Files
ai-taxonomy-literature-anal…/data/labels/o22.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
30 KiB
JSON

{"id": 14470, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Executive memory model managing dependency-aware memory over reasoning steps for tool agents"}], "pattern": "vary-scaffold", "fixes": ["base agent", "tasks"], "varies": ["memory mechanism"], "benchmarks": [], "results": [], "workflow": ["build dependency-aware memory", "prune and fold context", "feed back to agent"], "summary": "Executive memory model that builds dependency-aware memory over reasoning steps to sustain long-horizon tool-augmented reasoning."}
{"id": 3760, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "100 PhD-level deep research tasks across 22 fields"}, {"area": "search-research", "role": "eval-method", "why": "Two complementary evaluation methodologies for research reports"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks", "evaluation protocol"], "benchmarks": [{"name": "DeepResearch Bench", "purpose": "evaluating deep research agents", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 100 PhD-level research tasks with two complementary evaluation frameworks for deep research agents."}
{"id": 16180, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Adversarial multi-agent debate with visual falsification to reduce diagnostic hallucination"}, {"area": "domain-agents", "role": "application", "why": "Medical MLLM diagnosis system"}], "pattern": "vary-scaffold", "fixes": ["MLLM", "tasks"], "varies": ["debate structure"], "benchmarks": [], "results": [], "workflow": ["proponent proposes hypothesis", "opponent falsifies visually", "reconcile diagnosis"], "summary": "Proponent-opponent multi-agent debate with visual falsification module mitigating confirmation bias and hallucination in medical diagnosis."}
{"id": 3642, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free attention-guided layer-wise intervention to reduce LVLM hallucination"}], "pattern": "vary-scaffold", "fixes": ["LVLM weights"], "varies": ["decoding intervention"], "benchmarks": [], "results": [], "workflow": ["analyze token attention", "guide layer-wise intervention"], "summary": "Training-free AGE intervenes layer-wise using attention patterns of real tokens to mitigate hallucination in vision-language models."}
{"id": 14228, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "data", "why": "Multi-agent simulation synthesizing shopping trajectories for post-training"}, {"area": "social-sim", "role": "environment", "why": "Journey-aware stateful user simulator"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["synthetic trajectories"], "benchmarks": [], "results": [], "workflow": ["simulate user journey", "agent uses tools", "critic filters trajectories", "distill small model"], "summary": "Closed-loop multi-agent simulation with user simulator and critic generates shopping-dialogue trajectories to distill small agents."}
{"id": 5016, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "20 end-to-end LLM research tasks"}, {"area": "science-data", "role": "environment", "why": "ResearchGym environment for research agents"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["research tasks"], "benchmarks": [{"name": "InnovatorBench", "purpose": "end-to-end AI research agent evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark and ResearchGym platform with 20 LLM-research tasks assessing agents' end-to-end innovative research ability."}
{"id": 17477, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Dialectic two-agent framework for RTL code optimization"}, {"area": "domain-agents", "role": "application", "why": "EDA RTL PPA optimization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent workflow"], "benchmarks": [], "results": [], "workflow": ["articulate transformation plan", "predict outcomes", "reconcile deviations"], "summary": "Dialectic multi-agent framework with deterministic evaluation to automate RTL optimization for power, performance and area."}
{"id": 11915, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Dual-agent dynamic global planning for retrieval-augmented reasoning"}], "pattern": "vary-scaffold", "fixes": ["search-augmented LLM"], "varies": ["planning agent"], "benchmarks": [], "results": [], "workflow": ["plan global search chain", "execute retrieval reasoning"], "summary": "Dual-agent planning paradigm addressing flawed search chains and distractor hijacking in RL-trained search-augmented reasoning."}
{"id": 14144, "in_agent_scope": false, "types": [{"area": "eval/robust", "role": "eval-method", "why": "Answer instability across stochastic generations as label-free uncertainty signal"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["uncertainty metric"], "benchmarks": [], "results": [], "workflow": [], "summary": "Proposes answer variability across repeated samples as black-box label-free uncertainty estimate, tested on reasoning, MCQ and constraint tasks."}
{"id": 10152, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "LLM semantic agent with graph evaluator in evolutionary NAS loop"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search framework"], "benchmarks": [], "results": [], "workflow": ["agent proposes architectures", "graph evaluator scores", "evolve candidates"], "summary": "Closed-loop evolutionary NAS combining LLM semantic agent with graph-based zero-cost-proxy evaluator."}
{"id": 12868, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Train agentic graph retriever with synthetic stepwise supervision"}, {"area": "search-research", "role": "data", "why": "Offline golden stepwise supervision signals"}], "pattern": "fix-env-vary-agent", "fixes": ["graph QA tasks"], "varies": ["retriever training signal"], "benchmarks": [], "results": [], "workflow": ["extract golden steps offline", "train step-wise retriever", "answer with subgraph"], "summary": "Agentic textual-graph retriever trained with synthetic stepwise supervision instead of sparse answer rewards."}
{"id": 12729, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "method", "why": "Cross-modal steganography hiding images in text via discrete diffusion latents"}], "pattern": "n/a", "fixes": ["n/a"], "varies": ["steganography container"], "benchmarks": [], "results": [], "workflow": ["compress image to binary payload", "embed in text"], "summary": "Steganography framework using linguistic containers and discrete-diffusion latent compression to hide secret images in text."}
{"id": 5527, "in_agent_scope": false, "types": [{"area": "arch/attention", "role": "theory", "why": "Aitchison-geometry characterization of attention token distinguishability"}], "pattern": "n/a", "fixes": [], "varies": ["assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Formal analysis showing token distinguishability in softmax attention converges to non-zero limit; derives temperature scaling."}
{"id": 6268, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Controlled study of policy-gradient design choices for agentic RL stability"}, {"area": "planning-decision", "role": "train-agent", "why": "Stable agentic RL training recipe"}], "pattern": "vary-conditions", "fixes": ["testbed"], "varies": ["policy-gradient components"], "benchmarks": [], "results": [], "workflow": [], "summary": "ARLArena standardized testbed decomposing policy gradient to analyze and stabilize multi-turn agentic RL training."}
{"id": 17457, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "method", "why": "Agent-based iterative explanation of SAE features"}], "pattern": "vary-scaffold", "fixes": ["SAE", "LLM"], "varies": ["explainer procedure"], "benchmarks": [], "results": [], "workflow": ["propose explanations", "test them", "refine explanation"], "summary": "Agentic explainer that formulates and tests multiple explanations for SAE features rather than single-pass generation."}
{"id": 5738, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Reinforced self-evolving pseudo-labeling for semi-supervised referring segmentation"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["pseudo-label method"], "benchmarks": [], "results": [], "workflow": [], "summary": "RL-based self-evolving pseudo-label construction with MLLM priors for semi-supervised referring expression segmentation."}
{"id": 14923, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Economic environment grounding LLM agents for opinion dynamics simulation"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["economic environment feedback"], "benchmarks": [], "results": [], "workflow": [], "summary": "Opinion dynamics simulation where LLM agents act in an economic environment and receive feedback, producing coherent individual and population trends."}
{"id": 8443, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "environment", "why": "Automated pipeline harvesting executable SWE tasks and RL environments"}, {"area": "code-swe", "role": "data", "why": "Large multilingual SWE task collection"}], "pattern": "fix-agent-vary-env", "fixes": ["RL recipe"], "varies": ["task collection"], "benchmarks": [{"name": "SWE-bench", "purpose": "related benchmark for SWE tasks", "is_new": false}], "results": [], "workflow": ["harvest repos", "synthesize install/test", "validate tasks"], "summary": "Language-agnostic automated pipeline building executable SWE tasks with install and test procedures for RL training."}
{"id": 15832, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "benchmark", "why": "CheST dataset for chemical structure-text coreference"}, {"area": "apps/science", "role": "method", "why": "RULER rule-guided multimodal RL"}], "pattern": "vary-both", "fixes": ["MLLM"], "varies": ["dataset", "RL reward rules"], "benchmarks": [{"name": "CheST", "purpose": "chemical structure-text coreference", "is_new": true}], "results": [], "workflow": [], "summary": "New chemical structure-text coreference dataset and rule-guided RL method for multimodal models in patent documents."}
{"id": 16472, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Natural-language zero-code framework for building LLM agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent-building framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Fully automated zero-code framework letting users build LLM agents and workflows through natural language alone."}
{"id": 11995, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "data", "why": "Influence-score-guided tree search for MAS data synthesis and selection"}, {"area": "multi-agent", "role": "train-agent", "why": "Self-training multi-agent systems"}], "pattern": "vary-data", "fixes": ["MAS", "training recipe"], "varies": ["data selection"], "benchmarks": [], "results": [], "workflow": ["MCTS generates data", "score influence", "select data", "train agents"], "summary": "Data Influence-oriented Tree Search uses influence scores rather than Q-values to guide data synthesis for MAS self-training."}
{"id": 17003, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Multi-step API calling benchmark from NL2SQL"}, {"area": "tool-use", "role": "environment", "why": "Interactive API environments converted from BIRD-SQL"}], "pattern": "fix-env-vary-agent", "fixes": ["LLMs"], "varies": ["API formulations"], "benchmarks": [{"name": "Live API Bench", "purpose": "multi-step tool calling", "is_new": true}, {"name": "BIRD-SQL", "purpose": "source of SQL queries", "is_new": false}], "results": [], "workflow": [], "summary": "Benchmark of 2500+ live APIs converting BIRD-SQL queries into executable API sequences in three formulations."}
{"id": 13899, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "eval-method", "why": "RL-trained MLLM judge for element-level text-to-image alignment"}], "pattern": "vary-evaluation", "fixes": ["T2I models"], "varies": ["evaluation metric"], "benchmarks": [], "results": [], "workflow": ["ground elements", "reason", "conclude alignment"], "summary": "Reinforcement-guided grounding-reasoning-conclusion evaluator for fine-grained, interpretable text-image alignment."}
{"id": 13406, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "method", "why": "Causal diagnosis predicting reliability of dynamically generated tools"}], "pattern": "vary-scaffold", "fixes": ["agent"], "varies": ["tool reliability prediction"], "benchmarks": [], "results": [], "workflow": ["build causal model", "run sandbox interventions", "predict confidence"], "summary": "Causal Tool Diagnosis builds structural causal model with sandbox interventions to predict confidence and attribute failures of generated tools."}
{"id": 13864, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "LLM generates vocal feature plan decoupling instruction understanding from speech synthesis"}], "pattern": "vary-scaffold", "fixes": ["TTS"], "varies": ["planning interface"], "benchmarks": [], "results": [], "workflow": ["LLM plans vocal features", "TTS synthesizes"], "summary": "LLM conductor produces textual vocal-feature plan that a TTS model renders for controllable speech."}
{"id": 11439, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "ProFinR professional finance reasoning benchmark, 528 problems"}, {"area": "domain-agents", "role": "application", "why": "ProFinAgent structured agent with finance tool library"}], "pattern": "vary-both", "fixes": ["LLMs"], "varies": ["benchmark", "agent structure"], "benchmarks": [{"name": "ProFinR", "purpose": "professional finance reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "Finance benchmark with 528 problems, a 53-tool library, and a structured agent framework."}
{"id": 15289, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "GUITestBench with 143 tasks across 26 defects"}, {"area": "web-gui", "role": "scaffold", "why": "Multi-agent decoupling navigation from verification"}], "pattern": "vary-both", "fixes": ["MLLM"], "varies": ["benchmark", "agent framework"], "benchmarks": [{"name": "GUITestBench", "purpose": "exploratory GUI defect discovery", "is_new": true}], "results": [], "workflow": ["plan and execute navigation", "verify defects separately"], "summary": "Interactive benchmark and multi-agent framework for GUI agents to discover defects during exploratory testing."}
{"id": 4078, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "attack", "why": "Energy-latency attack inducing endless verbose loops in MLLMs"}], "pattern": "fix-agent-vary-env", "fixes": ["MLLM"], "varies": ["adversarial input"], "benchmarks": [], "results": [], "workflow": [], "summary": "LingoLoop attack exploits POS and sentence structure to make MLLMs generate excessively long repetitive output."}
{"id": 5326, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Theory-of-Mind-grounded rebuttal agent"}, {"area": "domain-agents", "role": "application", "why": "Academic rebuttal generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["ToM strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "RebuttalAgent grounds academic rebuttal in Theory of Mind for strategic persuasion."}
{"id": 13860, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Agentic system for very long egocentric video understanding"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["agentic framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Agentic framework for compositional multi-hop reasoning over days-long egocentric video streams."}
{"id": 14130, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "LLM agent post-processing diarization segmentation errors"}], "pattern": "vary-scaffold", "fixes": ["diarization system"], "varies": ["post-processing agent"], "benchmarks": [], "results": [], "workflow": ["detect errors heuristically", "LLM decides merges", "filter by uncertainty"], "summary": "Lightweight LLM agent refines speaker diarization segmentation with uncertainty-aware filtering."}
{"id": 8421, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Dynamic semantic guidance in iterative mask generation for LLM-conditioned segmentation"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["segmentation method"], "benchmarks": [], "results": [], "workflow": [], "summary": "FlowSeg lets language semantics participate in iterative mask generation to reduce semantic misalignment."}
{"id": 18302, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "data", "why": "LLM teacher-student agents synthesize educational feedback data"}], "pattern": "vary-data", "fixes": ["LLMs"], "varies": ["synthetic feedback data"], "benchmarks": [], "results": [], "workflow": ["student LLM completes", "teacher LLM critiques"], "summary": "SEFL generates 19.8K synthetic student-work and teacher-feedback pairs via two role-playing LLMs, used for fine-tuning."}
{"id": 16835, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "General and domain-specific agents annotating time series"}, {"area": "domain-agents", "role": "application", "why": "Time series annotation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": [], "summary": "TESSA multi-agent system generates general and domain-specific time-series annotations."}
{"id": 14646, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Co-learning with multi-agent interaction to improve single LLM reasoning"}], "pattern": "vary-both", "fixes": ["benchmarks"], "varies": ["training framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "ILR co-learning uses dynamic cooperative or competitive interaction so LLMs internalize multi-agent insights for independent reasoning."}
{"id": 12514, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Black-box inference of MAS communication topology"}], "pattern": "fix-agent-vary-env", "fixes": ["MAS"], "varies": ["adversarial queries"], "benchmarks": [], "results": [], "workflow": [], "summary": "Communication Inference Attack infers multi-agent communication topology from black-box queries, exposing privacy and IP risk."}
{"id": 17824, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "benchmark", "why": "ArtistMus 1,000 music QA questions and MusWikiDB corpus"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks and corpus"], "benchmarks": [{"name": "ArtistMus", "purpose": "retrieval-augmented music QA", "is_new": true}, {"name": "MusWikiDB", "purpose": "3.2M passage retrieval database", "is_new": true}], "results": [], "workflow": [], "summary": "Music QA benchmark of 1,000 questions on 500 artists with a 3.2M-passage Wikipedia vector database."}
{"id": 1994, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "222 web tasks paired with safety and trust policies"}, {"area": "web-gui", "role": "eval-method", "why": "Completion Under Policy and Risk Ratio metrics"}], "pattern": "fix-env-vary-agent", "fixes": ["web agents"], "varies": ["policies", "metrics"], "benchmarks": [{"name": "ST-WebAgentBench", "purpose": "web agent safety and trustworthiness", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 222 web tasks with safety policies scored on six dimensions, plus Completion Under Policy metric."}
{"id": 6171, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Learn to generate robust agentic system designs from primitives"}], "pattern": "vary-scaffold", "fixes": ["tasks"], "varies": ["system design generator"], "benchmarks": [], "results": [], "workflow": [], "summary": "AutoRAS learns symbolic primitive sequences to design robust agentic systems using execution-derived safety signals and GFlowNet."}
{"id": 16148, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "LLM agent integrates action and object features for video action recognition"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["feature learning"], "benchmarks": [], "results": [], "workflow": [], "summary": "LLM agent-guided framework combines action and object-detection features to reduce spurious correlations in action recognition."}
{"id": 12103, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Think-while-listening inference with unspoken CoT for spoken models"}], "pattern": "n/a", "fixes": ["SLM"], "varies": ["inference framework"], "benchmarks": [], "results": [], "workflow": ["chunk speech", "reason per chunk"], "summary": "SHANKS streams speech chunks and generates unspoken reasoning while listening to reduce latency."}
{"id": 13027, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Three-agent framework scaffolding pedagogical decisions for lesson generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": ["classify sources", "order objectives", "synthesize lesson"], "summary": "TeachCraft uses Explorer, Planner and synthesizer agents to select, sequence and combine content into better lessons."}
{"id": 1245, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "analysis", "why": "Study of LLM self-prediction of task success and overconfidence in agentic tasks"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["models", "task progress"], "benchmarks": [], "results": [], "workflow": [], "summary": "Finds LLMs overconfident in predicting their success, worsening over multi-step tasks, and examines in-context learning of self-assessment."}
{"id": 14125, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Compress accessibility-tree observations for GUI agents"}], "pattern": "vary-scaffold", "fixes": ["agent model"], "varies": ["observation representation"], "benchmarks": [{"name": "OSWorld", "purpose": "GUI agent efficiency evaluation", "is_new": false}], "results": [], "workflow": ["detect modals", "reduce redundancy", "structure semantics"], "summary": "A11y-Compressor restructures and reduces accessibility-tree observations to improve GUI agent efficiency."}
{"id": 12754, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Agent-data mutual evolution for RL training"}, {"area": "planning-decision", "role": "data", "why": "Failure-guided task synthesis"}], "pattern": "vary-data", "fixes": ["environment"], "varies": ["training data distribution"], "benchmarks": [], "results": [], "workflow": ["roll out agent", "extract feedback", "synthesize tasks", "update data"], "summary": "CoEvolve extracts forgetting and uncertainty signals from rollouts to guide task synthesis and update data during agent RL."}
{"id": 12349, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Response anonymization to reduce identity bias in debate"}, {"area": "multi-agent", "role": "theory", "why": "Bayesian formalization of identity-weighted debate"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["anonymization"], "benchmarks": [], "results": [], "workflow": [], "summary": "Formalizes debate as identity-weighted Bayesian update and proposes anonymization to reduce sycophancy and self-bias."}
{"id": 16135, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "TVWorld-N and TVWorld-G for TV remote-control agents"}, {"area": "web-gui", "role": "environment", "why": "Offline graph abstraction of TV navigation"}], "pattern": "fix-env-vary-agent", "fixes": ["LVLMs"], "varies": ["tasks"], "benchmarks": [{"name": "TVWorld", "purpose": "TV navigation and grounding", "is_new": true}], "results": [], "workflow": [], "summary": "Offline graph-based TV navigation environment with navigation and grounding benchmarks for remote-control agents."}
{"id": 2013, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "analysis", "why": "Entropy analysis of CoT and RL in autoregressive image generation"}, {"area": "arch/multimodal", "role": "method", "why": "Entropy-guided optimization for T2I"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["entropy regularization"], "benchmarks": [], "results": [], "workflow": [], "summary": "Entropy analysis of CoT plus RL for T2I motivates entropy-guided optimization for stable image generation."}
{"id": 9940, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "MT-AgentRisk multi-turn tool agent safety benchmark"}, {"area": "agent-safety", "role": "defense", "why": "Self-exploration defense"}], "pattern": "fix-env-vary-agent", "fixes": ["agents"], "varies": ["multi-turn attack tasks"], "benchmarks": [{"name": "MT-AgentRisk", "purpose": "multi-turn tool agent safety", "is_new": true}], "results": [], "workflow": [], "summary": "Taxonomy and benchmark turning harmful single-turn tasks into multi-turn attacks on tool agents, with a defense."}
{"id": 17902, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Persian cultural short-answer benchmark"}, {"area": "eval/judge", "role": "eval-method", "why": "Morphology-aware soft-match scoring"}], "pattern": "fix-agent-vary-env", "fixes": ["15 LLMs"], "varies": ["tasks", "scoring"], "benchmarks": [{"name": "TARAZ", "purpose": "Persian cultural competence", "is_new": true}], "results": [], "workflow": [], "summary": "Persian short-answer cultural benchmark with morphological normalization and semantic soft-match scoring over 15 models."}
{"id": 16517, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Introspective MCTS for agentic AutoML"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search procedure"], "benchmarks": [], "results": [], "workflow": ["expand nodes", "introspect solutions", "select via value"], "summary": "I-MCTS expands tree nodes via introspective analysis of parent/sibling solutions for LLM AutoML agents."}
{"id": 4414, "in_agent_scope": false, "types": [{"area": "arch/tokenization", "role": "method", "why": "Generative iterative refinement of text embeddings with soft tokens"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["embedding method"], "benchmarks": [], "results": [], "workflow": [], "summary": "GIRCSE produces soft-token sequences refined under iterative contrastive objective for LLM text embeddings."}
{"id": 13168, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Evolving principle-guided supervision for medical MLLM reasoning"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["supervision signal"], "benchmarks": [], "results": [], "workflow": [], "summary": "Evo-PI evolves language-based reasoning principles as adaptive supervision for medical multimodal reasoning."}
{"id": 16183, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "data", "why": "Self-refinement curriculum synthesizing data for ultra low-resource languages"}], "pattern": "vary-data", "fixes": ["LLM"], "varies": ["synthetic curriculum"], "benchmarks": [], "results": [], "workflow": [], "summary": "CodeRise generates progressive curricula using language syntax and biased library sampling for ULRPL code generation."}
{"id": 14731, "in_agent_scope": false, "types": [{"area": "train/continual", "role": "theory", "why": "Conformal prediction guarantees under continual domain pretraining"}], "pattern": "n/a", "fixes": [], "varies": ["assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Robust uncertainty quantification with conformal guarantees for LLMs under shifting domains in continual pretraining."}
{"id": 5909, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Emergence of biased consensus in LLM debates driven by noise"}, {"area": "multi-agent", "role": "theory", "why": "Physics-inspired phase transition model"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["temperature", "conformity"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows biased collective norms emerge in multi-agent debate, with noise driving a predicted phase transition."}
{"id": 14488, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Controlled test of community discussion effects on LLM humor writing"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Social memory from critic and audience threads"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["discussion condition"], "benchmarks": [], "results": [], "workflow": [], "summary": "Community discussion stored as social memory improves stand-up comedy generation versus baseline, winning 75.6% of paired comparisons."}
{"id": 8534, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Temporal-spatial consistency guided expert activation to accelerate MoE diffusion LMs"}], "pattern": "n/a", "fixes": ["dLLM"], "varies": ["expert activation"], "benchmarks": [], "results": [], "workflow": [], "summary": "TEAM reduces expert activations in MoE diffusion LMs using consistency to accelerate decoding."}
{"id": 5338, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "RL fine-tuning for GUI visual grounding"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training method"], "benchmarks": [], "results": [], "workflow": [], "summary": "GuirlVG explores rule-based reinforcement fine-tuning for GUI visual grounding as an efficient alternative to SFT."}
{"id": 15485, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "ABC-Bench with 224 backend coding tasks"}], "pattern": "fix-env-vary-agent", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "ABC-Bench", "purpose": "agentic backend coding", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 224 executable backend development tasks requiring environment configuration and service deployment."}