Files
ai-taxonomy-literature-anal…/data/labels/o15.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
27 KiB
JSON

{"id": 16661, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Bilingual English/Farsi dataset of true-but-misleading hints in factual QA"}], "pattern": "fix-env-vary-agent", "fixes": ["evaluated LLMs"], "varies": ["QA tasks with misleading true hints"], "benchmarks": [{"name": "TruthTrap", "purpose": "evaluate misleading true hints", "is_new": true}], "results": [], "workflow": [], "summary": "Introduces TruthTrap benchmark showing true yet persuasive hints degrade LLM factual QA accuracy."}
{"id": 13376, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Multi-turn RL with execution feedback for Text-to-SQL agent"}], "pattern": "fix-env-vary-agent", "fixes": ["BIRD-SQL benchmark"], "varies": ["RL training, multi-turn interaction"], "benchmarks": [{"name": "BIRD-SQL", "purpose": "evaluate Text-to-SQL", "is_new": false}], "results": [], "workflow": [], "summary": "Multi-turn RL agent for Text-to-SQL interacting with the database and refining queries from execution feedback."}
{"id": 9815, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Retrieval agent recommending baselines and datasets for experiment design"}], "pattern": "vary-scaffold", "fixes": ["general LLMs"], "varies": ["retrieval agent pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "Framework and knowledge base of 108,825 papers for recommending baselines and datasets to researchers."}
{"id": 13111, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "analysis", "why": "Evaluates LLM tutor answer-leakage robustness under adversarial student attacks"}, {"area": "agent-safety", "role": "attack", "why": "Adapts attacks to make tutors leak answers"}], "pattern": "vary-conditions", "fixes": ["tutor models"], "varies": ["adversarial student attacks"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies how LLM tutors leak answers under adversarial student attacks across model families and a multi-agent design."}
{"id": 16977, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "method", "why": "Neuron-guided expert allocation for MoE cross-lingual extension"}], "pattern": "vary-scaffold", "fixes": ["benchmark"], "varies": ["expert allocation per layer"], "benchmarks": [], "results": [], "workflow": [], "summary": "NeuronMoE allocates experts per layer using language-specific neuron diversity to extend LLMs to low-resource languages."}
{"id": 13267, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "analysis", "why": "Probes latent reasoning vectors with structural, causal, geometric analysis"}, {"area": "interp/steering", "role": "method", "why": "Training-free interventions guided by interpretability findings"}], "pattern": "vary-conditions", "fixes": ["latent reasoning models"], "varies": ["interventions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Interprets latent reasoning vectors and builds interpretability-guided interventions to control them."}
{"id": 3882, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "RL self-evolution of multi-agent LLMs through interaction rewards"}, {"area": "multi-agent", "role": "train-agent", "why": "Agents co-evolve via mutual discussion"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["interaction-based rewards"], "benchmarks": [], "results": [], "workflow": [], "summary": "CoMAS lets LLM agents self-evolve with RL using intrinsic rewards from inter-agent interaction."}
{"id": 15943, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Consensus-free debate removing majority voting and multi-round conformity"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["debate protocol"], "benchmarks": [], "results": [], "workflow": [], "summary": "Free-MAD replaces consensus-based multi-agent debate with a consensus-free protocol reducing token cost and conformity."}
{"id": 1145, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "defense", "why": "Answer-Then-Check alignment: model answers in thoughts then checks safety"}], "pattern": "vary-data", "fixes": ["threat model"], "varies": ["safety alignment method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Safety alignment training where the model drafts an answer and evaluates its safety before responding."}
{"id": 6088, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "Process-level diagnostic metrics for multi-agent debate"}, {"area": "multi-agent", "role": "analysis", "why": "Diagnoses sycophancy and premature convergence"}], "pattern": "vary-evaluation", "fixes": ["debate systems"], "varies": ["evaluation metrics"], "benchmarks": [], "results": [], "workflow": [], "summary": "Diagnostic framework scoring debate process (engagement, influence, balance) aligned better with human judgments than outcome proxies."}
{"id": 6356, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Agentic workflow calibrating agents' epistemic judgment of plan feasibility"}, {"area": "planning-decision", "role": "analysis", "why": "Identifies epistemic miscalibration in planning"}], "pattern": "vary-scaffold", "fixes": ["LLM agents"], "varies": ["calibration workflow"], "benchmarks": [], "results": [], "workflow": [], "summary": "EPC-AW workflow addresses epistemic miscalibration where multi-agent plans fail despite correct execution."}
{"id": 11965, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Training-free dual-cluster memory of approaches, checklists, pitfalls for optimization modeling"}], "pattern": "vary-scaffold", "fixes": ["frozen LLM"], "varies": ["memory construction and augmentation"], "benchmarks": [], "results": [], "workflow": [], "summary": "DCM-Agent uses clustered historical solutions as memory to guide optimization problem modeling without training."}
{"id": 14463, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Agentic knowledge warehouse for information seeking"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["knowledge warehousing agent"], "benchmarks": [], "results": [], "workflow": [], "summary": "AWARE agentically builds a knowledge warehouse to handle complex information-seeking beyond RAG and interleaved retrieval."}
{"id": 11662, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Curiosity-driven exploration for VLM agents via RL"}, {"area": "embodied-nav", "role": "train-agent", "why": "Visual-linguistic curiosity in partially observable environments"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["exploration reward, RL"], "benchmarks": [], "results": [], "workflow": [], "summary": "GLANCE drives VLM agent exploration with visual-linguistic curiosity to improve sparse-reward generalization."}
{"id": 280, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Mobile agent benchmark with environmental injection in on-device GUIs"}, {"area": "agent-safety", "role": "attack", "why": "Adversarial overlays and spoofed notifications"}], "pattern": "fix-agent-vary-env", "fixes": ["mobile VLM agents"], "varies": ["injected UI elements"], "benchmarks": [{"name": "GhostEI-Bench", "purpose": "environment injection robustness", "is_new": true}], "results": [], "workflow": [], "summary": "GhostEI-Bench tests mobile GUI agent resilience to environmental injection such as deceptive overlays."}
{"id": 14576, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "method", "why": "Multi-agent framework with symbolic solver for logical reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-agent logic framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "MAC-Reasoner uses translator and solver agents to build logic-augmented context for logical reasoning."}
{"id": 13850, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "RL algorithm for parallel exploration across multiple environments"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "DPEPO trains LLM agents to explore multiple environments in parallel and share cross-trajectory experience."}
{"id": 15648, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Cluster reasoning traces to cut critic cost in weighted self-consistency"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["answer selection method"], "benchmarks": [], "results": [], "workflow": [], "summary": "VecCISC reduces confidence-informed self-consistency overhead using trace clustering and candidate selection."}
{"id": 10744, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Self-evolving code agent extracts hidden system prompts"}], "pattern": "fix-agent-vary-env", "fixes": ["target LLMs"], "varies": ["extraction strategies"], "benchmarks": [], "results": [], "workflow": [], "summary": "JustAsk autonomously discovers strategies for extracting system prompts from frontier LLMs via code agent interaction."}
{"id": 16675, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent design editing for posters and websites"}, {"area": "domain-agents", "role": "benchmark", "why": "Introduces SMARTEdit-Bench"}], "pattern": "vary-scaffold", "fixes": ["MLLMs"], "varies": ["editing framework"], "benchmarks": [{"name": "SMARTEdit-Bench", "purpose": "evaluate design editing", "is_new": true}], "results": [], "workflow": [], "summary": "SMART-Editor is a multi-agent framework with reward-guided refinement and RewardDPO for structured design editing."}
{"id": 5335, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "data", "why": "Automated error injection building 9,533 annotated faulty trajectories"}, {"area": "multi-agent", "role": "analysis", "why": "Error attribution in multi-agent systems"}], "pattern": "vary-data", "fixes": ["MAS"], "varies": ["error attribution data"], "benchmarks": [], "results": [], "workflow": [], "summary": "Aegis builds a large dataset of annotated faulty multi-agent trajectories for training error attribution."}
{"id": 948, "in_agent_scope": false, "types": [{"area": "train/continual", "role": "method", "why": "Model merging with learned layer-wise coefficients from unlabeled data"}], "pattern": "vary-scaffold", "fixes": ["experts"], "varies": ["merging coefficients"], "benchmarks": [], "results": [], "workflow": [], "summary": "Expert Merging learns layer-wise merging coefficients from unlabeled calibration data with layer chunking."}
{"id": 2977, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Online RL-based speculative planning to cut agent latency and cost"}], "pattern": "vary-scaffold", "fixes": ["LLM agents"], "varies": ["speculative planning"], "benchmarks": [], "results": [], "workflow": [], "summary": "Dynamic Speculative Planning accelerates LLM agents with an asynchronous online RL controller trading speed and cost."}
{"id": 18503, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "analysis", "why": "Tests whether LLM embodied agents communicate effectively"}, {"area": "multi-agent", "role": "benchmark", "why": "Extends PARTNR with dialogue channel"}], "pattern": "vary-conditions", "fixes": ["PARTNR tasks"], "varies": ["communication channel"], "benchmarks": [{"name": "PARTNR", "purpose": "collaborative household robotics", "is_new": false}], "results": [], "workflow": [], "summary": "Extends PARTNR with dialogue to examine whether embodied LLM agents align world models by communication."}
{"id": 5138, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "data", "why": "Agentic multi-stage synthesis of video reasoning data for RLVR"}], "pattern": "vary-data", "fixes": ["RLVR recipe"], "varies": ["video reasoning data"], "benchmarks": [{"name": "ReWatch", "purpose": "video reasoning data", "is_new": true}], "results": [], "workflow": [], "summary": "ReWatch dataset synthesized by multi-agent pipeline to bootstrap RLVR for video reasoning in LVLMs."}
{"id": 3607, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "Data-lake-to-insight pipeline benchmark with 104 challenges"}], "pattern": "fix-env-vary-agent", "fixes": ["AI systems"], "varies": ["data pipeline tasks"], "benchmarks": [{"name": "KramaBench", "purpose": "data-to-insight pipelines", "is_new": true}], "results": [], "workflow": [], "summary": "KramaBench evaluates AI systems on end-to-end data pipelines over data lakes."}
{"id": 1535, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Agentic framework evaluating LLM-crafted heuristics for combinatorial optimization"}], "pattern": "fix-env-vary-agent", "fixes": ["LLMs"], "varies": ["optimization problems"], "benchmarks": [{"name": "HeuriGym", "purpose": "LLM heuristics evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "HeuriGym benchmarks LLMs proposing heuristics with execution feedback for combinatorial optimization."}
{"id": 14531, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "application", "why": "Full-lifecycle app synthesis with computer-use agent evaluator"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["synthesis framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Apeiron synthesizes apps autonomously using CUA persona evaluation, activity tracing and locality control."}
{"id": 12320, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "Self-improvement of unified multimodal models via proposer, solver, judge roles"}], "pattern": "vary-data", "fixes": ["UMM"], "varies": ["self-generated supervision"], "benchmarks": [], "results": [], "workflow": [], "summary": "UniCorn self-improves unified multimodal models with self-generated supervision, no external data."}
{"id": 14786, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL for agentic conversational search with contextual reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["RL training"], "benchmarks": [], "results": [], "workflow": [], "summary": "Trains an agent via RL to jointly handle query reformulation, retrieval and generation in multi-turn dialogue."}
{"id": 12653, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Poisons multimodal long-term memory of agentic recommenders via images"}], "pattern": "fix-agent-vary-env", "fixes": ["agentic recommender"], "varies": ["poisoned images"], "benchmarks": [], "results": [], "workflow": [], "summary": "Visual Inception injects triggers in user images that sleep in memory to compromise agent planning."}
{"id": 8572, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Budgeted controller deciding when and how to replan"}, {"area": "planning-decision", "role": "scaffold", "why": "Latency-aware replanning control"}], "pattern": "vary-scaffold", "fixes": ["LLM planner"], "varies": ["replanning policy"], "benchmarks": [], "results": [], "workflow": [], "summary": "BRACE controls LLM replanning as a budgeted loop to avoid latency tails in embodied agents."}
{"id": 7661, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Active information foraging over long texts with decoupled epistemic states"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["foraging agent"], "benchmarks": [], "results": [], "workflow": [], "summary": "SCOUT agent actively forages query-relevant subsets of million-token texts instead of full-context processing."}
{"id": 15829, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Answer-free RL with cross-prompt consensus rewards from paraphrases"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["reward design"], "benchmarks": [], "results": [], "workflow": [], "summary": "CPC-GRPO derives rewards from model probabilities across prompt paraphrases without ground-truth answers."}
{"id": 4171, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "attack", "why": "Multi-modal guided attack recovering concepts from unlearned diffusion models"}], "pattern": "fix-agent-vary-env", "fixes": ["unlearned image models"], "varies": ["adversarial multimodal inputs"], "benchmarks": [], "results": [], "workflow": [], "summary": "Attack showing image-generation unlearning is not robust under multimodal adversarial inputs."}
{"id": 7276, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "attack", "why": "Jailbreak weaving benign sub-queries with adaptive tree search"}], "pattern": "fix-agent-vary-env", "fixes": ["guardrailed LLMs"], "varies": ["jailbreak strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "Bypasses commercial LLM guardrails by weaving harmless sub-queries via adaptive tree search."}
{"id": 6802, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RL framework preventing interaction collapse in multimodal tool use"}, {"area": "web-gui", "role": "train-agent", "why": "Agentic vision models for image and video"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL rollout and tool reward"], "benchmarks": [], "results": [], "workflow": [], "summary": "PyVision-RL stabilizes RL for open multimodal agents with oversampling-filtering rollouts and accumulative tool reward."}
{"id": 11633, "in_agent_scope": false, "types": [{"area": "align/values", "role": "data", "why": "Data optimization for cultural alignment via representativeness and distinctiveness"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["cultural alignment data"], "benchmarks": [], "results": [], "workflow": [], "summary": "CAReDiO optimizes cultural alignment data for representativeness and distinctiveness."}
{"id": 2819, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Prompt injection exploiting chat templates against LLM agents"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["injection format"], "benchmarks": [], "results": [], "workflow": [], "summary": "ChatInject abuses structured chat templates for stronger indirect prompt injection on LLM agents."}
{"id": 11417, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Actor-critic edge prediction to prune communication cost"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["communication structure"], "benchmarks": [], "results": [], "workflow": [], "summary": "AgentTailor judges semantic contribution of communication edges to cut multi-agent token cost."}
{"id": 3991, "in_agent_scope": false, "types": [{"area": "interp/icl", "role": "scaffold", "why": "Agentic example selection for multimodal in-context learning"}], "pattern": "vary-scaffold", "fixes": ["MLLMs"], "varies": ["context construction"], "benchmarks": [], "results": [], "workflow": [], "summary": "ContextNav agentically builds clean contexts for multimodal ICL with noisy examples."}
{"id": 12721, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Metacognitive decomposition and refinement to improve RL rollout efficiency"}], "pattern": "vary-scaffold", "fixes": ["weak LLMs"], "varies": ["rollout procedure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Cog-Rethinker improves zero-RL sample efficiency via hierarchical metacognitive rollouts."}
{"id": 2404, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "GRPO variant using process-verifiable environment feedback for coding agents"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["advantage shaping RL"], "benchmarks": [], "results": [], "workflow": [], "summary": "GVPO shapes advantages with execution feedback to train interactive coding agents."}
{"id": 761, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "ToM-grounded hierarchical emotion benchmark HitEmotion"}, {"area": "arch/multimodal", "role": "method", "why": "ToM-guided reasoning chain"}], "pattern": "fix-env-vary-agent", "fixes": ["MLLMs"], "varies": ["emotion tasks"], "benchmarks": [{"name": "HitEmotion", "purpose": "ToM emotion reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "HitEmotion benchmark and ToM-guided reasoning for multimodal emotion understanding."}
{"id": 15408, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Training-free adaptive speculative decoding using linguistic priors"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["verification strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "LinguaSpec adapts drafting and verification by token linguistic type, training-free."}
{"id": 12473, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Branch-point trajectory expansion from seed demos for desktop GUI agents"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["synthetic trajectories"], "benchmarks": [], "results": [], "workflow": [], "summary": "Anchor expands verified seed demonstrations into new task variants for scalable GUI agent data."}
{"id": 15447, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "LLM agents as controllers over frozen backbone for time series forecasting"}], "pattern": "vary-scaffold", "fixes": ["frozen backbone"], "varies": ["LLM residual control"], "benchmarks": [], "results": [], "workflow": [], "summary": "CTRL decouples LLM reasoning from numerical forecasting using residual learning."}
{"id": 13680, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Neuro-symbolic FSM-grounded planning for wet-lab protocols"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["FSM planning workflow"], "benchmarks": [], "results": [], "workflow": [], "summary": "BioProAgent anchors LLM planning in a finite state machine for safe wet-lab planning."}
{"id": 6139, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "attack", "why": "Bandit-guided stylistic edits to inflate LLM judge scores"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM judges"], "varies": ["style edits"], "benchmarks": [], "results": [], "workflow": [], "summary": "BITE exploits judge stylistic biases via LinUCB-selected edits to inflate scores."}
{"id": 6068, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "method", "why": "Property-oriented structurally minimal feedback for code refinement"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["feedback type"], "benchmarks": [], "results": [], "workflow": [], "summary": "PGS improves code refinement with high-quality property-based feedback instead of noisy tests."}
{"id": 15322, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "method", "why": "Difference-aware contrastive learning for automatic answer grading"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["training objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "DiffCL uses difference labeling and dual-contrastive learning for answer grading."}
{"id": 11066, "in_agent_scope": true, "types": [{"area": "science-data", "role": "train-agent", "why": "Agentic LLM trained with curriculum for end-to-end data science"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["agentic training"], "benchmarks": [], "results": [], "workflow": [], "summary": "DeepAnalyze is an agentic LLM for autonomous data science from raw data to reports."}
{"id": 15098, "in_agent_scope": false, "types": [{"area": "train/sft", "role": "method", "why": "Probability-guided token selection in SFT"}], "pattern": "vary-data", "fixes": ["SFT recipe"], "varies": ["token selection"], "benchmarks": [], "results": [], "workflow": [], "summary": "ProFit selects high-value tokens by probability to reduce single-reference overfitting in SFT."}
{"id": 3994, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "RL-trained memory agent processing long text in segments"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["memory workflow, DAPO RL"], "benchmarks": [], "results": [], "workflow": [], "summary": "MemAgent processes long text in segments with overwrite memory, trained end-to-end with extended DAPO."}
{"id": 4250, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Reference-guided tool creation from textbooks for reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["tool creation"], "benchmarks": [], "results": [], "workflow": [], "summary": "RefTool creates executable tools from external references for knowledge-intensive reasoning."}
{"id": 11785, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Two-stage multi-agent spreadsheet understanding across formats"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "SpreadsheetAgent reads massive spreadsheets step by step with multi-format reasoning."}
{"id": 15818, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Training-free verifier-free test-time scaling with confidence-guided sampling"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["sampling and aggregation"], "benchmarks": [], "results": [], "workflow": [], "summary": "ConMA reallocates a fixed inference budget through iterative sample-filter-diversify-select cycles."}
{"id": 613, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Self-critiquing region-guided chest X-ray report generation"}], "pattern": "vary-scaffold", "fixes": ["baselines"], "varies": ["self-critique mechanism"], "benchmarks": [], "results": [], "workflow": [], "summary": "Learns self-critiquing reasoning for region-grounded radiology report generation."}
{"id": 15083, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "RL emotion planning module for streaming emotional TTS"}], "pattern": "vary-scaffold", "fixes": ["TTS"], "varies": ["emotion planner"], "benchmarks": [], "results": [], "workflow": [], "summary": "Self-EmoQ plans emotions with RL before generation to drive streaming emotional TTS."}
{"id": 12494, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Disagreement-guided instruction refinement for zero-shot NER"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["instructions"], "benchmarks": [], "results": [], "workflow": [], "summary": "DiZiNER refines NER instructions by simulating pilot annotation disagreement."}