Files
ai-taxonomy-literature-anal…/data/labels/o29.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
33 KiB
JSON

{"id": 7588, "in_agent_scope": false, "types": [{"area": "arch/attention", "role": "method", "why": "Dense differentiable communication between LLM nodes trained end-to-end"}], "pattern": "vary-both", "fixes": ["pretrained LLMs"], "varies": ["communication protocol", "topology"], "benchmarks": [], "results": [], "workflow": ["strip LLMs as nodes", "train seq2seq edges", "end-to-end optimize"], "summary": "LMNet connects pretrained LLM nodes via trainable dense-vector edges, enabling end-to-end gradient training of language model networks."}
{"id": 15673, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "survey", "why": "Surveys evaluation methods and benchmarks for LLM-based agents"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "First comprehensive survey of LLM agent evaluation across capabilities, application benchmarks, generalist agents, and frameworks."}
{"id": 9592, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Conformal counterfactual generation over agent-environment causal model with guarantees"}], "pattern": "vary-scaffold", "fixes": ["LLM agent"], "varies": ["counterfactual generation procedure"], "benchmarks": [], "results": [], "workflow": ["model SCM", "sample candidate counterfactuals", "conformal calibration"], "summary": "Conformal counterfactual generation gives guaranteed-coverage sets of what-if outcomes for LLM agent control, shown on wireless network control."}
{"id": 10424, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "method", "why": "Adapts model vocabulary from tool-calling traces to cut decoding latency"}], "pattern": "vary-scaffold", "fixes": ["tool-calling performance"], "varies": ["vocabulary"], "benchmarks": [{"name": "tau-bench", "purpose": "tool-calling performance and latency", "is_new": false}, {"name": "tau2-bench", "purpose": "tool-calling performance and latency", "is_new": false}], "results": [], "workflow": ["mine tool-calling traces", "derive vocabulary entries", "adapt model vocabulary"], "summary": "AgentVocab adds structure-aware vocabulary entries from tool-call traces, preserving tool-calling accuracy while lowering latency."}
{"id": 16652, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Multi-objective RL with query-level safety shaping for search agents"}, {"area": "search-research", "role": "analysis", "why": "Shows search agents are less safe than base LLMs"}], "pattern": "vary-both", "fixes": ["search agent task", "red-teaming data"], "varies": ["reward design", "safety training"], "benchmarks": [], "results": [], "workflow": ["evaluate on red-teaming", "safety/utility reward", "query-level shaping RL"], "summary": "SafeSearch finds search agents less safe than base LLMs and trains them with a multi-objective RL reward that preserves utility."}
{"id": 14299, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Compiles guidelines into fixed review tree for medical record review"}], "pattern": "vary-scaffold", "fixes": ["LLM backbones"], "varies": ["review workflow"], "benchmarks": [], "results": [], "workflow": ["compile guideline tree", "route checks", "verify with full text"], "summary": "GuideTree turns textual guidelines into an offline-induced review tree for long medical records, improving Macro-F1 and cutting cost."}
{"id": 12206, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "analysis", "why": "Audits linguistic diversity of VLA instruction datasets"}], "pattern": "vary-conditions", "fixes": ["datasets"], "varies": ["linguistic dimensions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Dataset audit showing common VLA corpora contain repetitive, template-like instructions with limited linguistic variety."}
{"id": 1176, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "analysis", "why": "Controlled study of CoT reasoning vs memory retrieval in answer generation"}, {"area": "align/rlvr", "role": "method", "why": "Introduces FARL fine-tuning to curb retrieval shortcut"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["misleading cues", "corrupted answers", "fine-tuning type"], "benchmarks": [], "results": [], "workflow": ["perturb reasoning/retrieval", "compare RL vs distill", "FARL fine-tuning"], "summary": "Shows reasoning models mix CoT and retrieval to answer, with RL enabling shortcuts, and proposes FARL fine-tuning."}
{"id": 18536, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Induces procedural instructions from dialogues to prompt TOD agents without fine-tuning"}, {"area": "social-sim", "role": "eval-method", "why": "Interactive evaluation with constrained goal-aware user simulator"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["instructions", "evaluation protocol"], "benchmarks": [], "results": [], "workflow": ["induce instructions", "generate dialogues", "refine with feedback"], "summary": "Training-free task-oriented dialogue via LLM-induced instructions refined iteratively, plus a simulator-based interactive evaluation."}
{"id": 17700, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "analysis", "why": "Compares prompting, SFT, DPO, retrieval hybrid for character dialogue"}], "pattern": "vary-conditions", "fixes": ["QA data"], "varies": ["learning strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "Comparative multi-dimensional evaluation of learning strategies for character-authentic dialogue, finding retrieval-plus-DPO hybrid strongest."}
{"id": 8757, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Test-time prompt tuning with stability and suitability scores for robust CLIP"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["adaptation method"], "benchmarks": [], "results": [], "workflow": ["score augmented views", "guided consistency loss", "weighted prediction"], "summary": "SS-TPT weights augmented views by stability and suitability to improve adversarial robustness of CLIP with less slowdown."}
{"id": 14053, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "application", "why": "Open-source modular framework for production multi-agent systems"}], "pattern": "n/a", "fixes": [], "varies": ["framework abstraction"], "benchmarks": [], "results": [], "workflow": ["unified Oxy abstraction", "runtime execution graphs", "OxyBank evolution"], "summary": "OxyGent is an open-source multi-agent framework with unified component abstraction, observability, and evolution engine."}
{"id": 4252, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Speculative reward-guided decoding approximating soft best-of-n"}], "pattern": "vary-scaffold", "fixes": ["base model"], "varies": ["decoding algorithm"], "benchmarks": [{"name": "MATH500", "purpose": "accuracy", "is_new": false}, {"name": "OlympiadBench", "purpose": "accuracy", "is_new": false}, {"name": "Minerva Math", "purpose": "accuracy", "is_new": false}, {"name": "MMLU-STEM", "purpose": "accuracy", "is_new": false}, {"name": "GSM8K", "purpose": "accuracy", "is_new": false}], "results": [], "workflow": ["small-model speculative samples", "reward scoring", "soft best-of-n"], "summary": "Guided Speculative Inference gives provable efficient reward-guided decoding, improving accuracy and cutting latency up to 28%."}
{"id": 1271, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Debate agents mask memories from prior round to resist wrong reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["memory selection"], "benchmarks": [], "results": [], "workflow": ["debate rounds", "select memories", "mask bad memories"], "summary": "MAD-M2 improves multi-agent debate robustness by letting agents mask unreliable previous-round memories, with theory on memory quality."}
{"id": 3085, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Self-evolving policy and generative process reward model for OR modeling"}, {"area": "align/pref", "role": "method", "why": "Weighted DPO with generative process supervision"}], "pattern": "vary-data", "fixes": ["OR problems"], "varies": ["reward", "training loop"], "benchmarks": [], "results": [], "workflow": ["solver verification", "GenPRM evaluation", "weighted DPO"], "summary": "StepORLM co-evolves a policy and generative process reward model with solver feedback for operations research LLMs."}
{"id": 15753, "in_agent_scope": false, "types": [{"area": "train/data", "role": "method", "why": "Adversarially evolves constraint-graph blueprints to synthesize math reasoning data"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": ["evolve constraint graph", "instantiate scenarios", "train on data"], "summary": "MathAgent synthesizes complex math data via adversarial constraint graph evolution and semantic instantiation, tested on ten models."}
{"id": 3657, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Document expansion pipeline enriching tool docs for retrieval"}, {"area": "tool-use", "role": "benchmark", "why": "Tool-DE benchmark for tool retrieval"}], "pattern": "vary-data", "fixes": ["retrieval models"], "varies": ["tool documentation"], "benchmarks": [{"name": "ToolRet", "purpose": "tool retrieval evaluation", "is_new": false}, {"name": "Tool-DE", "purpose": "tool retrieval with enriched docs", "is_new": true}], "results": [], "workflow": ["generate enriched profiles", "validate and refine", "train Tool-Embed/Rank"], "summary": "Tool-DE expands tool documentation with structured fields and trains Tool-Embed and Tool-Rank, reaching state-of-the-art tool retrieval."}
{"id": 5219, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Fast-slow planning with A* and mined subroutines for image editing toolpaths"}], "pattern": "vary-scaffold", "fixes": ["LLM", "image tools"], "varies": ["planning procedure"], "benchmarks": [], "results": [], "workflow": ["LLM subtask planning", "A* toolpath search", "mine subroutines"], "summary": "FaSTA* combines LLM planning, A* search, and reusable mined subroutines for cost-efficient multi-turn image editing."}
{"id": 14854, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Dual-phase tree search separating planning and execution with dynamic budget"}], "pattern": "vary-scaffold", "fixes": ["LLM", "reward model"], "varies": ["search procedure"], "benchmarks": [], "results": [], "workflow": ["separate plan/execute", "search each phase", "reallocate budget"], "summary": "Test-time scaling framework searching planning and execution separately with reward-driven adaptive budget allocation."}
{"id": 14699, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Triton dataset with hard negative mining and consensus synthesis"}, {"area": "web-gui", "role": "train-agent", "why": "Progressive SFT, ORPO, GRPO curriculum for web agents"}], "pattern": "vary-data", "fixes": ["web navigation task"], "varies": ["training data", "training stage"], "benchmarks": [], "results": [], "workflow": ["mine hard negatives", "synthesize tasks", "SFT-ORPO-GRPO curriculum"], "summary": "Triton dataset and progressive SFT/ORPO/GRPO curriculum yield robust text-based web navigation agents."}
{"id": 16848, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Attention-guided feature amplification and hybrid decoding for captioning"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["captioning method"], "benchmarks": [{"name": "Flickr8k", "purpose": "captioning", "is_new": false}, {"name": "Flickr30k", "purpose": "captioning", "is_new": false}, {"name": "MSCOCO", "purpose": "captioning", "is_new": false}], "results": [], "workflow": ["amplify salient regions", "hybrid decoding"], "summary": "AGIC amplifies salient visual features and uses hybrid decoding for relevant, faster image captioning."}
{"id": 15167, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Agent-driven clustering of memory with cluster-targeted retrieval for small models"}], "pattern": "vary-scaffold", "fixes": ["SLM backbones"], "varies": ["memory organization"], "benchmarks": [], "results": [], "workflow": ["route to cluster", "evolve within cluster", "retrieve relevant clusters"], "summary": "CLAG organizes agent memory into clusters via an SLM router, reducing interference and improving QA for small models."}
{"id": 15709, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Reflective collaborating text and citation agents for attributed summaries"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [{"name": "ALCE", "purpose": "attributed summary correctness and citation", "is_new": false}], "results": [], "workflow": ["summarize", "reflect on consistency", "recalibrate inputs"], "summary": "RAAC uses reflective TextAgent and CitationAgent collaboration to improve factuality and citation quality in attributed summarization."}
{"id": 5528, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "Thumbnail reasoning then high-res crop answering for grounded visual reasoning"}], "pattern": "vary-both", "fixes": ["MLLM task"], "varies": ["reasoning paradigm", "training"], "benchmarks": [], "results": [], "workflow": ["reason on thumbnail", "crop high-res", "answer"], "summary": "RTAF reasons over low-resolution thumbnails to pick regions, then answers from high-resolution crops, cutting grounded reasoning cost."}
{"id": 15023, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Suffix-restart curriculum replay from successful trajectories for GRPO"}], "pattern": "fix-env-vary-agent", "fixes": ["environments", "GRPO"], "varies": ["rollout curriculum"], "benchmarks": [{"name": "ALFWorld", "purpose": "long-horizon RL training eval", "is_new": false}, {"name": "WebShop", "purpose": "long-horizon RL training eval", "is_new": false}], "results": [], "workflow": ["buffer successes", "restart from suffix", "adapt k"], "summary": "CRPS replays progressive suffixes of successful trajectories to fix degenerate GRPO groups for long-horizon agents."}
{"id": 1478, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Trains StateLM to manage its own context via memory tools"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["model state management"], "benchmarks": [], "results": [], "workflow": ["index/prune/note tools", "train management loop"], "summary": "StateLM is a model trained to manipulate its own context with memory tools, decoupling performance from context window size."}
{"id": 17540, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "Dutch legal QA dataset with attributions and formalizations for RAG"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["RAG configuration"], "benchmarks": [{"name": "QuALA-NL", "purpose": "legal RAG with attribution", "is_new": true}], "results": [], "workflow": [], "summary": "QuALA-NL provides 101 Dutch legal QA pairs with attributions, used to compare LLM-only, RAG, and knowledge-based RAG."}
{"id": 15816, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "Progressive argumentation-mining diagnostic isolating LLM judgment from debate noise"}], "pattern": "vary-evaluation", "fixes": ["models"], "varies": ["evaluation protocol"], "benchmarks": [], "results": [], "workflow": ["distill best rationale", "control context", "diagnose judgment"], "summary": "A diagnostic framework controlling debate context to test whether LLMs truly discriminate valid reasoning from plausible errors."}
{"id": 10867, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Two-stage RL trains proactive intent-clarifying deep research agents"}, {"area": "search-research", "role": "data", "why": "Seed-expansion pipeline building intent dialogue data"}], "pattern": "fix-env-vary-agent", "fixes": ["deep research task"], "varies": ["training data", "RL stages"], "benchmarks": [], "results": [], "workflow": ["expand seeds to dialogues", "offline RL", "online RL with simulator"], "summary": "IntentRL trains deep research agents to clarify user intent first, using synthesized dialogues and two-stage RL."}
{"id": 16277, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "M-SAEA probes audit multi-agent LLM teams for risk"}, {"area": "multi-agent", "role": "benchmark", "why": "Risk-first evaluation of multi-agent finance tasks"}], "pattern": "vary-evaluation", "fixes": ["six models"], "varies": ["probes", "risk metrics"], "benchmarks": [], "results": [], "workflow": ["deploy ten probes", "aggregate risk vector"], "summary": "M-SAEA audits LLM teams in finance with probes across four layers, yielding risk vectors beyond standard task metrics."}
{"id": 4186, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Training-free sparse attention estimating scores via representative heads"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["attention estimation"], "benchmarks": [], "results": [], "workflow": ["pool representative heads", "estimate token importance", "block-aware sparsity"], "summary": "ProxyAttn speeds long-context pre-filling with token-level sparse attention approximated from pooled representative heads."}
{"id": 17802, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "benchmark", "why": "Japanese LLM safety QA dataset reflecting local context"}, {"area": "train/sft", "role": "data", "why": "Dataset used to fine-tune Japanese LLM for safety"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["safety data"], "benchmarks": [{"name": "JLLMSafety", "purpose": "Japanese LLM safety evaluation and tuning", "is_new": true}], "results": [], "workflow": [], "summary": "JLLMSafety is a 1,800-pair Japanese safety dataset used to fine-tune and benchmark 12 Japanese LLMs."}
{"id": 8434, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Text-guided video compression with Transformer-CNN block and entropy model"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["codec architecture"], "benchmarks": [], "results": [], "workflow": ["text fusion", "history-conditioned coding", "entropy model"], "summary": "MoVie is a multimodal video codec using text guidance and history-conditioned coding for better perceptual quality."}
{"id": 14551, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "environment", "why": "AVACraft multimodal StarCraft II environment for MARL and VLM agents"}, {"area": "embodied-nav", "role": "benchmark", "why": "21 scenarios with standardized protocols"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["scenarios", "observation modalities"], "benchmarks": [{"name": "AVACraft", "purpose": "StarCraft II decision-making benchmark", "is_new": true}, {"name": "SMAC", "purpose": "prior environment compared", "is_new": false}], "results": [], "workflow": [], "summary": "AVACraft is a multimodal StarCraft II benchmark environment supporting MARL and VLM decision-makers, with an attentive VLM agent baseline."}
{"id": 7126, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Planner-Coder-Critic agents generate educational videos via code"}, {"area": "domain-agents", "role": "benchmark", "why": "MMMC benchmark of professional educational videos"}], "pattern": "fix-agent-vary-env", "fixes": ["video domain"], "varies": ["agent pipeline"], "benchmarks": [{"name": "MMMC", "purpose": "educational video generation evaluation", "is_new": true}], "results": [], "workflow": ["plan storyboard", "write code", "critic refines layout"], "summary": "Code2Video generates educational videos through code-writing planner, coder, and critic agents, evaluated on new benchmark MMMC."}
{"id": 14326, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent error-driven refinement of class descriptions for classification prompts"}], "pattern": "vary-scaffold", "fixes": ["LLM", "instructions"], "varies": ["class descriptions"], "benchmarks": [], "results": [], "workflow": ["classify", "analyze errors", "refine descriptions"], "summary": "Multi-agent framework iteratively refines class descriptions from classification errors, targeting contact-center labeling."}
{"id": 10606, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "theory", "why": "Game-theoretic framework with Nash equilibrium for answer-level fine-tuning"}], "pattern": "n/a", "fixes": ["answer-level objective"], "varies": ["assumptions", "algorithm"], "benchmarks": [], "results": [], "workflow": ["policy-target game", "optimize jointly"], "summary": "Distributional Alignment Games formulate answer-level fine-tuning as a two-player game whose equilibrium solves the marginalized objective."}
{"id": 12217, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Multi-horizon preference optimization of tutor agents in simulated classrooms"}], "pattern": "fix-env-vary-agent", "fixes": ["simulation environment"], "varies": ["reward", "preference optimization"], "benchmarks": [], "results": [], "workflow": ["simulate student", "multi-horizon reward", "iterative preference tuning"], "summary": "MHPO trains Socratic tutor agents with turn-level and trajectory-level rewards in teacher-student simulation."}
{"id": 11435, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Plug-in safety model correcting unsafe agent thoughts before action"}], "pattern": "fix-agent-vary-env", "fixes": ["agent", "threat model"], "varies": ["thought-level guardrail"], "benchmarks": [], "results": [], "workflow": ["detect unsafe thought", "correct thought", "feed back to agent"], "summary": "Thought-Aligner is a lightweight model-agnostic plug-in that corrects unsafe agent thoughts before execution."}
{"id": 12680, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "RL-trained unified long and short-term memory tool actions"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["memory policy training"], "benchmarks": [], "results": [], "workflow": ["memory ops as tools", "three-stage RL", "step-wise GRPO"], "summary": "AgeMem exposes memory operations as tools and trains agents with progressive RL to manage long- and short-term memory jointly."}
{"id": 15105, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Code Reffix evaluates reflection generation separate from code repair"}], "pattern": "fix-agent-vary-env", "fixes": ["14 LLMs"], "varies": ["tasks", "dual-task protocol"], "benchmarks": [{"name": "Code Reffix", "purpose": "reflection-guided code repair", "is_new": true}], "results": [], "workflow": ["oracle reflections", "dual-task protocol"], "summary": "Code Reffix benchmarks reflection quality and repair separately across 14 LLMs, showing small models' reflection is the bottleneck."}
{"id": 8762, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Faithfulness SFT and RL with guided advantage estimator on GRPO"}], "pattern": "fix-env-vary-agent", "fixes": ["GUI tasks"], "varies": ["training objective", "advantage estimator"], "benchmarks": [], "results": [], "workflow": ["faithfulness SFT", "RFT with GuAE"], "summary": "Faithful-Agent trains mobile GUI agents to ground actions in screen evidence via faithfulness SFT and GuAE-tempered GRPO."}
{"id": 7096, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Schema-guided world model for visual dynamics prediction and simulation"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["world model"], "benchmarks": [], "results": [], "workflow": ["state-transition sequences", "MoE modeling"], "summary": "DynaVieW learns interleaved state-transition sequences with hierarchical schema to predict and simulate visual dynamics."}
{"id": 3627, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Lightweight dual-stream CNN-Transformer for 3D medical segmentation"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["architecture"], "benchmarks": [], "results": [], "workflow": ["window attention", "JL-guided convolution"], "summary": "VeloxSeg is an efficient 3D medical segmentation network using paired window attention and JL-lemma-guided convolution."}
{"id": 18362, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "Evaluates fidelity of LLM persona agents and persuasion dynamics"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["persona", "rhetorical strategy"], "benchmarks": [], "results": [], "workflow": ["compare personas to human data", "simulate 1080 dialogues"], "summary": "Dual-level evaluation of LLM agents simulating political personas and persuasion, finding reduced variability and stereotypes."}
{"id": 13423, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Answer-guided GRPO rollouts for masked diffusion LMs"}, {"area": "align/rlvr", "role": "method", "why": "Adds answer-guided rollouts to RLVR"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["answer-free rollouts", "answer-guided rollouts", "re-predict answer"], "summary": "AG-GRPO adds ground-truth-suffix-guided rollouts to GRPO for diffusion language models, improving reasoning-answer consistency."}
{"id": 16367, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "RL-trained planner selects reasoning actions for frozen executor"}], "pattern": "vary-scaffold", "fixes": ["executor LLM"], "varies": ["planner"], "benchmarks": [], "results": [], "workflow": ["select action", "execute step", "logit mask"], "summary": "PG-HAP trains an RL planner to choose high-level reasoning actions for a frozen LLM, with masks preventing degenerate loops."}
{"id": 1291, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "ResearchRubrics prompts and expert rubrics for deep research agents"}], "pattern": "fix-agent-vary-env", "fixes": ["DR agents"], "varies": ["prompts", "rubrics"], "benchmarks": [{"name": "ResearchRubrics", "purpose": "deep research evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "ResearchRubrics pairs realistic prompts with 2,500+ expert rubrics and a complexity framework to evaluate deep research agents."}
{"id": 5135, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "OpenAgentSafety evaluates agents with real tools across eight risk categories"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks", "risk categories"], "benchmarks": [{"name": "OpenAgentSafety", "purpose": "real-tool agent safety evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Modular framework with 350+ multi-turn tasks evaluating real-tool agent safety under benign and adversarial intents."}
{"id": 5057, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Studies greediness, frequency bias, knowing-doing gap in LLM decision-making"}, {"area": "planning-decision", "role": "train-agent", "why": "RL fine-tuning on self-generated CoT rationales"}], "pattern": "vary-conditions", "fixes": ["bandits and Tic-tac-toe"], "varies": ["model", "RL fine-tuning"], "benchmarks": [], "results": [], "workflow": ["diagnose failure modes", "RL on CoT"], "summary": "Identifies failure modes of LLM decision-making and mitigates them with RL fine-tuning on self-generated rationales."}
{"id": 12188, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Type-constrained KGQA over ontology graph with planner-judge"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["KGQA framework"], "benchmarks": [], "results": [], "workflow": ["type planner", "judge verifies", "generative backoff"], "summary": "OntGQA constrains KGQA reasoning using a relation-centric ontology graph with planner, judge, and generator."}
{"id": 12248, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Bipartite matching turn-level rewards for tool-integrated reasoning RL"}], "pattern": "fix-env-vary-agent", "fixes": ["tool tasks"], "varies": ["credit assignment"], "benchmarks": [], "results": [], "workflow": ["match predicted to truth", "turn-level reward", "dual-level advantage"], "summary": "MatchTIR gives fine-grained turn-level credit for tool calls via bipartite matching in RL."}
{"id": 5413, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "SciNet relation-aware scientific literature retrieval dataset for agents"}], "pattern": "fix-agent-vary-env", "fixes": ["retrieval agents"], "varies": ["relational tasks"], "benchmarks": [{"name": "SciNet", "purpose": "relation-aware literature retrieval", "is_new": true}], "results": [], "workflow": [], "summary": "SciNet is a relation-aware scientific paper retrieval dataset over 269M papers for evaluating information retrieval agents."}
{"id": 5536, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Fully connected coordinator of specialized agents for engineering problem solving"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": ["analyze", "model", "verify and solve"], "summary": "EngiAgent coordinates specialized LLM agents to produce feasible solutions for open-ended engineering problems."}
{"id": 12158, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Trains end-to-end spoken dialogue model with agentic tool use"}, {"area": "tool-use", "role": "data", "why": "470-hour AgentChat dataset"}], "pattern": "fix-env-vary-agent", "fixes": ["spoken dialogue"], "varies": ["training data", "think-before-speak"], "benchmarks": [{"name": "AgentChat", "purpose": "spoken agentic training data", "is_new": true}], "results": [], "workflow": ["think before speak", "tool use"], "summary": "VoxMind equips spoken dialogue models with tool use and think-before-speak reasoning using the AgentChat dataset."}
{"id": 14961, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Trains structured reflection as learnable error-repair action for tool calls"}], "pattern": "fix-env-vary-agent", "fixes": ["tool tasks"], "varies": ["reflection training"], "benchmarks": [], "results": [], "workflow": ["diagnose error", "repair call", "RL train"], "summary": "Structured reflection treats error-to-repair as trainable action, improving recovery in multi-turn tool interaction."}
{"id": 7927, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Non-portable obfuscated prompts protecting agent prompt IP"}], "pattern": "fix-agent-vary-env", "fixes": ["agent systems"], "varies": ["prompt protection"], "benchmarks": [], "results": [], "workflow": ["anchor semantics with code", "inject noise via feedback"], "summary": "PragLocker obfuscates agent prompts so they only work on the target LLM, preventing reuse."}
{"id": 16567, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "analysis", "why": "Linguistic taxonomy for analyzing prompts, with refinement applications"}], "pattern": "n/a", "fixes": [], "varies": ["taxonomy"], "benchmarks": [], "results": [], "workflow": [], "summary": "PromptPrism taxonomy analyzes prompts by structure, semantics, and syntax, with applications such as prompt refinement."}
{"id": 14656, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "RL with single-turn anchors stabilizes multi-turn interaction"}, {"area": "reason/cot", "role": "analysis", "why": "Identifies contextual inertia in multi-turn LLMs"}], "pattern": "fix-env-vary-agent", "fixes": ["multi-turn tasks"], "varies": ["training approach"], "benchmarks": [], "results": [], "workflow": ["single-turn anchors", "RL training"], "summary": "RLSTA addresses contextual inertia in multi-turn interaction using single-turn anchor rewards in RL."}
{"id": 4011, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "PACEbench of realistic cyber-exploitation scenarios"}, {"area": "agent-safety", "role": "attack", "why": "PACEagent penetration-testing agent"}], "pattern": "fix-agent-vary-env", "fixes": ["seven frontier LLMs"], "varies": ["scenarios"], "benchmarks": [{"name": "PACEbench", "purpose": "cyber-exploitation evaluation", "is_new": true}], "results": [], "workflow": ["reconnaissance", "analysis", "exploitation"], "summary": "PACEbench evaluates LLM cyber-exploitation across four scenarios, with PACEagent emulating human penetration testers."}