, {"area": "domain-agents", "role": "application", "why": "Legislation and institutional simulation"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["legal scenarios"], "benchmarks": [], "results": [], "workflow": [], "summary": "Law in Silico simulates legal societies with LLM agents."}
{"id": 1274, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "MCTS-inspired tool planning with dual feedback and pruning"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["tool planning"], "benchmarks": [], "results": [], "workflow": ["Explore tool trajectories", "Dual LLM evaluation", "Bidirectional pruning"], "summary": "ToolTree plans tool use via Monte Carlo tree search with pruning."}
{"id": 15033, "in_agent_scope": false, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Experience replay of corrections for text-to-optimization modeling"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["experience memory"], "benchmarks": [], "results": [], "workflow": ["Distill interaction histories", "Index correction mappings", "Retrieve"], "summary": "EOM reuses causal correction experience to improve text-to-optimization modeling."}
{"id": 13007, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Normalized-entropy step-level supervision for LLM agent RL"}], "pattern": "fix-env-vary-agent", "fixes": ["long-horizon tasks"], "varies": ["RL objective"], "benchmarks": [], "results": [], "workflow": ["Compute normalized entropy", "Select trajectory steps", "Policy optimization"], "summary": "STAPO uses normalized entropy for selective trajectory-aware agent RL."}
{"id": 13415, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "Multi-turn video misinformation reasoning benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["deception categories"], "benchmarks": [{"name": "MisVideoQA", "purpose": "video misinformation analysis", "is_new": true}], "results": [], "workflow": [], "summary": "MisVideoQA evaluates multi-turn reasoning on video misinformation across 12 categories."}
{"id": 15486, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "SWE-level coding benchmark for Ascend CANN"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "LiveCANNBench", "purpose": "CANN SWE coding", "is_new": true}], "results": [], "workflow": [], "summary": "LiveCANNBench evaluates AI coding agents on the Ascend CANN software stack."}
{"id": 6684, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Internalizes tool representations into parameters rather than context"}], "pattern": "fix-env-vary-agent", "fixes": ["tool benchmarks"], "varies": ["tool representation"], "benchmarks": [], "results": [], "workflow": ["Encode tools into parameters", "Call without documentation"], "summary": "ParaTool moves tool knowledge from context into model parameters."}
{"id": 4454, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "scaffold", "why": "Training-free guidance decoding lifts reasoning to omni-modal models"}], "pattern": "vary-scaffold", "fixes": ["OLLM weights"], "varies": ["decoding guidance"], "benchmarks": [], "results": [], "workflow": ["Guide with reasoning model", "Decode"], "summary": "ThinkOmni is training-free guidance decoding that transfers textual reasoning to omni-modal LLMs."}
{"id": 15274, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RL co-evolving policy and tool metadata for dialogue agents"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "Autonomous tool metadata evolution"}], "pattern": "fix-env-vary-agent", "fixes": ["dialogue tasks"], "varies": ["policy", "tool metadata"], "benchmarks": [], "results": [], "workflow": ["RL rollouts", "Evolve metadata", "Refine rewards"], "summary": "ToolDNA co-evolves policy and tool metadata via RL for task-oriented dialogue."}
{"id": 14271, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Verbal epsilon-greedy exploration for multi-turn RL agents"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["exploration strategy"], "benchmarks": [], "results": [], "workflow": ["Use external feedback", "Balance explore/exploit"], "summary": "VEG improves exploration in multi-turn LLM agent RL with verbal epsilon-greedy."}
{"id": 3595, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "data", "why": "Intent-aware elicitation improves long-form attributed QA and synthetic data"}], "pattern": "vary-data", "fixes": ["LLMs"], "varies": ["intent schemes"], "benchmarks": [], "results": [], "workflow": ["Extract intent tags", "Generate reports", "Synthesize training data"], "summary": "Intent awareness improves attributed long-form QA zero-shot and via synthetic data."}
{"id": 13074, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Multilingual GUI perception and reasoning benchmark"}, {"area": "web-gui", "role": "data", "why": "Enhancement for multilingual GUI agents"}], "pattern": "fix-agent-vary-env", "fixes": ["LVLMs"], "varies": ["language", "capability"], "benchmarks": [{"name": "MPR-GUI", "purpose": "multilingual GUI P&R", "is_new": true}], "results": [], "workflow": [], "summary": "MPR-GUI benchmarks and improves multilingual perception and reasoning in GUI agents."}
{"id": 15955, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Self-evolving agent strategy library for text steganography"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["strategy library"], "benchmarks": [], "results": [], "workflow": ["Generate", "Evaluate", "Summarize", "Update library"], "summary": "Auto-Stega evolves steganographic strategies at inference time via an agent loop."}
{"id": 12965, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Out-of-sample classical Chinese poetry reasoning benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["poetry tasks"], "benchmarks": [{"name": "Neo-Classic", "purpose": "aesthetic reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "Neo-Classic evaluates linguistic-aesthetic reasoning on contemporary expert-written classical poetry."}
{"id": 7048, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Milestone-based credit assignment for long-horizon agent RL"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["reward shaping"], "benchmarks": [], "results": [], "workflow": ["Partition at milestones", "Shape rewards", "Policy update"], "summary": "BEACON uses milestones for precise credit assignment in long-horizon language agent RL."}
{"id": 13806, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Training-free role-state structured reasoning for multi-turn Text-to-SQL"}], "pattern": "vary-scaffold", "fixes": ["small LRMs"], "varies": ["in-context framework"], "benchmarks": [], "results": [], "workflow": ["Build Role-State", "Link schema to SQL", "Generate"], "summary": "Rose-SQL uses Role-State representation for training-free multi-turn Text-to-SQL."}
{"id": 13638, "in_agent_scope": false, "types": [{"area": "train/pretrain", "role": "analysis", "why": "Replicates studies varying hyperparameters to test conclusion robustness"}], "pattern": "vary-conditions", "fixes": ["prior studies"], "varies": ["hyperparameters"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows hyperparameter choices can change qualitative conclusions in computational linguistics LM studies."}
{"id": 9931, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "CLIP-guided diffusion transformer generating M/EEG from images"}], "pattern": "n/a", "fixes": [], "varies": ["generation model"], "benchmarks": [], "results": [], "workflow": ["CLIP guidance", "Diffusion transformer"], "summary": "Image-to-brain-signal generation with CLIP-guided diffusion for visual prostheses."}
{"id": 14888, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "scaffold", "why": "Routes clinical NLI pairs to specialised solvers with verification"}], "pattern": "vary-scaffold", "fixes": ["backbones"], "varies": ["agentic routing"], "benchmarks": [{"name": "CTNLI", "purpose": "clinical NLI", "is_new": false}], "results": [], "workflow": ["Route to family", "Specialised solver", "Verify and refine"], "summary": "CARENLI routes clinical NLI to reasoning-family solvers with verification."}
{"id": 1458, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "Interactive scientific law discovery benchmark with 324 tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["tasks"], "benchmarks": [{"name": "NewtonBench", "purpose": "law discovery", "is_new": true}], "results": [], "workflow": [], "summary": "NewtonBench evaluates LLM agents on interactive scientific law discovery."}
{"id": 16757, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "analysis", "why": "Query linguistic features and hallucination risk across 369,837 queries"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["query features"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analyzes how query linguistic features affect hallucination likelihood."}
{"id": 13419, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "application", "why": "E-commerce risk management GUI agent framework"}, {"area": "web-gui", "role": "data", "why": "Builds trajectory dataset for GUI agents"}], "pattern": "n/a", "fixes": ["general model"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": ["Collect trajectories", "Train GUI agent", "Deploy"], "summary": "RISK builds and deploys GUI agents for e-commerce risk management with a trajectory dataset."}