Files
ai-taxonomy-literature-anal…/data/labels/o09.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
39 KiB
JSON

{"id": 12715, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Augments VLA with temporal memory hubs, latent reasoning tokens, parallel action decoding"}], "pattern": "fix-env-vary-agent", "fixes": ["LIBERO benchmark", "pretrained VLA backbone"], "varies": ["architecture", "memory hubs", "action decoding"], "benchmarks": [{"name": "LIBERO", "purpose": "evaluate VLA performance", "is_new": false}, {"name": "LeRobot", "purpose": "real-world evaluation", "is_new": false}], "results": [], "workflow": ["Temporal memory hubs", "Latent reasoning tokens", "Parallel action decoding"], "summary": "MIRTH augments a VLA with temporal memory, MI-optimized reasoning tokens and parallel decoding, achieving state-of-the-art on LIBERO."}
{"id": 11595, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Studies LLM cooperation failures in costless collaboration across eight LLMs"}, {"area": "multi-agent", "role": "environment", "why": "Builds turn-based environment making cooperation costless and optimal"}], "pattern": "vary-conditions", "fixes": ["environment", "instructions"], "varies": ["LLM models", "interventions", "communication"], "benchmarks": [], "results": [{"benchmark": "zero-cost collaboration environment", "system": "OpenAI o3", "base_model": null, "setting": "baseline", "metric": "percent of optimal collective performance", "score": 17, "unit": "%"}, {"benchmark": "zero-cost collaboration environment", "system": "o3-mini", "base_model": null, "setting": "baseline", "metric": "percent of optimal collective performance", "score": 50, "unit": "%"}], "workflow": ["Build costless cooperation environment", "Decompose cooperation vs competence failures", "Test protocols and incentives"], "summary": "Shows capable LLMs often fail zero-cost cooperation, separates cooperation from competence failures, and tests protocol and incentive interventions."}
{"id": 1284, "in_agent_scope": false, "types": [{"area": "train/pretrain", "role": "method", "why": "Runs multiple controlled pretraining experiments simultaneously in one training run"}, {"area": "train/pretrain", "role": "analysis", "why": "Investigates knowledge acquisition, math reasoning, watermarking, contamination, poisoning"}], "pattern": "vary-conditions", "fixes": ["single 1.5B training run"], "varies": ["embedded experiments"], "benchmarks": [], "results": [], "workflow": ["Embed ten experiments in one run", "Check interactions via continual pretraining"], "summary": "Proposes running many controlled pretraining experiments in a single training run, showing minimal interference at 1.5B parameters and 210B tokens."}
{"id": 15128, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Trains a retriever for agentic search using local relevance and global answer correctness"}], "pattern": "fix-env-vary-agent", "fixes": ["search agents", "QA benchmarks"], "varies": ["retriever training", "utility measure"], "benchmarks": [{"name": "seven single-hop and multi-hop QA benchmarks", "purpose": "evaluate retriever", "is_new": false}], "results": [], "workflow": ["Measure passage utility locally and globally", "Iteratively train agent and retriever"], "summary": "Agentic-R trains retrievers for agentic search using answer correctness and iterative agent-retriever co-optimization, beating baselines across agents."}
{"id": 11346, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Decouples context summarization into asynchronous parallel memory model"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "Reward-driven training of the memory model"}], "pattern": "vary-scaffold", "fixes": ["agent task"], "varies": ["context management pipeline"], "benchmarks": [{"name": "SWE-Bench-Verified", "purpose": "latency and performance", "is_new": false}], "results": [], "workflow": ["k-step-off async summarization", "Reward-trained memory model"], "summary": "CoMem decouples memory summarization from the agent to run in parallel, giving 1.4x latency improvement on SWE-Bench-Verified."}
{"id": 9152, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "MLLM guardrail evaluating observation-action pairs for physical hazards"}, {"area": "agent-safety", "role": "benchmark", "why": "Contributes EMBGuardTest of 329 scenarios and EMBHazard training data"}], "pattern": "vary-conditions", "fixes": ["threat model: physical hazards"], "varies": ["guardrail"], "benchmarks": [{"name": "EMBGuardTest", "purpose": "evaluate hazard detection", "is_new": true}, {"name": "EMBHazard", "purpose": "training data", "is_new": true}], "results": [], "workflow": ["Take observation and action", "Identify hazards", "Explain risk"], "summary": "EMBGuard is a small MLLM safety guardrail for embodied agents, with a new hazard dataset and benchmark, rivaling proprietary MLLMs."}
{"id": 9107, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "Benchmarks user agents coordinating with website content agents"}, {"area": "search-research", "role": "benchmark", "why": "Tasks include web search, recommendation, QA, deep research"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs under test"], "varies": ["tasks", "coordination strategies"], "benchmarks": [{"name": "AgentWebBench", "purpose": "multi-agent coordination in agentic web", "is_new": true}], "results": [], "workflow": [], "summary": "AgentWebBench evaluates decentralized user-agent and content-agent coordination across four web tasks, finding it lags centralized retrieval but gap shrinks with scale."}
{"id": 3196, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "MENTOR gives expert guidance only at critical decision points in RLVR"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL exploration guidance"], "benchmarks": [], "results": [], "workflow": ["Identify critical decision points", "Mix expert guidance token-level"], "summary": "MENTOR improves RLVR exploration by providing expert guidance only at critical decision points, retaining diversity."}
{"id": 16745, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Enriches arguments with lexical cues before fine-tuning LLMs for discourse relations"}, {"area": "apps/classic", "role": "analysis", "why": "Shows fine-tuned LLMs match encoder-based models on IDRC"}], "pattern": "vary-data", "fixes": ["fine-tuning recipe"], "varies": ["input cues"], "benchmarks": [], "results": [], "workflow": ["Add lexical semantic cues", "Fine-tune LLM"], "summary": "Fine-tuned LLMs rival encoders on implicit discourse relation classification, and lexical cue enrichment improves F1 by over 10 points."}
{"id": 13046, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "TravelBench for multi-turn tool-using travel planning with sandbox"}, {"area": "planning-decision", "role": "benchmark", "why": "Evaluates preference elicitation and recognizing unsolvable tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs under test"], "varies": ["travel tasks"], "benchmarks": [{"name": "TravelBench", "purpose": "real-world travel planning", "is_new": true}], "results": [], "workflow": [], "summary": "TravelBench is a real-world travel planning benchmark with single-turn, multi-turn and unsolvable subtasks and cached tool sandbox."}
{"id": 8327, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Context Preference Learning to reduce conversational inertia"}, {"area": "planning-decision", "role": "analysis", "why": "Identifies conversational inertia via attention analysis"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Inference-time context management strategies"}], "pattern": "vary-both", "fixes": ["environments"], "varies": ["training preference", "context length"], "benchmarks": [], "results": [], "workflow": ["Detect inertia", "Build preference pairs", "Preference-train model"], "summary": "Identifies conversational inertia in multi-turn agents and mitigates it with Context Preference Learning and context management."}
{"id": 12572, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "scaffold", "why": "Ontology-guided symbolic reasoning with retrieval for temporal reasoning in LLMs"}], "pattern": "vary-scaffold", "fixes": ["LLM backbones"], "varies": ["reasoning pipeline"], "benchmarks": [], "results": [], "workflow": ["Learn temporal ontology", "Symbolic deductive reasoning", "Logic-constrained retrieval"], "summary": "ODL-TempLLM models temporal structure with ontologies and description logic reasoning, outperforming prior methods on three datasets."}
{"id": 16581, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "ETOM five-level benchmark for MCP tool orchestration"}, {"area": "tool-use", "role": "eval-method", "why": "Equal function sets give objective F1 ground truth"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["tool orchestration levels"], "benchmarks": [{"name": "ETOM", "purpose": "MCP multi-hop tool orchestration", "is_new": true}], "results": [], "workflow": [], "summary": "ETOM benchmarks multi-hop tool orchestration in hierarchical MCP ecosystems with objective metrics, exposing robustness weaknesses."}
{"id": 17315, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent LLM tutoring system with centralized learner state"}, {"area": "multi-agent", "role": "application", "why": "Orchestrator coordinates six specialized tutoring agents"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["tutoring pipeline"], "benchmarks": [], "results": [], "workflow": ["Orchestrate six agents", "Update shared learner state", "Give graduated hints"], "summary": "IntelliCode is a multi-agent tutoring system with a shared versioned learner state, validated with simulated learners."}
{"id": 5742, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "data", "why": "Collects LLM coordination trajectories via task-level tool interfaces"}, {"area": "multi-agent", "role": "train-agent", "why": "Distills coordination policies into lightweight agents"}], "pattern": "fix-env-vary-agent", "fixes": ["cooperative environments"], "varies": ["policy training via distillation"], "benchmarks": [{"name": "Overcooked-AI", "purpose": "cooperation evaluation", "is_new": false}, {"name": "Level-Based Foraging", "purpose": "cooperation evaluation", "is_new": false}], "results": [], "workflow": ["LLM coordinates via tool interfaces", "Collect trajectories", "Distill compact policies"], "summary": "SynCoord distills LLM task-level coordination into compact policies, matching RL baselines and generalizing to unseen teams."}
{"id": 12140, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Evolving agent using implicit user feedback with KTO alignment"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Inference-time prompt evolution from feedback matrix"}], "pattern": "vary-both", "fixes": ["benchmarks"], "varies": ["training with feedback", "prompts"], "benchmarks": [{"name": "MT-Bench-101", "purpose": "evaluate multi-turn", "is_new": false}, {"name": "WildBench", "purpose": "evaluate", "is_new": false}, {"name": "FB-Bench", "purpose": "evaluate", "is_new": false}, {"name": "LMSYS-Chat-1M", "purpose": "feedback data source", "is_new": false}, {"name": "WildChat", "purpose": "feedback data source", "is_new": false}], "results": [], "workflow": ["Annotate feedback patterns", "Estimate feedback matrix", "KTO plus prompt evolution"], "summary": "IEvoAgent mines implicit user feedback to evolve conversational agents via offline KTO and inference-time prompt evolution."}
{"id": 5720, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "TopBench for implicit predictive reasoning in table QA"}, {"area": "science-data", "role": "analysis", "why": "Analyzes intent recognition failures in text and agentic workflows"}], "pattern": "fix-agent-vary-env", "fixes": ["models under test"], "varies": ["tasks"], "benchmarks": [{"name": "TopBench", "purpose": "implicit predictive tabular QA", "is_new": true}], "results": [], "workflow": [], "summary": "TopBench has 779 samples testing implicit predictive reasoning over tables; models often default to lookups."}
{"id": 15165, "in_agent_scope": false, "types": [{"area": "interp/icl", "role": "scaffold", "why": "Online training-free adaptation with confidence-guided exemplar cache for multi-label classification"}], "pattern": "vary-scaffold", "fixes": ["frozen LLM"], "varies": ["exemplar caching", "confidence measure"], "benchmarks": [], "results": [], "workflow": ["Compute L3R confidence", "Submodular exemplar cache", "Adapt in-context"], "summary": "SCOTTA adapts LLM multi-label classification online via label-level confidence and a submodular in-context exemplar cache."}
{"id": 577, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "LLM-guided evolutionary program synthesis for QMC point sets and Sobol parameters"}], "pattern": "vary-scaffold", "fixes": ["QMC objectives"], "varies": ["evolved programs"], "benchmarks": [], "results": [], "workflow": ["Mutate code with LLM", "Select by fitness", "Numerical refinement"], "summary": "LLM-guided evolutionary program synthesis discovers improved low-discrepancy point sets and Sobol direction numbers."}
{"id": 15275, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Sociological expert agents reconstruct WVS responses into individual value profiles"}, {"area": "social-sim", "role": "eval-method", "why": "Multi-agent debate to evaluate value consistency"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["profile construction"], "benchmarks": [{"name": "World Values Survey", "purpose": "individual value data", "is_new": false}], "results": [{"benchmark": "World Values Survey", "system": "ExpertIVS", "base_model": null, "setting": "proposed", "metric": "value restoration fidelity", "score": 90.78, "unit": "%"}], "workflow": ["Expert agents interpret WVS", "Build profiles", "Debate evaluation"], "summary": "ExpertIVS uses sociological expert agents to simulate individual values, reaching 90.78% value restoration fidelity."}
{"id": 12759, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "PEAP dataset for proactive embodied planning from vision and audio"}, {"area": "embodied-nav", "role": "analysis", "why": "Ablation shows joint multimodal understanding helps"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "PEAP", "purpose": "proactive embodied planning", "is_new": true}], "results": [], "workflow": [], "summary": "PEAP is a multimodal dataset for proactive embodied action planning; LLMs show deficiencies in perception."}
{"id": 16586, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "benchmark", "why": "BiasInEar speech benchmark from Global MMLU Lite"}, {"area": "safety/fairness", "role": "analysis", "why": "Evaluates nine audio LLMs under language, accent, gender, option order"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["language", "accent", "gender", "option order"], "benchmarks": [{"name": "BiasInEar", "purpose": "speech bias in audio LLMs", "is_new": true}, {"name": "Global MMLU Lite", "purpose": "source questions", "is_new": false}], "results": [], "workflow": [], "summary": "BiasInEar benchmarks speech bias in audio LLMs, finding sensitivity to language and option order but robustness to gender."}
{"id": 15901, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "defense", "why": "Inference-time cognitive analysis graph guides safe multi-turn responses"}], "pattern": "vary-scaffold", "fixes": ["threat model"], "varies": ["safety mechanism"], "benchmarks": [], "results": [], "workflow": ["Retrieve question graph", "Retain safety nodes from history", "Guide response"], "summary": "CogGSE uses retrieved cognitive analysis graphs at inference to improve multi-turn LLM safety."}
{"id": 11182, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Agentic framework synthesizing and refining epidemiological simulators"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["epidemic modeling pipeline"], "benchmarks": [], "results": [], "workflow": ["Synthesize flow graph", "Verify graph", "Compile and calibrate model"], "summary": "EpiAgent iteratively synthesizes, verifies and calibrates epidemiological simulators via a flow graph representation."}
{"id": 6238, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "LiveNewsBench with fresh news questions for agentic web search"}, {"area": "search-research", "role": "data", "why": "Pipeline supports large-scale training data generation"}], "pattern": "fix-agent-vary-env", "fixes": ["systems under test"], "varies": ["questions"], "benchmarks": [{"name": "LiveNewsBench", "purpose": "agentic web search with fresh news", "is_new": true}], "results": [], "workflow": ["Generate QA from news", "Human-verify subset"], "summary": "LiveNewsBench is a regularly updated benchmark of fresh news questions for evaluating agentic web search."}
{"id": 9028, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "data", "why": "PDB-wide mutation augmentation dataset"}, {"area": "apps/science", "role": "method", "why": "Unsupervised multi-source mutation preference distillation"}], "pattern": "vary-data", "fixes": ["ProteinGym eval"], "varies": ["training signals"], "benchmarks": [{"name": "ProteinGym", "purpose": "zero-shot mutation effect evaluation", "is_new": false}], "results": [], "workflow": ["Enumerate mutations", "Align model signals", "Distill preferences"], "summary": "MutAtlas builds a PDB-wide mutation dataset and distills multi-source preferences without experimental labels."}
{"id": 7089, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Diffusion-based generation of query-adaptive communication topologies"}, {"area": "multi-agent", "role": "scaffold", "why": "Reduces communication overhead in LLM multi-agent systems"}], "pattern": "vary-scaffold", "fixes": ["LLM agents"], "varies": ["communication topology"], "benchmarks": [{"name": "six benchmarks", "purpose": "evaluate accuracy and tokens", "is_new": false}], "results": [], "workflow": ["Condition on query", "Discrete graph diffusion", "Redundancy-aware generation"], "summary": "RADAR generates redundancy-aware communication graphs with diffusion, improving accuracy and lowering token use."}
{"id": 12164, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "scaffold", "why": "Test-time corrective vector at uncertain tokens for self-reflection"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["decoding procedure"], "benchmarks": [{"name": "mathematical reasoning benchmarks", "purpose": "evaluate reasoning", "is_new": false}], "results": [], "workflow": ["Detect high-entropy tokens", "Train corrective vector", "Correct distribution"], "summary": "SRGen reflects at uncertain tokens during decoding with a learned corrective vector, improving math reasoning."}
{"id": 3663, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "KAIROS simulates quiz contests with peers of varying reliability"}, {"area": "multi-agent", "role": "analysis", "why": "Studies rapport and social influence by model size"}, {"area": "multi-agent", "role": "defense", "why": "Evaluates prompting, SFT and GRPO mitigation"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["peer reliability", "mitigation"], "benchmarks": [{"name": "KAIROS", "purpose": "rapport bias in multi-agent interaction", "is_new": true}], "results": [], "workflow": [], "summary": "KAIROS measures rapport bias in LLM multi-agent settings and tests prompting, SFT and GRPO mitigation."}
{"id": 13272, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Tool merging and context-aware retrieval to select tools"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["toolset management"], "benchmarks": [{"name": "three tool-use benchmarks", "purpose": "tool selection accuracy", "is_new": false}], "results": [], "workflow": ["Merge redundant tools", "Auto-correct merges", "Retrieve relevant tools"], "summary": "ToolScope merges redundant tools and filters by relevance, improving tool selection accuracy by 8.38% to 38.6%."}
{"id": 17214, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Schema-free entity structure extraction via enrichment, refinement, unification"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["extraction method"], "benchmarks": [], "results": [], "workflow": ["Enrich", "Refine", "Unify"], "summary": "ZOES extracts entity attribute-value structures zero-shot without schemas or annotations, across three domains."}
{"id": 459, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Disentangled hierarchical VAE latent diffusion for 3D human-human interaction generation"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["generative model"], "benchmarks": [], "results": [], "workflow": ["Disentangle global and individual latents", "Contrastive contact constraints", "DDIM diffusion"], "summary": "DHVAE disentangles interaction context and individual motion for plausible 3D human-human interaction generation."}
{"id": 15281, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "P-BRIDGE distills reflective strategies from student-teacher trajectories into small agents"}, {"area": "memory-selfevolve", "role": "data", "why": "Contrastive failure-recovery trajectories for distillation"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["distillation data and training"], "benchmarks": [{"name": "2WikiMultiHopQA", "purpose": "small model accuracy", "is_new": false}], "results": [{"benchmark": "2WikiMultiHopQA", "system": "P-BRIDGE", "base_model": "0.6B model", "setting": "finetuned", "metric": "accuracy", "score": 34.2, "unit": "%"}, {"benchmark": "2WikiMultiHopQA", "system": "0.6B baseline", "base_model": "0.6B model", "setting": "baseline", "metric": "accuracy", "score": 6.2, "unit": "%"}], "workflow": ["Reflect during execution", "Contrast student-teacher trajectories", "Abstract strategies"], "summary": "P-BRIDGE distills failure-recovery strategies into small agents, raising a 0.6B model on 2WikiMultiHopQA from 6.2% to 34.2%."}
{"id": 3197, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "UIS-QA benchmark for unindexed information seeking"}, {"area": "search-research", "role": "train-agent", "why": "UIS-Digger multi-agent framework trained with SFT and RL"}], "pattern": "vary-both", "fixes": ["search agents"], "varies": ["tasks", "agent framework"], "benchmarks": [{"name": "UIS-QA", "purpose": "unindexed information seeking", "is_new": true}, {"name": "GAIA", "purpose": "compare agent performance", "is_new": false}, {"name": "BrowseComp-zh", "purpose": "compare agent performance", "is_new": false}], "results": [{"benchmark": "UIS-QA", "system": "state-of-the-art agents", "base_model": null, "setting": "baseline", "metric": "score", "score": 24.55, "unit": "other"}, {"benchmark": "GAIA", "system": "state-of-the-art agents", "base_model": null, "setting": "baseline", "metric": "score", "score": 70.9, "unit": "other"}, {"benchmark": "BrowseComp-zh", "system": "state-of-the-art agents", "base_model": null, "setting": "baseline", "metric": "score", "score": 46.7, "unit": "other"}], "workflow": ["Dual-mode browsing", "Search and file parsing", "SFT and RL"], "summary": "Introduces UIS-QA and UIS-Digger for information not indexed by search engines."}
{"id": 14870, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Compact hyperbolic Lorentzian model for EHR question answering"}], "pattern": "fix-env-vary-agent", "fixes": ["EHR-QA benchmarks"], "varies": ["model geometry"], "benchmarks": [{"name": "MIMIC-IV EHR-QA benchmarks", "purpose": "evaluate QA", "is_new": false}], "results": [], "workflow": ["Embed in hyperbolic space", "Cross-attention pointer heads", "Pretrain on next-visit diagnosis"], "summary": "HypEHR is a compact hyperbolic model that nears LLM-based EHR QA with far fewer parameters."}
{"id": 14685, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "CORBA contagious recursive blocking attacks cause denial-of-collaboration in LLM multi-agent systems"}], "pattern": "fix-agent-vary-env", "fixes": ["multi-agent system"], "varies": ["adversarial instructions", "topologies"], "benchmarks": [], "results": [], "workflow": ["Inject benign recursive instruction", "Spread across agents", "Cause message loops"], "summary": "Formalizes Denial-of-Collaboration and introduces CORBA attacks that paralyze LLM multi-agent systems with benign recursive instructions."}
{"id": 7231, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "ORLoopBench with ORDebug repair instances and ORBias for OR decisions"}, {"area": "science-data", "role": "train-agent", "why": "Solver-verified RLVR training of 8B model"}], "pattern": "vary-both", "fixes": ["solver feedback"], "varies": ["tasks", "model training"], "benchmarks": [{"name": "ORLoopBench", "purpose": "OR self-correction", "is_new": true}, {"name": "ORDebug", "purpose": "LP/MILP repair", "is_new": true}, {"name": "ORBias", "purpose": "decision rationality", "is_new": true}], "results": [{"benchmark": "ORDebug LP repair", "system": "8B RLVR model", "base_model": "8B", "setting": "finetuned", "metric": "RR@5", "score": 95.3, "unit": "%"}, {"benchmark": "ORDebug LP repair", "system": "frontier APIs", "base_model": null, "setting": "baseline", "metric": "RR@5", "score": 92.4, "unit": "%"}], "workflow": ["Solver re-execution with IIS", "RLVR on repair"], "summary": "ORLoopBench benchmarks solver-in-the-loop infeasible model repair; RLVR lets an 8B model beat frontier APIs on LP repair."}
{"id": 7794, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "PSBench of 600 Photoshop image editing tasks for GUI agents"}], "pattern": "fix-agent-vary-env", "fixes": ["GUI agents"], "varies": ["Photoshop tasks"], "benchmarks": [{"name": "PSBench", "purpose": "Photoshop GUI agents", "is_new": true}], "results": [{"benchmark": "PSBench hard", "system": "Agent S3", "base_model": null, "setting": "baseline", "metric": "success rate", "score": 18.09, "unit": "%"}], "workflow": [], "summary": "PSBench is the first Photoshop GUI agent benchmark with 600 tasks; best agent gets 18.09% on hard tasks."}
{"id": 8602, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Training-free draft-guided contiguous verification for faster dLLM decoding"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["decoding strategy"], "benchmarks": [], "results": [], "workflow": ["Contiguous verification", "Draft-guided leaping"], "summary": "DC-Leap accelerates diffusion LLM parallel decoding without training via contiguous verification and draft guidance."}
{"id": 6821, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Inference-time residual-guided multi-resolution refinement for frozen time series foundation models"}], "pattern": "vary-scaffold", "fixes": ["frozen backbones"], "varies": ["inference wrapper"], "benchmarks": [{"name": "SPEI drought forecasting", "purpose": "evaluate MSE", "is_new": false}], "results": [], "workflow": ["Coarse-to-fine refinement", "Residual diagnosis"], "summary": "RGMR refines frozen time series foundation models for drought forecasting, cutting TimesFM SPEI MSE by up to 18.9%."}
{"id": 5449, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "data", "why": "Clean-PR corpus of 2M pull requests converted to search/replace edits"}, {"area": "code-swe", "role": "train-agent", "why": "Mid-training plus agentless SFT with error-driven augmentation"}], "pattern": "vary-data", "fixes": ["SWE-bench"], "varies": ["training data"], "benchmarks": [{"name": "SWE-bench Lite", "purpose": "evaluate", "is_new": false}, {"name": "SWE-bench Verified", "purpose": "evaluate", "is_new": false}], "results": [], "workflow": ["Convert PR diffs", "Mid-train", "Agentless SFT"], "summary": "Clean-PR trains repo-level code editing from pull requests, improving SWE-bench Lite by 13.6% and Verified by 12.3%."}
{"id": 7738, "in_agent_scope": true, "types": [{"area": "science-data", "role": "train-agent", "why": "Dual-loop agentic captioning optimized by proxy reward without caption labels"}], "pattern": "fix-env-vary-agent", "fixes": ["no labels"], "varies": ["captioning agent training"], "benchmarks": [], "results": [], "workflow": ["Agent explores captions", "Downstream reasoning reward", "Feedback optimizes agent"], "summary": "Caption label-free learning optimizes a time series captioning agent from downstream-reasoning proxy rewards."}
{"id": 1685, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "analysis", "why": "Shows VLMs attend to evidence yet answer incorrectly"}, {"area": "interp/steering", "role": "scaffold", "why": "Training-free attention-masking intervention highlighting evidence regions"}], "pattern": "vary-conditions", "fixes": ["VLM weights"], "varies": ["attention intervention"], "benchmarks": [], "results": [], "workflow": ["Analyze layer attention", "Mask to highlight evidence"], "summary": "VLMs often see but do not use visual evidence; attention-based masking improves accuracy without training."}
{"id": 10441, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Semiparametric estimator combining labels with pairwise comparison signals for math evaluation"}, {"area": "reason/math", "role": "theory", "why": "Efficiency bound and variance reduction guarantees"}], "pattern": "vary-evaluation", "fixes": ["models"], "varies": ["estimator"], "benchmarks": [], "results": [], "workflow": ["Obtain comparison signals", "Use as control variates", "One-step estimator"], "summary": "Efficient evaluation of math reasoning combining labeled outcomes with LLM pairwise comparison signals."}
{"id": 15387, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "data", "why": "OmniCoT interleaved guided-token reasoning traces"}, {"area": "arch/multimodal", "role": "train-agent", "why": "SFT then DyME-GRPO to build EmoOmni"}], "pattern": "vary-data", "fixes": ["Qwen2.5-Omni-7B"], "varies": ["data and RL reward"], "benchmarks": [], "results": [], "workflow": ["SFT on OmniCoT", "DyME-GRPO"], "summary": "EmoOmni addresses modality collapse in emotion reasoning with guided-token data and entropy-regulated GRPO."}
{"id": 14842, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Adversarial Environmental Injection poisons tool outputs; Potemkin harness"}, {"area": "agent-safety", "role": "analysis", "why": "Evaluates frontier models across 11,000+ runs"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tool output poisoning"], "benchmarks": [{"name": "Potemkin", "purpose": "MCP robustness testing harness", "is_new": true}], "results": [], "workflow": ["Poison retrieval", "Structural traps"], "summary": "Potemkin tests tool-integrated agents against lying tools via breadth and depth attacks."}
{"id": 14107, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Uses linguistic attestation to constrain AES key search"}], "pattern": "n/a", "fixes": [], "varies": ["cryptanalysis method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Proposes using linguistic plausibility to prune keys in AES cryptanalysis."}
{"id": 17092, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "eval-method", "why": "Evaluates faithfulness and plausibility of ICD coding rationales"}, {"area": "apps/science", "role": "data", "why": "New multi-granular rationale-annotated ICD dataset from MIMIC-IV"}], "pattern": "vary-evaluation", "fixes": ["ICD coding models"], "varies": ["rationale types", "metrics"], "benchmarks": [{"name": "MIMIC-IV", "purpose": "rationale annotated dataset", "is_new": true}], "results": [], "workflow": [], "summary": "Evaluates ICD coding rationales for faithfulness and plausibility, with a new annotated dataset and LLM-guided learning."}
{"id": 5748, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "wGraph-based workflow management and KV cache state management for agent serving"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow generation", "KV cache"], "benchmarks": [], "results": [], "workflow": ["Build wGraph", "Generate workflow adaptively", "Manage KV state"], "summary": "GraphFlow represents workflows as graph for adaptive generation and efficient KV cache in agent serving."}
{"id": 5157, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Meta-learned soft prompts distilled from visual features for few-shot VQA"}], "pattern": "fix-env-vary-agent", "fixes": ["VL-ICL Bench"], "varies": ["prompt learning"], "benchmarks": [{"name": "VL-ICL Bench", "purpose": "few-shot VQA", "is_new": false}], "results": [], "workflow": ["Distill soft prompts", "Attention-mapper", "Test-time gradient adaptation"], "summary": "Meta-adaptive prompt distillation yields few-shot VQA adaptation, outperforming ICL by 21.2%."}
{"id": 9190, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "TRIP-Bench long-horizon travel planning with 18 tools"}, {"area": "planning-decision", "role": "train-agent", "why": "GTPO online multi-turn RL"}], "pattern": "vary-both", "fixes": ["agents"], "varies": ["tasks", "RL algorithm"], "benchmarks": [{"name": "TRIP-Bench", "purpose": "long-horizon interactive travel", "is_new": true}], "results": [{"benchmark": "TRIP-Bench easy", "system": "advanced models", "base_model": null, "setting": "baseline", "metric": "success rate", "score": 50, "unit": "%"}], "workflow": ["Reward normalization", "Reward differencing"], "summary": "TRIP-Bench tests long-horizon travel agents with up to 150+ tool calls, plus GTPO multi-turn RL."}
{"id": 17581, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "analysis", "why": "Pilot study of LLM pseudonym generation and evaluation for English and Swedish"}, {"area": "safety/privacy", "role": "eval-method", "why": "Explores metrics for pseudonym evaluation"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["methods", "constraints"], "benchmarks": [], "results": [], "workflow": [], "summary": "Pilot experiments generating pseudonyms for redacted PII spans, examining evaluation metrics and errors."}
{"id": 7075, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "PPT-Eval of 120 PowerPoint tasks"}, {"area": "web-gui", "role": "eval-method", "why": "Rubric-based partial credit evaluation framework"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "PPT-Eval", "purpose": "PowerPoint computer-use", "is_new": true}], "results": [], "workflow": [], "summary": "PPT-Eval benchmarks computer-use agents on 120 PowerPoint tasks with rubric-based partial-credit evaluation."}
{"id": 9441, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Aligns decoder manifolds for quantized and continuous embeddings in neural speech codecs"}], "pattern": "fix-env-vary-agent", "fixes": ["XCodec2"], "varies": ["training loss"], "benchmarks": [{"name": "XCodec2", "purpose": "base codec", "is_new": false}], "results": [], "workflow": ["Feature-mapping loss"], "summary": "Self-guidance improves neural codec fidelity and permits 4x smaller codebook."}
{"id": 10916, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "analysis", "why": "CUDAnalyst attributes planning decisions to feedback components in CUDA kernel generation"}, {"area": "memory-selfevolve", "role": "analysis", "why": "Studies feedback-conditioned planning in self-evolving agents"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["feedback components"], "benchmarks": [], "results": [], "workflow": ["Freeze trajectories", "Inject selected feedback"], "summary": "CUDAnalyst enables controlled attribution of planning decisions to feedback in CUDA kernel agents."}
{"id": 3568, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Hierarchical automaton multi-agent visual reasoning with trainable hyper agent"}, {"area": "multi-agent", "role": "train-agent", "why": "SFT of transition policy on MATA-SFT-90K"}], "pattern": "vary-both", "fixes": ["agents"], "varies": ["automaton", "policy"], "benchmarks": [{"name": "MATA-SFT-90K", "purpose": "SFT data", "is_new": true}], "results": [], "workflow": ["Hyper agent picks state", "Rule-based sub-automata", "Shared memory"], "summary": "MATA is a multi-agent hierarchical automaton for visual reasoning with a fine-tuned transition policy."}
{"id": 14550, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "One-shot topology generation with diverse interaction modes"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["communication topology"], "benchmarks": [], "results": [], "workflow": ["Generate topology once", "Heterogeneous interaction modes"], "summary": "TopoDIM generates communication topologies in one shot, cutting tokens 46.41% with 1.50% better performance."}
{"id": 9730, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Interactive agentic framework with exploration policies to extract LLM knowledge"}, {"area": "memory-selfevolve", "role": "analysis", "why": "Finds knowledge scaling law and Pass@1 vs Pass@k tradeoff"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["exploration policy"], "benchmarks": [], "results": [], "workflow": ["Explore policies", "Dedup and adjudicate", "Audit relevance"], "summary": "Interactive agentic framework probes LLM knowledge boundaries; Recursive Taxonomy works best."}
{"id": 1174, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "TimeSeriesExam and agent-generated time series reasoning benchmarks"}, {"area": "science-data", "role": "environment", "why": "LLM agents generate benchmarks at scale"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["benchmarks"], "benchmarks": [{"name": "TimeSeriesExam", "purpose": "time series reasoning", "is_new": true}], "results": [], "workflow": ["Templates plus agents", "Generate from real datasets"], "summary": "Scalable methods using templates and LLM agents to create time series reasoning benchmarks."}
{"id": 270, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "environment", "why": "LH-Deception simulation framework with performer, supervisor, auditor"}, {"area": "agent-safety", "role": "analysis", "why": "Quantifies deception across 11 frontier models"}], "pattern": "vary-conditions", "fixes": ["framework"], "varies": ["models", "pressure"], "benchmarks": [{"name": "LH-Deception", "purpose": "long-horizon deception", "is_new": true}], "results": [], "workflow": ["Performer completes tasks", "Supervisor gives feedback", "Auditor reviews"], "summary": "LH-Deception simulates long-horizon LLM deception and finds it model-dependent, rising with pressure and eroding trust."}