d6fc7c1dfa
- 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
61 lines
32 KiB
JSON
61 lines
32 KiB
JSON
{"id": 15240, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Intent-aware context folding keeping full dialogue history for user-centric tool agents"}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["context folding method"], "benchmarks": [], "results": [], "workflow": ["keep full history", "fold context per turn", "track evolving user intent"], "summary": "U-Fold dynamically folds context with intent tracking for multi-turn user-centric tool-using agents."}
|
|
{"id": 12409, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Adaptive preference margins for reward models on subjective role-play tasks"}], "pattern": "vary-data", "fixes": ["Bradley-Terry framework"], "varies": ["margin calibration"], "benchmarks": [], "results": [], "workflow": ["estimate model knowledge", "calibrate preference margin", "train reward model"], "summary": "AAM dynamically calibrates reward model preference margins for subjective ambiguity without extra annotation."}
|
|
{"id": 10866, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Local sanitization minimizing UI state leaked to remote agents"}, {"area": "web-gui", "role": "application", "why": "Privacy-aware minimal UI view for GUI agents"}], "pattern": "vary-scaffold", "fixes": ["threat model", "agent"], "varies": ["observation disclosure policy"], "benchmarks": [], "results": [], "workflow": ["score sensitivity and necessity", "ternary disclosure policy", "send minimal view"], "summary": "MINIM is a trusted local broker that minimizes sensitive UI state sent to remote GUI agents."}
|
|
{"id": 16823, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "method", "why": "AST-guided translation of natural language into first-order logic"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["translation method"], "benchmarks": [], "results": [], "workflow": ["parse to AST", "guide LLM translation"], "summary": "NL2Logic translates natural language to first-order logic using AST guidance with LLMs."}
|
|
{"id": 15802, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "method", "why": "Prompt optimization with historical contrastive reflection to debias black-box MLLMs"}], "pattern": "vary-scaffold", "fixes": ["MLLM weights"], "varies": ["debiasing prompts"], "benchmarks": [], "results": [], "workflow": ["optimize debias prompt", "contrastive self-reflection on history"], "summary": "HRPO automatically optimizes debiasing prompts for black-box multimodal LLMs using reflection over optimization history."}
|
|
{"id": 6837, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "ICRL4AHT benchmark of in-context RL for ad-hoc teamwork on Overcooked-V2"}], "pattern": "fix-agent-vary-env", "fixes": ["ICRL algorithms"], "varies": ["teammates", "train-test shift"], "benchmarks": [{"name": "ICRL4AHT", "purpose": "evaluate in-context RL in ad-hoc teamwork", "is_new": true}, {"name": "Overcooked-V2", "purpose": "base environment", "is_new": false}], "results": [], "workflow": [], "summary": "A JAX Overcooked-V2 benchmark revealing limitations of AD and DPT for ad-hoc teamwork."}
|
|
{"id": 6978, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Evolutionary counterfactual planning with memory for multi-agent embodied collaboration"}, {"area": "embodied-nav", "role": "scaffold", "why": "Rule library induced from failures"}], "pattern": "vary-scaffold", "fixes": ["LLM planner"], "varies": ["planning framework"], "benchmarks": [], "results": [], "workflow": ["induce symbolic constraints", "mutate counterfactual plans", "memory-grounded evaluation"], "summary": "EvoCF improves LLM multi-agent embodied planning via constraint induction and evolutionary counterfactual plans."}
|
|
{"id": 5515, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Stealthy cooperative memory-injection poisoning of LLM agents"}], "pattern": "vary-conditions", "fixes": ["agent"], "varies": ["injected query sets"], "benchmarks": [], "results": [], "workflow": ["generate query set", "bidirectional evolutionary optimization", "inject into memory"], "summary": "MemIncept poisons agent memory using cooperative benign-looking queries optimized evolutionarily in black-box settings."}
|
|
{"id": 3514, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Residual RL specialists and distillation to self-improve VLAs"}, {"area": "embodied-nav", "role": "data", "why": "Distribution-aware hybrid rollout data collection"}], "pattern": "fix-env-vary-agent", "fixes": ["VLA backbone", "environment"], "varies": ["training data", "residual actors"], "benchmarks": [], "results": [], "workflow": ["train residual RL specialists", "hybrid rollout collection", "fine-tune VLA"], "summary": "PLD improves VLA models by residual RL specialists and distribution-aware data for fine-tuning."}
|
|
{"id": 1652, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Structured graph-search exploration to collect web agent trajectories"}, {"area": "web-gui", "role": "train-agent", "why": "Fine-tune 7B model on collected data"}], "pattern": "vary-data", "fixes": ["WebArena env", "training recipe"], "varies": ["training trajectories"], "benchmarks": [{"name": "WebArena", "purpose": "collect data and evaluate", "is_new": false}], "results": [{"benchmark": "WebArena", "system": "Go-Browse 7B", "base_model": null, "setting": "finetuned", "metric": "success rate", "score": 21.7, "unit": "%"}], "workflow": ["graph search exploration", "collect trajectories", "fine-tune 7B model"], "summary": "Go-Browse collects diverse web trajectories by structured exploration; a fine-tuned 7B model reaches 21.7% on WebArena."}
|
|
{"id": 2575, "in_agent_scope": false, "types": [{"area": "knowledge/editing", "role": "method", "why": "Difficulty metric TangleScore and adaptive purge-imprint editing for unstructured knowledge"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["editing method"], "benchmarks": [], "results": [], "workflow": ["score edit difficulty", "adaptive purge and imprint"], "summary": "TangleScore quantifies edit difficulty and drives PIPE for unstructured knowledge editing."}
|
|
{"id": 9593, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Benchmark of 200 tasks for agentic upward deception under constraints"}, {"area": "agent-safety", "role": "analysis", "why": "Evaluates 11 LLMs showing deceptive behaviors"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs under test"], "varies": ["constrained task scenarios"], "benchmarks": [{"name": "Upward deception benchmark", "purpose": "200 tasks, 5 types, 8 scenarios", "is_new": true}], "results": [], "workflow": [], "summary": "Defines agentic upward deception and benchmarks 11 LLMs in constrained environments, finding action-based deception."}
|
|
{"id": 15691, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Black-box auditing to detect Trojan agents in multi-agent systems"}, {"area": "multi-agent", "role": "attack", "why": "Evolved conversational strategies trigger hidden behaviors"}], "pattern": "vary-conditions", "fixes": ["black-box agents"], "varies": ["probing conversation strategies"], "benchmarks": [], "results": [], "workflow": ["evolve conversational probes", "query suspect agent", "detect Trojan behavior"], "summary": "CTUS uses evolutionary conversation strategies to audit and detect Trojan LLMs in multi-agent systems."}
|
|
{"id": 14967, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Study of collective intelligence factor across 108 LLM agent groups"}], "pattern": "vary-conditions", "fixes": ["tasks"], "varies": ["group size", "LLM composition", "topology"], "benchmarks": [], "results": [], "workflow": [], "summary": "Constructs 108 LLM agent groups and tests whether a general collective intelligence factor exists."}
|
|
{"id": 15721, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "ParaCook benchmark for time-efficient parallel collaborative planning"}, {"area": "planning-decision", "role": "analysis", "why": "LLMs show suboptimal parallel planning"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated LLMs"], "varies": ["cooking tasks"], "benchmarks": [{"name": "ParaCook", "purpose": "time-efficient multi-agent planning", "is_new": true}, {"name": "Overcooked", "purpose": "inspiration", "is_new": false}], "results": [], "workflow": [], "summary": "Overcooked-inspired benchmark showing LLMs produce suboptimal plans for parallel, asynchronous collaboration."}
|
|
{"id": 17827, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "InQA+ and GenIQA corpora for indirect question answering in English, German, Bavarian"}], "pattern": "fix-agent-vary-env", "fixes": ["multilingual encoders"], "varies": ["datasets", "languages"], "benchmarks": [{"name": "InQA+", "purpose": "hand-annotated IQA evaluation set", "is_new": true}, {"name": "GenIQA", "purpose": "synthetic GPT-4o-mini training set", "is_new": true}], "results": [], "workflow": [], "summary": "Two multilingual indirect QA corpora; low performance even for English with mBERT, XLM-R, mDeBERTa."}
|
|
{"id": 15930, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Dual-layer working and experience memory for KG reasoning agent"}, {"area": "search-research", "role": "scaffold", "why": "Multi-hop KG question answering agent"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory mechanism"], "benchmarks": [], "results": [], "workflow": ["compress triplets in working memory", "abstract experience strategies", "reuse across tasks"], "summary": "EvoMemKG gives KG reasoning agents evolvable working and experience memory for multi-hop QA."}
|
|
{"id": 13176, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "ReviewBench rubric-based evaluation of review text"}, {"area": "domain-agents", "role": "scaffold", "why": "Rubric-guided tool-integrated multi-agent reviewer"}], "pattern": "vary-scaffold", "fixes": ["Phi-4-14B drafter"], "varies": ["review generation framework"], "benchmarks": [{"name": "ReviewBench", "purpose": "rubric-based review evaluation", "is_new": true}], "results": [], "workflow": ["draft review", "ground with evidence", "consolidate"], "summary": "ReviewGrounder, a rubric-guided tool-integrated multi-agent reviewer, evaluated on new ReviewBench."}
|
|
{"id": 4033, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Natural-language PDDL lifted regression planning for open-world embodied agents"}], "pattern": "vary-scaffold", "fixes": ["LLM/VLM"], "varies": ["planning representation"], "benchmarks": [], "results": [], "workflow": ["express goals in NL-PDDL", "lifted regression", "plan sensing actions"], "summary": "NL-PDDL combines LLMs with lifted regression planning for open-world goal-oriented embodied planning."}
|
|
{"id": 10767, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Training-free nuclear instance segmentation via prototype-guided prompting"}], "pattern": "n/a", "fixes": ["foundation model"], "varies": ["prompting method"], "benchmarks": [], "results": [], "workflow": ["build reference prototypes", "optimal transport alignment", "prompt segmenter"], "summary": "SPROUT is a training- and annotation-free prompting framework for nuclear instance segmentation."}
|
|
{"id": 8705, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "method", "why": "Self-Debias trajectory-level training to interrupt bias propagation in CoT"}], "pattern": "vary-data", "fixes": ["LLM"], "varies": ["training objective"], "benchmarks": [], "results": [], "workflow": ["reallocate probability mass", "trajectory-level objective"], "summary": "Self-Debias instills self-correction to stop social bias propagation in chain-of-thought."}
|
|
{"id": 14079, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Hierarchical multi-agent system generating educational games"}, {"area": "multi-agent", "role": "scaffold", "why": "Phase-based LangGraph subgraphs with quality gates"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["architecture"], "benchmarks": [], "results": [{"benchmark": "200 questions", "system": "GameDAI", "base_model": null, "setting": "proposed", "metric": "validation pass rate", "score": 90, "unit": "%"}], "workflow": ["phase-based subgraphs", "quality gates", "schema validation"], "summary": "GameDAI multi-agent framework converts instructor questions into validated educational games."}
|
|
{"id": 12489, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "Multi-agent persona debate to simulate multi-dimensional human evaluation"}, {"area": "eval/judge", "role": "eval-method", "why": "LLM-as-judge framework"}], "pattern": "vary-evaluation", "fixes": ["evaluated systems"], "varies": ["judge protocol"], "benchmarks": [], "results": [], "workflow": ["build personas from documents", "instantiate agents", "group debate", "aggregate"], "summary": "MAJ-EVAL builds evaluator personas from documents and debates to align with human multi-dimensional evaluation."}
|
|
{"id": 5578, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "COBA pipeline auditing validity of agent benchmarks"}], "pattern": "vary-evaluation", "fixes": ["agent benchmarks"], "varies": ["auditing method"], "benchmarks": [], "results": [], "workflow": ["decompose into four components", "rule and LLM detectors", "filter flaws"], "summary": "COBA automatically diagnoses and filters validity issues in agent benchmarks across six benchmarks."}
|
|
{"id": 16251, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "RevCI annotated benchmark for review contradiction with graded intensity"}, {"area": "apps/science", "role": "method", "why": "IMPACT structured method for contradiction analysis"}], "pattern": "n/a", "fixes": ["models"], "varies": ["task formulation"], "benchmarks": [{"name": "RevCI", "purpose": "review contradiction evidence and intensity", "is_new": true}], "results": [], "workflow": [], "summary": "Fine-grained peer review contradiction analysis with RevCI benchmark and IMPACT method."}
|
|
{"id": 12336, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Toulmin-guided curriculum goal-conditioned training for clinical diagnostic reasoning"}], "pattern": "vary-data", "fixes": ["LLM"], "varies": ["training pipeline"], "benchmarks": [], "results": [], "workflow": ["adapt Toulmin model", "curriculum training", "goal-conditioned learning"], "summary": "Trains LLMs for trustworthy clinical argumentation using a Toulmin-guided curriculum goal-conditioned pipeline."}
|
|
{"id": 8215, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Dual-graph outline and knowledge memory for open-ended deep research"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Co-evolving knowledge graph memory"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["maintain outline graph", "maintain knowledge graph", "targeted exploration"], "summary": "DualGraph separates outline structure from knowledge exploration for open-ended deep research agents."}
|
|
{"id": 6394, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "benchmark", "why": "Plausibility-based retrosynthesis benchmark with ChemCensor metric"}, {"area": "apps/science", "role": "eval-method", "why": "ChemCensor chemical plausibility metric"}, {"area": "apps/science", "role": "data", "why": "CREED dataset for LLM training"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["metric", "dataset"], "benchmarks": [{"name": "CREED", "purpose": "train retrosynthesis model", "is_new": true}], "results": [], "workflow": [], "summary": "New retrosynthesis benchmarking with ChemCensor plausibility metric and CREED dataset."}
|
|
{"id": 746, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "CounselBench expert evaluation and adversarial benchmark for mental health QA"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["questions", "ratings"], "benchmarks": [{"name": "CounselBench", "purpose": "mental health QA expert evaluation", "is_new": true}, {"name": "CounselChat", "purpose": "patient question source", "is_new": false}], "results": [], "workflow": [], "summary": "Expert-annotated benchmark with 2,000 evaluations stress-testing LLMs on mental health questions."}
|
|
{"id": 11485, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Multi-agent framework for full-stack web development"}, {"area": "code-swe", "role": "data", "why": "Repository back-translation for self-improvement data"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": ["plan and edit code", "dev-oriented testing", "back-translate repositories"], "summary": "FullStack-Agent builds full-stack web apps with multi-agent dev framework and repository back-translation learning."}
|
|
{"id": 13588, "in_agent_scope": false, "types": [{"area": "train/data", "role": "method", "why": "Iterative selection and refinement of instruction tuning data"}], "pattern": "vary-data", "fixes": ["instruction tuning recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": ["quality differentiation", "evaluate-refine-review", "output alignment"], "summary": "Iterative framework to select and refine instruction data rather than pruning low-quality samples."}
|
|
{"id": 12304, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Dual-objective RL for multimodal agentic search"}, {"area": "tool-use", "role": "train-agent", "why": "Reduces unnecessary search calls"}], "pattern": "fix-env-vary-agent", "fixes": ["VQA tasks"], "varies": ["RL reward objective"], "benchmarks": [], "results": [], "workflow": ["reward search quality", "penalize unnecessary calls", "RL post-train MLLM"], "summary": "DORA uses dual-objective RL to improve MLLM search quality and efficiency in VQA."}
|
|
{"id": 712, "in_agent_scope": false, "types": [{"area": "train/data", "role": "method", "why": "Critic-Adviser-Reviser cyclic refinement for synthetic EMR generation"}], "pattern": "vary-data", "fixes": ["LLM"], "varies": ["generation pipeline"], "benchmarks": [], "results": [], "workflow": ["critic evaluates", "adviser feedback", "reviser refines"], "summary": "LLM-CARe iteratively refines synthetic EMRs at corpus, section and document levels."}
|
|
{"id": 15149, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Tree-guided CoT with RL for aspect sentiment quad prediction"}], "pattern": "vary-data", "fixes": ["ASQP task"], "varies": ["reasoning and RL"], "benchmarks": [], "results": [], "workflow": ["tree reasoning templates", "generate chains", "RL refine"], "summary": "Tree-CoT-RT guides LLM explainable reasoning and RL for aspect sentiment quad prediction."}
|
|
{"id": 15212, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Meta-tool synthesis and reflective replanning for agents"}, {"area": "tool-use", "role": "benchmark", "why": "ToolHop-Pro diagnostic benchmark with pruned tools"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent architecture"], "benchmarks": [{"name": "ToolHop-Pro", "purpose": "diagnose missing-tool handling", "is_new": true}], "results": [], "workflow": ["synthesize tools", "reflective replanning"], "summary": "CAR dynamically creates tools and replans globally; releases ToolHop-Pro benchmark."}
|
|
{"id": 11851, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Hierarchical main agent, meta-thinker and context manager for long-horizon tasks"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Evolving context management"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["framework"], "benchmarks": [], "results": [], "workflow": ["main agent acts", "meta-thinker oversees", "context manager condenses"], "summary": "COMPASS manages evolving context with a hierarchical agent framework for long-horizon reasoning."}
|
|
{"id": 13362, "in_agent_scope": false, "types": [{"area": "arch/longctx", "role": "method", "why": "Attention-guided context compression for personalized LLMs"}], "pattern": "vary-scaffold", "fixes": ["task"], "varies": ["compression method"], "benchmarks": [], "results": [], "workflow": ["use attention signals", "compress profile"], "summary": "Attn-GS uses attention patterns to compress user histories for efficient personalization."}
|
|
{"id": 4327, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Swap-guided preference learning to fix posterior collapse in VPL"}], "pattern": "vary-data", "fixes": ["VPL framework"], "varies": ["learning objective"], "benchmarks": [], "results": [], "workflow": ["construct swap annotators", "guide encoder"], "summary": "SPL prevents posterior collapse in variational preference learning for personalized RLHF."}
|
|
{"id": 15070, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "Self-evolving multi-agent safety evaluation from policy documents"}, {"area": "agent-safety", "role": "benchmark", "why": "Generated evolving safety benchmark"}], "pattern": "vary-evaluation", "fixes": ["evaluated LLMs"], "varies": ["evaluation protocol"], "benchmarks": [], "results": [], "workflow": ["ingest policy docs", "generate test cases", "evolve from results"], "summary": "AgenticEval generates and evolves safety benchmarks from policy documents via multi-agent loop."}
|
|
{"id": 7913, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Disentangled alignment and structure-guided reasoning for table images"}], "pattern": "vary-data", "fixes": ["LVLM"], "varies": ["alignment and reasoning"], "benchmarks": [], "results": [], "workflow": ["disentangled alignment", "global-to-local reasoning"], "summary": "DiSCo and Table-GLS adapt LVLMs for table reasoning with minimal annotation."}
|
|
{"id": 16834, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Turn-level MDP PPO advantage estimation for multi-turn agent RL"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["advantage estimation"], "benchmarks": [{"name": "WebShop", "purpose": "multi-turn RL evaluation", "is_new": false}, {"name": "Sokoban", "purpose": "multi-turn RL evaluation", "is_new": false}], "results": [], "workflow": ["turn-level MDP", "PPO advantage"], "summary": "Turn-PPO uses turn-level PPO for stable multi-turn agent RL on WebShop and Sokoban."}
|
|
{"id": 14158, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Optimize conditioning context in latent space to steer coding agents"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["context search"], "benchmarks": [{"name": "SWE-Bench Verified", "purpose": "raise weak 24B model to frontier parity", "is_new": false}], "results": [], "workflow": ["optimize context", "execution feedback", "steer agent"], "summary": "Plug-in layer optimizing conditioning contexts raises a 24B model's SWE-Bench Verified resolution to frontier parity."}
|
|
{"id": 16406, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "application", "why": "Schema-guided multi-frame dialogue state for motivational interviewing responses"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["dialogue method"], "benchmarks": [], "results": [], "workflow": ["update dialogue states", "decide strategy", "generate response"], "summary": "Schema-guided dialogue method generates MI-aligned counselor responses."}
|
|
{"id": 8420, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "attack", "why": "Diffusion-guided coherent adversarial suffix attacks on LLMs"}], "pattern": "vary-conditions", "fixes": ["target models"], "varies": ["attack method"], "benchmarks": [], "results": [], "workflow": ["diffusion guidance", "greedy coordinate search"], "summary": "GCD generates low-perplexity semantically coherent adversarial attacks against safety-aligned LLMs."}
|
|
{"id": 4136, "in_agent_scope": false, "types": [{"area": "social-sim", "role": "train-agent", "why": "Hint-guided RL with verifiable role-awareness reward for role-play agents"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["reward"], "benchmarks": [{"name": "RAIDEN", "purpose": "role-awareness", "is_new": false}, {"name": "CharacterEval", "purpose": "role-play eval", "is_new": false}], "results": [], "workflow": ["extract hints", "verifiable reward", "RL"], "summary": "VeriRole improves role-awareness of role-play agents using hints and verifiable reward."}
|
|
{"id": 12649, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Citation-aware rubric rewards for RL of deep search agents"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": ["decompose rubrics", "verify citations", "citation-aware GRPO"], "summary": "CaRR provides fine-grained citation-aware rewards for RL training of deep search agents."}
|
|
{"id": 8203, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Multi-round agentic RAG for medical reasoning"}, {"area": "search-research", "role": "scaffold", "why": "Conflict-driven query retrieval"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["test-time framework"], "benchmarks": [], "results": [], "workflow": ["identify conflict", "retrieve evidence", "refine history"], "summary": "MA-RAG iteratively turns response conflicts into queries for test-time scaling in medical reasoning."}
|
|
{"id": 13159, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Divide-and-conquer subtable selection for table QA"}], "pattern": "vary-scaffold", "fixes": ["LM"], "varies": ["selection method"], "benchmarks": [], "results": [], "workflow": ["window-based selection", "aggregate evidence"], "summary": "PieTa selects subtables via iterative multi-resolution window selection for table QA."}
|
|
{"id": 1572, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "Interactive Video Dataset for real-time face-to-face questions"}], "pattern": "fix-agent-vary-env", "fixes": ["VLMs"], "varies": ["dataset"], "benchmarks": [{"name": "Interactive Video Dataset (IVD)", "purpose": "assess real-time VLM conversation", "is_new": true}], "results": [], "workflow": [], "summary": "IVD benchmark assesses VLMs on live question answering and studies fine-tuning."}
|
|
{"id": 15278, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Stackelberg multi-agent contract revision"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent structure"], "benchmarks": [], "results": [], "workflow": ["leader sets risk budgets", "revision agent edits", "verifier checks"], "summary": "RCBSF uses leader-follower agents for risk-constrained contract revision with convergence proof."}
|
|
{"id": 7032, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "method", "why": "Graph-guided LLM malicious code detection"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": ["parse code graph", "train GNN", "guide LLM attention"], "summary": "GNN-guided LLM attention localizes malicious behavior across files."}
|
|
{"id": 14610, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "LLM multi-perspective location completion for next location prediction"}], "pattern": "vary-data", "fixes": ["predictor"], "varies": ["augmentation"], "benchmarks": [], "results": [], "workflow": ["translate coordinates to text", "dual completion agents"], "summary": "LaMDA uses LLM agents to complete check-in data for next location prediction."}
|
|
{"id": 16360, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Joint reflective optimization of prompts and tool schemas"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["prompts", "tool descriptions"], "benchmarks": [], "results": [], "workflow": ["rollout reflection", "co-optimize instructions and schemas"], "summary": "JTPRO jointly optimizes instructions and tool descriptions for reliable tool calling."}
|
|
{"id": 8888, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Exploration-aware RL for adaptive agent exploration"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["RL reward and grouping"], "benchmarks": [], "results": [], "workflow": ["variational reward", "exploration-aware grouping"], "summary": "Exploration-aware RL lets agents explore only when uncertain."}
|
|
{"id": 16805, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "RADAR chart attribution benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["MLLMs"], "varies": ["dataset"], "benchmarks": [{"name": "RADAR", "purpose": "chart region attribution", "is_new": true}], "results": [], "workflow": [], "summary": "RADAR benchmark and approach for attributing MLLM chart reasoning to regions."}
|
|
{"id": 1604, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Metacognitive policy for deferring to human experts"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["policy"], "benchmarks": [], "results": [], "workflow": ["metacognitive policy", "dual-loop optimization"], "summary": "LIMA trains multi-agent LLMs to decide when to defer to humans via DLPO."}
|
|
{"id": 16193, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Step-level guardrail for tool invocation safety"}, {"area": "agent-safety", "role": "benchmark", "why": "TS-Bench detection benchmark"}], "pattern": "vary-conditions", "fixes": ["agent"], "varies": ["guardrail"], "benchmarks": [{"name": "TS-Bench", "purpose": "step-level tool safety detection", "is_new": true}], "results": [], "workflow": ["detect unsafe action", "feedback to agent"], "summary": "ToolSafe provides TS-Bench, RL-trained TS-Guard, and TS-Flow guardrail feedback."}
|
|
{"id": 371, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "eval-method", "why": "Human-preference arena for computer-use agents"}, {"area": "web-gui", "role": "analysis", "why": "Analysis of CUA behavior"}], "pattern": "vary-evaluation", "fixes": ["CUAs"], "varies": ["evaluation protocol"], "benchmarks": [], "results": [], "workflow": [], "summary": "Open-source platform for head-to-head CUA evaluation using human preferences."}
|
|
{"id": 5427, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "General coding agent with Lean MCP for formal math"}, {"area": "tool-use", "role": "scaffold", "why": "MCP tools"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["agent system"], "benchmarks": [{"name": "Putnam2025", "purpose": "formal theorem proving", "is_new": false}], "results": [], "workflow": ["Claude Code", "Lean MCP tools"], "summary": "Numina-Lean-Agent uses Claude Code with Lean MCP for formal theorem proving."}
|
|
{"id": 10266, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "VerdictBench e-commerce dispute benchmark"}, {"area": "domain-agents", "role": "scaffold", "why": "CyberJurors multi-agent verdict framework"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["framework"], "benchmarks": [{"name": "VerdictBench", "purpose": "6,000 multimodal dispute cases", "is_new": true}], "results": [], "workflow": ["individual chain-of-thought", "jury deliberation"], "summary": "VerdictBench benchmark and CyberJurors multi-agent framework for e-commerce disputes."}
|