Files
ai-taxonomy-literature-anal…/data/labels/o34.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
32 KiB
JSON

{"id": 12345, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Bandit-based start-URL selection with multi-agent navigation and episodic memory, training-free"}], "pattern": "vary-scaffold", "fixes": ["base LLM", "WebVoyager tasks"], "varies": ["start URL selection", "budget allocation", "episodic memory"], "benchmarks": [{"name": "WebVoyager", "purpose": "evaluate navigation success", "is_new": false}], "results": [{"benchmark": "WebVoyager", "system": "Mango", "base_model": "GPT-5-mini", "setting": "proposed", "metric": "success rate", "score": 63.6, "unit": "%"}], "workflow": ["Select start URLs via Thompson Sampling", "Allocate budget across URLs", "Store episodic navigation history"], "summary": "Multi-agent web navigation choosing starting URLs with a multi-armed bandit plus episodic memory to improve success."}
{"id": 1429, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Planner-Composer-Evaluator turns LLM reasoning assumptions into scored decision tree for uncertainty-aware planning"}], "pattern": "vary-scaffold", "fixes": ["LLM", "embodied environments"], "varies": ["planning framework"], "benchmarks": [], "results": [], "workflow": ["Extract assumptions from reasoning", "Build decision tree", "Score paths, pick action"], "summary": "PCE converts LLM reasoning assumptions into a decision tree for uncertainty-aware multi-agent embodied planning."}
{"id": 17851, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Portuguese discursive QA benchmark from Brazilian university exams, baselines on nine models"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated models"], "varies": ["tasks"], "benchmarks": [{"name": "EduBench", "purpose": "Portuguese open-ended QA", "is_new": true}], "results": [], "workflow": [], "summary": "Portuguese benchmark of 3,149 discursive entrance-exam questions with reference answers, evaluating nine LLMs."}
{"id": 1471, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL with hierarchical knowledge-grounded process rewards against over- and under-search"}], "pattern": "fix-env-vary-agent", "fixes": ["QA benchmarks"], "varies": ["reward design"], "benchmarks": [], "results": [], "workflow": ["Detect search steps", "Assess necessity of each search", "Add process reward to RL"], "summary": "HiPRAG adds fine-grained process rewards to RL training of agentic RAG to curb over-search and under-search."}
{"id": 8643, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Training-free hierarchical global/local working memory for small LM agents"}], "pattern": "vary-scaffold", "fixes": ["small LM", "tasks"], "varies": ["memory structure"], "benchmarks": [], "results": [], "workflow": ["Decompose into subgoals", "Keep global task memory", "Keep local subtask memory"], "summary": "LightWM is a training-free hierarchical working memory for small-model agents on long-horizon procedural tasks."}
{"id": 11818, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Hierarchical reasoning agent and search agent coordinating for multi-hop claim verification"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent structure"], "benchmarks": [], "results": [], "workflow": ["Reasoning agent builds chain", "Search agent retrieves evidence", "Refine queries iteratively"], "summary": "HARIS coordinates a high-level reasoning agent and low-level search agent for multi-hop claim verification."}
{"id": 13038, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "LLM agent simulation of prehistoric society manipulating moral type observability and communication"}], "pattern": "vary-conditions", "fixes": ["agent framework"], "varies": ["moral dispositions", "observability", "communication bandwidth"], "benchmarks": [], "results": [], "workflow": ["Simulate hunter-gatherer society", "Vary settings", "Observe emergent mechanisms"], "summary": "LLM agent simulation to study moral evolution, finding cooperation and mutual help central to survival."}
{"id": 4125, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Self-evolving agentic RL for alpha factor mining"}], "pattern": "fix-env-vary-agent", "fixes": ["alpha mining task"], "varies": ["agent training"], "benchmarks": [], "results": [], "workflow": ["Generate alphas", "Evaluate and reflect", "Train with agentic RL"], "summary": "AlphaAgentEvo trains an agent with self-evolving agentic RL for systematic alpha factor mining."}
{"id": 11045, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "data", "why": "Execution-grounded Docker-building trajectories to train builder model"}, {"area": "code-swe", "role": "environment", "why": "Agentic Docker builder constructs scalable coding environments"}], "pattern": "fix-env-vary-agent", "fixes": ["training recipe"], "varies": ["training trajectories"], "benchmarks": [{"name": "Multi-Docker-Eval", "purpose": "evaluate Docker building", "is_new": false}], "results": [{"benchmark": "Multi-Docker-Eval", "system": "DockSmith", "base_model": "30B-A3B", "setting": "finetuned", "metric": "Fail-to-Pass", "score": 39.72, "unit": "%"}, {"benchmark": "Multi-Docker-Eval", "system": "DockSmith", "base_model": "30B-A3B", "setting": "finetuned", "metric": "Commit Rate", "score": 58.28, "unit": "%"}], "workflow": ["Generate build trajectories", "Loop detection and success memory", "Train 30B-A3B"], "summary": "DockSmith trains an agentic Docker builder on synthesized trajectories to scale reliable SWE environments."}
{"id": 1857, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "environment", "why": "Time-accelerated Matter-based smart home simulator"}, {"area": "domain-agents", "role": "benchmark", "why": "Smart home agent benchmark with temporal and device constraints"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated LLM agents"], "varies": ["tasks", "simulated environment"], "benchmarks": [{"name": "SimuHome", "purpose": "smart home agent evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "SimuHome is a Matter-protocol smart home simulator and benchmark for LLM agents."}
{"id": 14714, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "Supply chain benchmark of domain knowledge and SOP-grounded tool orchestration"}, {"area": "tool-use", "role": "scaffold", "why": "SupChain-ReAct synthesizes executable procedures without SOPs"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "SupChain-Bench", "purpose": "supply chain evaluation", "is_new": true}], "results": [], "workflow": ["Synthesize procedures", "Execute tools"], "summary": "SupChain-Bench evaluates LLMs on supply chain knowledge and tool orchestration, plus SupChain-ReAct baseline."}
{"id": 6972, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Entropy-based metric HE-SNR guiding mid-training for SWE performance"}], "pattern": "vary-evaluation", "fixes": ["models"], "varies": ["metric"], "benchmarks": [{"name": "SWE-bench", "purpose": "downstream SWE performance", "is_new": false}], "results": [], "workflow": ["Filter data", "Entropy analysis", "Define metric"], "summary": "HE-SNR, an entropy-based metric to guide mid-training for software engineering ability."}
{"id": 16765, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "data", "why": "Multi-agent feedback loop to augment data for domain NER"}], "pattern": "vary-data", "fixes": ["NER model"], "varies": ["augmentation"], "benchmarks": [{"name": "SciERC", "purpose": "NER augmentation", "is_new": false}, {"name": "NCBI-disease", "purpose": "NER augmentation", "is_new": false}], "results": [], "workflow": ["Abstract sentences", "Refine guidance", "Calibrate samples"], "summary": "R-GDA uses reflective multi-agent feedback to improve LLM data augmentation for domain NER."}
{"id": 10545, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Empowerment-maximizing fine-tuning for assistive coding LMs, with user study"}], "pattern": "fix-env-vary-agent", "fixes": ["user study setting"], "varies": ["training objective"], "benchmarks": [], "results": [], "workflow": ["Self-supervised empowerment objective", "Fine-tune LM", "User study"], "summary": "Empower fine-tunes assistive language models to maximize human empowerment using offline text only."}
{"id": 14645, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Cognitive-diagnosis guided assignment of LLM backbones to simulate students by ability"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["LLM assignment"], "benchmarks": [], "results": [], "workflow": ["Extract ability with NeuralCD", "Match LLM to student", "Simulate behavior"], "summary": "Ability-aware student simulation matching LLM backbones to students via cognitive diagnosis, reducing ability bias."}
{"id": 6247, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Orchestrator spawns sub-agents from (Model, Task, Tools, Context) tuple"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["orchestration"], "benchmarks": [], "results": [], "workflow": ["Curate context", "Select model and tools", "Spawn sub-agent"], "summary": "AOrchestra dynamically creates specialized sub-agents via a unified tuple abstraction."}
{"id": 11982, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "Interpretability-guided data selection via SAE task features"}], "pattern": "vary-data", "fixes": ["fine-tuning recipe"], "varies": ["selected data"], "benchmarks": [], "results": [], "workflow": ["Identify causal task features", "Select resonant data", "Fine-tune"], "summary": "IGDS selects fine-tuning data that maximally activates SAE-identified task features, improving data efficiency."}
{"id": 13756, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Token-level RLVR updating only mutual-information-critical tokens"}], "pattern": "vary-both", "fixes": ["benchmarks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Score tokens by mutual information", "Update critical tokens", "Blend info gain and correctness"], "summary": "AIPO focuses RLVR policy gradients on high-information tokens, improving reasoning with 10% token updates."}
{"id": 10723, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Repository-level benchmark for proactive bug discovery via test generation"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "TestExplora", "purpose": "proactive bug discovery", "is_new": true}], "results": [], "workflow": [], "summary": "TestExplora benchmarks LLMs on 2,389 repo-level tasks for proactive bug discovery through test generation."}
{"id": 12687, "in_agent_scope": true, "types": [{"area": "search-research", "role": "data", "why": "Graph-based pruning of web agent trajectories for continued training"}], "pattern": "vary-data", "fixes": ["agent training recipe"], "varies": ["trajectory data"], "benchmarks": [], "results": [], "workflow": ["Model search as state graph", "Mine minimal DAG", "Continue training"], "summary": "WebClipper prunes redundant web-agent trajectories, cutting tool calls about 20% while improving accuracy."}
{"id": 401, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Test-time external module plants reasoning to help conversational agents recover from errors"}], "pattern": "vary-scaffold", "fixes": ["model weights", "prompts"], "varies": ["inception module"], "benchmarks": [], "results": [], "workflow": ["Diagnose flawed context", "Plant recovery reasoning", "Agent acts"], "summary": "ReIn is a test-time intervention that injects reasoning to help tool-using dialogue agents recover from user-induced errors."}
{"id": 3047, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Transformer network with proposal-centric modules for LiDAR 3D detection"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["detector architecture"], "benchmarks": [], "results": [], "workflow": ["Align proposal features", "Collaborative refinement"], "summary": "PTNet improves two-stage LiDAR 3D object detection with proposal-centric transformer modules."}
{"id": 15963, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Multi-agent comment generation and example retrieval for code refactoring"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["multi-agent framework"], "benchmarks": [], "results": [], "workflow": ["Generate comments", "Retrieve expert examples", "Iterative feedback validation"], "summary": "MACOR uses multi-agent comment-guided refactoring with retrieval and feedback loop."}
{"id": 11024, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "environment", "why": "Stateful tool-value cache speeding up agent RL rollouts"}], "pattern": "fix-env-vary-agent", "fixes": ["RL training"], "varies": ["tool execution caching"], "benchmarks": [], "results": [], "workflow": ["Tree of tool sequences", "Longest-prefix match", "Reuse cached outputs"], "summary": "TVCache caches tool outputs by tool-call history to cut RL post-training time for agents."}
{"id": 15565, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Semi-online RL integrating trajectory-level signals into offline GUI agent training"}], "pattern": "fix-env-vary-agent", "fixes": ["GUI tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Reconstruct rollout candidates", "Detect first failure", "Assign long-horizon reward"], "summary": "SOLAR-RL brings global trajectory insights into offline RL for GUI agents."}
{"id": 6363, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Structured memory construction and reasoning agent for advertising video understanding"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": ["Build structured video memory", "Structured reasoning agent"], "summary": "AD-MIR builds structured memory and a reasoning agent to decode advertising intent in videos."}
{"id": 13700, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "environment", "why": "Environment creator with 5,571 tools, task synthesis, state controller and rewards for computer-use agents"}, {"area": "web-gui", "role": "environment", "why": "Synthesized world engine approximates tool behavior"}], "pattern": "fix-agent-vary-env", "fixes": ["agent"], "varies": ["environment"], "benchmarks": [], "results": [], "workflow": ["Define action space", "Synthesize tasks", "Simulate transitions and rewards"], "summary": "C-World creates scalable computer-use agent environments with tools, tasks, failures and rewards."}
{"id": 17308, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Eight-sub-agent real estate assistant with RAG and tools"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": ["Coordinate sub-agents", "Retrieve knowledge", "Execute calculations"], "summary": "PropGenie is a multi-agent conversational system for real estate assistance."}
{"id": 16067, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Stage-wise test-time scaling with process judgers in a data science pipeline"}], "pattern": "vary-scaffold", "fixes": ["model"], "varies": ["inference compute allocation"], "benchmarks": [], "results": [], "workflow": ["Distribute compute across stages", "Prune with judgers"], "summary": "Selective TTS scales inference across multi-agent pipeline stages, pruning early to generate visual insights."}
{"id": 10133, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Program-aware scheduling system for agentic workloads managing KV cache and tool resources"}], "pattern": "n/a", "fixes": ["models"], "varies": ["serving system"], "benchmarks": [], "results": [], "workflow": ["Abstract workflows as LLM programs", "Unified resource scheduling"], "summary": "ThunderAgent is a program-aware inference system improving KV cache and tool environment management for agent workflows."}
{"id": 15080, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Personalized QA benchmark from data-mined cognitive factors, 1,985 user profiles"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated models"], "varies": ["tasks"], "benchmarks": [{"name": "CoPA", "purpose": "personalized QA", "is_new": true}], "results": [], "workflow": ["Mine preference divergence", "Distill six factors"], "summary": "CoPA benchmarks personalized QA with six data-informed cognitive factors."}
{"id": 2675, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Unsupervised meta-RL with self-imposed goals for efficient exploration"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["pretraining method"], "benchmarks": [], "results": [], "workflow": ["Generate goals", "Meta-learn exploration", "Curriculum by estimated performance"], "summary": "ULEE is an unsupervised meta-learning method pre-training adaptive exploration policies via self-set goals."}
{"id": 12448, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Temporal property-graph memory with multi-tool retrieval agent"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory system"], "benchmarks": [{"name": "LOCOMO", "purpose": "long-term memory QA", "is_new": false}, {"name": "LongMemEval", "purpose": "long-term memory QA", "is_new": false}], "results": [{"benchmark": "LOCOMO", "system": "APEX-MEM", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 88.88, "unit": "%"}, {"benchmark": "LongMemEval", "system": "APEX-MEM", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 86.2, "unit": "%"}], "workflow": ["Build property graph", "Append-only storage", "Retrieval agent resolves conflicts"], "summary": "APEX-MEM structures conversations in a temporal graph with agentic retrieval for long-term memory."}
{"id": 13661, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Turn-level reward RL algorithm for multi-turn tool-integrated reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["TIR tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Turn-level rewards", "Return-based advantage", "Self-supervised reward shaping"], "summary": "GTPO gives turn-level rewards and returns for RL training of multi-turn tool-integrated reasoning."}
{"id": 14032, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "application", "why": "Open-source annotation platform with AI assistance and agent trace support"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [{"name": "Potato 2.0", "purpose": "annotation platform", "is_new": true}], "results": [], "workflow": [], "summary": "Potato 2.0 is an annotation platform supporting many task types, agentic traces and AI-assisted labeling."}
{"id": 13345, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Hierarchical agentic reasoning with preference optimization for conversational recommendation"}], "pattern": "fix-env-vary-agent", "fixes": ["CRS datasets"], "varies": ["training objective"], "benchmarks": [], "results": [], "workflow": ["Decompose quality dimensions", "Hierarchical reasoning", "Preference optimization"], "summary": "HARPO optimizes agentic conversational recommenders for multi-dimensional recommendation quality."}
{"id": 181, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "LLM-generated reward shaping with affordance-based exploration bonuses for robot skills"}], "pattern": "n/a", "fixes": ["robot tasks"], "varies": ["reward and exploration"], "benchmarks": [], "results": [], "workflow": ["LLM generates rewards", "Build affordance space", "Exploration bonuses"], "summary": "PoRSE combines LLM reward design with task-aware exploration for robot skill RL."}
{"id": 12833, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Paper-fabrication agents tested against LLM reviewers"}, {"area": "agent-safety", "role": "defense", "why": "Mitigation strategies for reviewer vulnerabilities"}], "pattern": "vary-conditions", "fixes": ["LLM reviewers"], "varies": ["fabrication strategy"], "benchmarks": [], "results": [], "workflow": ["Generate fabricated papers", "Review with LLMs", "Measure acceptance"], "summary": "BadScientist shows fabricated papers fool multi-model LLM reviewers, reaching up to 82% acceptance."}
{"id": 8047, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "RL with differential reward to learn when and which audio tools to invoke"}], "pattern": "fix-env-vary-agent", "fixes": ["audio benchmarks"], "varies": ["training reward"], "benchmarks": [], "results": [], "workflow": ["Select tools", "Differential reward", "RL training"], "summary": "AuTAgent learns selective tool invocation for audio reasoning via RL."}
{"id": 14432, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Experiential priors and narrative paths for multi-hop KGQA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["reasoning framework"], "benchmarks": [], "results": [], "workflow": ["Narrate reasoning paths", "Abstract exploration priors", "Dual-feedback re-ranking"], "summary": "TRACE improves multi-hop KGQA using path narratives and reusable exploration priors."}
{"id": 17905, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Reasoning-graph-structured multi-hop QA dataset"}, {"area": "reason/planning", "role": "analysis", "why": "Evaluates LLMs across context structures and prompting"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["context structure", "prompting"], "benchmarks": [{"name": "GRS-QA", "purpose": "multi-hop QA with reasoning graphs", "is_new": true}], "results": [], "workflow": [], "summary": "GRS-QA encodes reasoning graphs for multi-hop QA and benchmarks LLMs across structures."}
{"id": 12000, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Visually-guided RLVR amplifying visual focus for VLM reasoning"}], "pattern": "vary-both", "fixes": ["benchmarks"], "varies": ["policy optimization"], "benchmarks": [], "results": [], "workflow": ["Visual attention compensation", "Advantage re-weighting"], "summary": "VGPO reinforces visual focus in RLVR to counter visual forgetting in VLMs."}
{"id": 13728, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "RL with language-conditioned hints for multilingual reasoning"}], "pattern": "vary-both", "fixes": ["multilingual math benchmarks"], "varies": ["RL method"], "benchmarks": [], "results": [], "workflow": ["Add hints", "Decay hints", "Language-adaptive switch"], "summary": "LANG uses decaying language-adaptive hints in RL to improve multilingual reasoning."}
{"id": 14639, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Recursive framework with visual layout language for deep research report generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow"], "benchmarks": [], "results": [], "workflow": ["Hierarchical recursive planning", "Abstract visual representation"], "summary": "CogGen recursively plans and restructures deep research reports with multimodal layout refinement."}
{"id": 8144, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Contrastive-loss guidance for diffusion sampling replacing negative prompting"}], "pattern": "n/a", "fixes": ["diffusion model"], "varies": ["guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "ContrastiveCFG guides diffusion sampling by contrasting positive and negative concepts."}
{"id": 15511, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Unsupervised step-level error detection and correction agent for multi-agent systems"}], "pattern": "vary-scaffold", "fixes": ["MAS"], "varies": ["error correction"], "benchmarks": [], "results": [], "workflow": ["Reconstruct next execution", "Anomaly scoring with prototypes", "Correction agent revises"], "summary": "MASC detects anomalous steps in multi-agent trajectories and triggers correction."}
{"id": 184, "in_agent_scope": false, "types": [{"area": "arch/attention", "role": "method", "why": "Two-step router with explicit guidance for MoE diffusion transformers"}], "pattern": "n/a", "fixes": ["DiT"], "varies": ["routing"], "benchmarks": [], "results": [], "workflow": ["Conditional routing", "Expert specialization"], "summary": "ProMoE improves MoE in diffusion transformers with explicit routing guidance."}
{"id": 8355, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Blue-teaming agent using automated red-teaming examples for code safety"}], "pattern": "fix-agent-vary-env", "fixes": ["threat model"], "varies": ["defense"], "benchmarks": [], "results": [], "workflow": ["Red team generates risks", "Constitution summarization", "Dynamic code analysis"], "summary": "BlueCodeAgent defends CodeGen AI using red-team-derived examples and code analysis."}
{"id": 15570, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Role-specific agents selecting interaction protocols for fake news detection"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["protocol"], "benchmarks": [{"name": "LIAR-RAW+", "purpose": "extended fake news dataset", "is_new": true}], "results": [], "workflow": ["Decompose verification", "Select protocol", "Debate"], "summary": "PARD uses role-specific multi-agent protocols for fake news detection."}
{"id": 13141, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Benchmark of adversarial instances for multi-agent LLM systems"}, {"area": "agent-safety", "role": "attack", "why": "300 instances across six attack types"}], "pattern": "fix-agent-vary-env", "fixes": ["multi-agent systems"], "varies": ["attacks"], "benchmarks": [{"name": "TAMAS", "purpose": "multi-agent adversarial robustness", "is_new": true}], "results": [], "workflow": [], "summary": "TAMAS benchmarks robustness of multi-agent LLM systems with 300 adversarial instances."}
{"id": 4687, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "Self-play on generated visual deduction games for VLM training"}], "pattern": "fix-env-vary-agent", "fixes": ["VLM"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["Generate games", "Self-play", "RL"], "summary": "Vision-Zero trains VLMs via self-play in 'Who Is the Spy' games without human annotation."}
{"id": 16910, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "analysis", "why": "MMGrader approach and evaluation of VLMs inferring mental model quality"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["models"], "benchmarks": [], "results": [], "workflow": ["Concept graphs"], "summary": "Evaluates nine VLMs on inferring student mental-model quality; best reach about 40% accuracy."}
{"id": 12500, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "attack", "why": "Two-stage jailbreak with hidden-state-guided local edits"}], "pattern": "fix-agent-vary-env", "fixes": ["target LLM"], "varies": ["adversarial prompt"], "benchmarks": [], "results": [], "workflow": ["Rephrase query", "Hidden-state-guided edits"], "summary": "AGILE jailbreaks LLMs via context rephrasing and activation-guided edits."}
{"id": 13522, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "defense", "why": "Training-free contrastive decoding to reduce over-refusal"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["decoding"], "benchmarks": [], "results": [], "workflow": ["Compare distributions", "Adaptive contrast"], "summary": "AdaCD mitigates LLM over-refusal while maintaining safety."}
{"id": 11996, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "method", "why": "Training-free LLM plus data validation for regulation-guided anomaly detection"}, {"area": "apps/domain", "role": "benchmark", "why": "New task and three datasets"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["LLM generates ideas", "Validate with data"], "summary": "RegValidator detects tabular anomalies from regulations; introduces task and datasets."}
{"id": 12708, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "WebDecept deceptive interface injection to evaluate web agents"}, {"area": "web-gui", "role": "analysis", "why": "Controlled evaluation shows susceptibility"}], "pattern": "fix-agent-vary-env", "fixes": ["web agents"], "varies": ["deceptive patterns"], "benchmarks": [{"name": "WebDecept", "purpose": "deceptive interface safety", "is_new": true}], "results": [], "workflow": ["Inject patterns", "Evaluate agents"], "summary": "WebDecept injects seven deceptive patterns into web environments; agents prove highly susceptible."}
{"id": 18643, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Model-based correctness judge for open-ended audio QA with 9,663 annotations"}], "pattern": "vary-evaluation", "fixes": ["LALMs"], "varies": ["judge"], "benchmarks": [{"name": "ORCA", "purpose": "audio QA correctness assessment", "is_new": true}], "results": [], "workflow": ["Annotate", "Curriculum training"], "summary": "ORCA is a judge model for open-ended audio QA answers, 0.91 Spearman with humans."}
{"id": 14301, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "eval-method", "why": "Snapshot-based diversity-guided user simulation for efficient agent evaluation"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["evaluation protocol"], "benchmarks": [], "results": [], "workflow": ["Capture state snapshots", "Branch trajectories", "Reuse prefixes"], "summary": "DIVERT branches from conversation snapshots to evaluate agents efficiently."}
{"id": 5178, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Large cross-platform computer-use dataset"}, {"area": "web-gui", "role": "train-agent", "why": "ScaleCUA models trained on data"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training data"], "benchmarks": [{"name": "WebArena-Lite-v2", "purpose": "GUI agent eval", "is_new": false}, {"name": "ScreenSpot-Pro", "purpose": "grounding", "is_new": false}, {"name": "MMBench-GUI L1-Hard", "purpose": "GUI eval", "is_new": false}, {"name": "OSWorld-G", "purpose": "grounding", "is_new": false}], "results": [{"benchmark": "MMBench-GUI L1-Hard", "system": "ScaleCUA", "base_model": null, "setting": "proposed", "metric": "score", "score": 94.4, "unit": "%"}, {"benchmark": "OSWorld-G", "system": "ScaleCUA", "base_model": null, "setting": "proposed", "metric": "score", "score": 60.6, "unit": "%"}, {"benchmark": "WebArena-Lite-v2", "system": "ScaleCUA", "base_model": null, "setting": "proposed", "metric": "score", "score": 47.4, "unit": "%"}], "workflow": ["Closed-loop data pipeline", "Train"], "summary": "ScaleCUA provides large cross-platform data and trained open-source computer use agents."}
{"id": 26, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Proactive self-refinement during generation trained via RL"}], "pattern": "n/a", "fixes": ["tasks"], "varies": ["refinement"], "benchmarks": [], "results": [], "workflow": ["Decide when to refine"], "summary": "PASR lets LLMs refine outputs during generation."}