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
26 KiB
JSON
61 lines
26 KiB
JSON
{"id": 15531, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "attack", "why": "Experience-driven multi-agent framework for black-box jailbreak prompt generation"}], "pattern": "vary-conditions", "fixes": ["target LLMs"], "varies": ["jailbreak generation method"], "benchmarks": [], "results": [], "workflow": ["agents propose prompts", "reuse experience", "iterate attacks"], "summary": "Multi-agent experience-driven black-box jailbreak method with better transfer and query efficiency."}
|
|
{"id": 14509, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Slide-editing agent using structured object model manipulation instead of pixels"}], "pattern": "vary-scaffold", "fixes": ["MLLM"], "varies": ["interaction modality"], "benchmarks": [], "results": [], "workflow": ["parse slide object model", "LLM edits structure"], "summary": "Slide editing agent manipulating structured object model via language instead of GUI screenshots, cutting cost and latency."}
|
|
{"id": 16945, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Controlled generation with linguistic attribute encoder and Pareto masking"}], "pattern": "n/a", "fixes": ["LM"], "varies": ["control method"], "benchmarks": [], "results": [], "workflow": ["encode attributes", "inject at BOS", "Pareto masking"], "summary": "Controlled text generation over many linguistic attributes using BOS injection and P-masking training."}
|
|
{"id": 18114, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "benchmark", "why": "Multimodal KGQA benchmark on household daily activities"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "HOME-KGQA", "purpose": "multimodal KGQA", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark for multimodal KGQA on household activities with spatiotemporal data."}
|
|
{"id": 13705, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "analysis", "why": "Studies multi-party dialogue generation and judge bias vs humans"}, {"area": "eval/judge", "role": "eval-method", "why": "Shows LLM judges diverge from human preferences in social dialogue"}], "pattern": "vary-evaluation", "fixes": ["generator"], "varies": ["judges"], "benchmarks": [], "results": [], "workflow": ["generate dialogues", "human eval", "compare judges"], "summary": "Lightweight multi-party dialogue generator used to show LLM judges are biased relative to human evaluation."}
|
|
{"id": 12051, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "train-agent", "why": "RL training of MLLM for video-guided translation with dynamic modality switching"}], "pattern": "fix-env-vary-agent", "fixes": ["VMT benchmarks"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["detect ambiguity", "RL train", "switch modality"], "summary": "RL framework for video-guided MT that switches between text-only and multimodal reasoning."}
|
|
{"id": 5956, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Training-free image-to-video method preserving early-step motion phase"}], "pattern": "n/a", "fixes": ["I2V model"], "varies": ["denoising procedure"], "benchmarks": [], "results": [], "workflow": ["spectral analysis", "lock motion phase"], "summary": "Finds 2-step generation is more physical than 50-step; proposes training-free PhaseLock."}
|
|
{"id": 18377, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "survey", "why": "Position framework for sociolinguistic variation in NLP with Luxembourgish case study"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Position paper arguing NLP should embrace sociolinguistic variation, with Luxembourgish case study."}
|
|
{"id": 4881, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "train-agent", "why": "Two-stage finetuning with RL for efficient multi-hop QA retrieval"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training recipe"], "benchmarks": [], "results": [], "workflow": ["finetune", "RL on search budget"], "summary": "Two-stage finetuning (SFT then RL) for small LMs balancing accuracy and retrieval efficiency in multi-hop QA."}
|
|
{"id": 13131, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Experience-driven automatic creation and adaptation of domain agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent design"], "benchmarks": [], "results": [], "workflow": ["analyze trajectories", "revise agent"], "summary": "Framework that automatically creates and adapts domain agents using experience on why agents succeed or fail."}
|
|
{"id": 1561, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "SFT plus dual-environment RL for web agents"}, {"area": "search-research", "role": "data", "why": "SailorFog-QA-2 knowledge-graph synthetic data"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["data", "training"], "benchmarks": [], "results": [], "workflow": ["synthesize QA", "SFT", "dual-env RL"], "summary": "Post-training pipeline for web agents with graph-based synthetic data and simulator-plus-real dual-environment RL."}
|
|
{"id": 1803, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Cooperative LLM agents for network-wide traffic signal control"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent cooperation design"], "benchmarks": [], "results": [], "workflow": ["spatiotemporal reasoning", "cooperate"], "summary": "Cooperative LLM agent framework for network-wide traffic signal control."}
|
|
{"id": 12130, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Hybrid validation safety detector for mobile GUI agents"}, {"area": "agent-safety", "role": "benchmark", "why": "Introduces mobile agent safety benchmark"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["safety detector"], "benchmarks": [], "results": [], "workflow": ["rule check", "VLM validate trajectory"], "summary": "Mobile agent safety benchmark plus hybrid formal and model-based unsafe-action detector."}
|
|
{"id": 8638, "in_agent_scope": false, "types": [{"area": "interp/steering", "role": "analysis", "why": "Studies model scale and compression effects on brain encoding"}], "pattern": "vary-conditions", "fixes": ["methods"], "varies": ["model size", "precision"], "benchmarks": [], "results": [], "workflow": [], "summary": "Systematic study of how model scale and precision affect brain-encoding alignment."}
|
|
{"id": 15741, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "LLM-based preference estimation via opponent modeling in negotiation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["estimation method"], "benchmarks": [], "results": [], "workflow": ["parse utterances", "estimate preferences"], "summary": "Preference estimation integrating natural language utterance understanding for opponent modeling in multi-party negotiation."}
|
|
{"id": 11687, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Calibrated dynamic job-shop scheduling benchmark for LLM agents"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["instances"], "benchmarks": [{"name": "DynaSchedBench", "purpose": "dynamic scheduling", "is_new": true}], "results": [], "workflow": [], "summary": "Calibrated DFJSP benchmark with stress index; reports observability paradox for LLM scheduling agents."}
|
|
{"id": 3549, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Lane-level navigation by associating SD and online perception maps"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "New task refining SD-map routes into lane-level guidance using online perception maps."}
|
|
{"id": 670, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Self-play multi-turn RL on zero-sum games improving reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["games"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["self-play", "online RL"], "summary": "Self-play RL on zero-sum games yields transferable reasoning without human supervision."}
|
|
{"id": 12814, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "analysis", "why": "Studies visual ambiguity in VQA via answer-supporting regions"}, {"area": "eval/bench", "role": "benchmark", "why": "Operationalizes ambiguity for eliciting and evaluating"}], "pattern": "vary-conditions", "fixes": ["VLLMs"], "varies": ["ambiguity"], "benchmarks": [], "results": [], "workflow": [], "summary": "Region-centric operationalization of visual question ambiguity and elicitation in VLLMs."}
|
|
{"id": 7742, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "train-agent", "why": "RL for LLMs generating executable KBQA logical forms"}], "pattern": "fix-env-vary-agent", "fixes": ["KBQA benchmarks"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["interact with KB", "RL"], "summary": "RL framework for KBQA moving from template reasoning to environment-verified logical form generation."}
|
|
{"id": 8841, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Mutual-information-guided RL for VLM perception-then-reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": ["describe image", "MI reward", "RL"], "summary": "RLVR for VLMs with mutual-information rewards separating perception from reasoning errors."}
|
|
{"id": 15556, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "scaffold", "why": "Training-free entropy-guided inference control for NL2FOL"}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["inference"], "benchmarks": [], "results": [], "workflow": ["perturb input", "entropy guide"], "summary": "Fine-tuning-free inference control improving logical consistency in NL2FOL."}
|
|
{"id": 10135, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "analysis", "why": "Studies how information access affects LLM monitors detecting sabotage"}, {"area": "agent-safety", "role": "defense", "why": "Monitoring agents for sabotage"}], "pattern": "vary-conditions", "fixes": ["agent environments"], "varies": ["monitor information access"], "benchmarks": [], "results": [], "workflow": [], "summary": "Study showing LLM monitors often perform better with less information (less-is-more) across sabotage environments."}
|
|
{"id": 16296, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Entity-aware retrieval with reasoning navigator for multi-hop QA"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Structured entity-aware retrieval with chain-of-reasoning navigator for multi-hop QA."}
|
|
{"id": 12982, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "application", "why": "Multi-agent LLM framework for distributed training strategy optimization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "LLM multi-agent framework automatically optimizing distributed training strategies."}
|
|
{"id": 9980, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Triangular fixed-point refinement coupling observable, latent and memory for persona consistency"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["refinement"], "benchmarks": [], "results": [], "workflow": [], "summary": "Method coupling response, latent state and memory to keep long-horizon dialogue persona consistent, with proof."}
|
|
{"id": 13891, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "User study of LLM simulated users versus humans on tau-Bench"}, {"area": "web-gui", "role": "eval-method", "why": "Questions validity of simulated-user agent evaluation"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["user type"], "benchmarks": [{"name": "tau-Bench", "purpose": "retail tasks with simulated vs human users", "is_new": false}], "results": [], "workflow": [], "summary": "User study showing LLM-simulated users are unreliable proxies for humans in agent evaluation."}
|
|
{"id": 16801, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "Romanian driving exam multimodal QA dataset"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["tasks"], "benchmarks": [{"name": "RoD-TAL", "purpose": "Romanian driving law QA", "is_new": true}], "results": [], "workflow": [], "summary": "Multimodal Romanian driving-law QA dataset evaluating LLMs and VLMs."}
|
|
{"id": 13459, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "scaffold", "why": "Iterative refinement of annotation guidelines for LLM NER annotation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["guidelines"], "benchmarks": [], "results": [], "workflow": ["add guidelines", "moderate", "refine"], "summary": "Reusing and iteratively refining annotation guidelines to align LLM annotators on biomedical NER."}
|
|
{"id": 17130, "in_agent_scope": false, "types": [{"area": "arch/tokenization", "role": "method", "why": "Morphologically guided tokenization for Latin encoders"}], "pattern": "n/a", "fixes": ["pretraining"], "varies": ["tokenizer"], "benchmarks": [], "results": [], "workflow": [], "summary": "Morphology-aware tokenization for Latin encoder models."}
|
|
{"id": 15477, "in_agent_scope": false, "types": [{"area": "multiling/multi", "role": "benchmark", "why": "Massively multilingual table QA benchmark and instruction data"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "M3TQA-BENCH", "purpose": "multilingual table QA", "is_new": true}, {"name": "M3TQA-INSTRUCT", "purpose": "training data", "is_new": true}], "results": [], "workflow": [], "summary": "Multilingual table QA benchmark across 97 languages with instruction set."}
|
|
{"id": 6112, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Trains clarifier with information gain reward"}], "pattern": "fix-env-vary-agent", "fixes": ["tools env"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": ["ask question", "info gain reward", "train"], "summary": "Information-gain reward for training LLM agents to ask clarifying questions under underspecified instructions."}
|
|
{"id": 1765, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Distributed self-routing among LLMs"}], "pattern": "n/a", "fixes": ["LLM pool"], "varies": ["routing"], "benchmarks": [], "results": [], "workflow": [], "summary": "Distributed self-routing where each LLM decides whether to answer, replacing centralized routers."}
|
|
{"id": 13593, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Generative engine optimization with latent user demand"}], "pattern": "n/a", "fixes": ["GSE"], "varies": ["content optimization"], "benchmarks": [], "results": [], "workflow": [], "summary": "GEO method optimizing content using latent user demands to improve visibility in generative search answers."}
|
|
{"id": 16901, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Self-correcting noise-augmented few-shot bio/chem NER"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Few-shot generative NER with self-correction and noise augmentation for complex entities."}
|
|
{"id": 4361, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Scalable end-to-end online RL for computer-use agents"}, {"area": "web-gui", "role": "environment", "why": "Scalable desktop RL infrastructure"}], "pattern": "fix-env-vary-agent", "fixes": ["desktop tasks"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["API-GUI actions", "scale RL"], "summary": "Online RL framework with API-GUI paradigm for desktop agents and scalable training infrastructure."}
|
|
{"id": 18659, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "method", "why": "Linguistic steganography using self-adjusting asymmetric numeral system"}], "pattern": "n/a", "fixes": ["LM"], "varies": ["coding"], "benchmarks": [], "results": [], "workflow": [], "summary": "Adaptive linguistic steganography with user-specified embedding rates."}
|
|
{"id": 10416, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "data", "why": "Parlay-Instruct corpus via Critic-Refine"}, {"area": "code-swe", "role": "train-agent", "why": "Specialized models for parallel code with agentic evolution"}], "pattern": "vary-data", "fixes": ["LLMs"], "varies": ["data", "agent"], "benchmarks": [], "results": [], "workflow": ["build corpus", "finetune", "evolve code"], "summary": "Corpus and fine-tuned models plus evolutionary agent for high-performance parallel code on irregular data."}
|
|
{"id": 15811, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "Self-evolving LMM with trajectory-aware intrinsic reasoning rewards"}], "pattern": "fix-env-vary-agent", "fixes": ["unlabeled images"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": ["propose", "solve CoT", "reward agreement"], "summary": "Proposer-Solver self-evolution for LMMs adding reasoning-agreement rewards beyond outcomes."}
|
|
{"id": 792, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Repository memory for code localization agents"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Long-term repository memory"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory"], "benchmarks": [], "results": [], "workflow": [], "summary": "Augments localization agents with long-term repository memory of modules and bug-fix associations."}
|
|
{"id": 15119, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Continuous pre-training for tool knowledge"}], "pattern": "fix-env-vary-agent", "fixes": ["tool benchmarks"], "varies": ["training stage"], "benchmarks": [], "results": [], "workflow": [], "summary": "Continuous pre-training to build tool knowledge in LLM agents for OOD tool use."}
|
|
{"id": 12070, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "train-agent", "why": "Answer-aligned retriever fine-tuning with KG-augmented curriculum"}], "pattern": "fix-env-vary-agent", "fixes": ["generator"], "varies": ["retriever training"], "benchmarks": [], "results": [], "workflow": [], "summary": "Retriever fine-tuning aligned to answer sufficiency with KG-based curriculum for long-context RAG."}
|
|
{"id": 9973, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "analysis", "why": "Inductive biases for transformers to learn Newtonian world models"}], "pattern": "vary-conditions", "fixes": ["task"], "varies": ["inductive biases"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows spatial smoothness and regression formulation help transformers learn veridical world models."}
|
|
{"id": 18485, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Tangible child-AI storytelling dialogue system"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["conversation design"], "benchmarks": [], "results": [], "workflow": [], "summary": "Tangible dialogue system with board, NFC toys and app for child-AI co-creative storytelling."}
|
|
{"id": 18181, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "English-Swedish translationese preference probing dataset"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["sentences"], "benchmarks": [{"name": "Translationese Swedish dataset", "purpose": "probing preferences", "is_new": true}], "results": [], "workflow": [], "summary": "First English-Swedish dataset contrasting translationese with idiomatic sentences to probe LLM preferences."}
|
|
{"id": 16886, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "LLM-GNN integration for open-world KGQA"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Combines LLMs and GNNs to answer questions over incomplete knowledge graphs."}
|
|
{"id": 13295, "in_agent_scope": false, "types": [{"area": "arch/tokenization", "role": "method", "why": "Learned corpus-dependent subcharacter encoding via HMM"}], "pattern": "n/a", "fixes": ["subword tokenizer"], "varies": ["encoding"], "benchmarks": [], "results": [], "workflow": [], "summary": "HMM-learned fixed-length character codes as alternative to byte encoding."}
|
|
{"id": 14925, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Holistic security benchmark for LVLM web agents"}], "pattern": "fix-agent-vary-env", "fixes": ["web agents"], "varies": ["attacks"], "benchmarks": [{"name": "SecureWebArena", "purpose": "web agent security", "is_new": true}], "results": [], "workflow": [], "summary": "Security benchmark with six web environments and 2,970 adversarial trajectories over six attack vectors."}
|
|
{"id": 456, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "data", "why": "VisCode-Multi-679K visualization coding dataset"}, {"area": "code-swe", "role": "benchmark", "why": "Multi-language visualization benchmark"}, {"area": "code-swe", "role": "train-agent", "why": "VisCoder2 models"}], "pattern": "vary-data", "fixes": ["LLMs"], "varies": ["data", "models"], "benchmarks": [{"name": "VisCode-Multi-679K", "purpose": "training dataset", "is_new": true}], "results": [], "workflow": ["build data", "train", "evaluate"], "summary": "Dataset, benchmark and models for multi-language visualization coding agents with iterative correction."}
|
|
{"id": 7696, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Self-reflection policy optimization for long-horizon reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["reflect", "patch rollouts", "distill"], "summary": "RL framework converting reflections on own trajectories into on-policy distillation targets."}
|
|
{"id": 4699, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Hierarchical preference learning for long-horizon agents"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["preference objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "Hierarchical DPO combining trajectory and step-level preference signals for long-horizon agents."}
|
|
{"id": 17266, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Explicit rollback mechanism for web agent search"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search procedure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Gives web agents ability to revert to earlier states for efficient search and recovery."}
|
|
{"id": 14503, "in_agent_scope": true, "types": [{"area": "search-research", "role": "data", "why": "Hierarchical Experience for training search agents"}, {"area": "search-research", "role": "train-agent", "why": "RL search agent with experience"}], "pattern": "vary-data", "fixes": ["search env"], "varies": ["experience/data"], "benchmarks": [], "results": [], "workflow": [], "summary": "HiExp extracts empirical knowledge to stabilize RL training of search agents."}
|
|
{"id": 1897, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "LLM preference-guided Bayesian optimization"}], "pattern": "vary-scaffold", "fixes": ["BO"], "varies": ["LLM guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "LLM preference-guided Bayesian optimization improving cold-start for scientific discovery."}
|
|
{"id": 14364, "in_agent_scope": true, "types": [{"area": "search-research", "role": "analysis", "why": "Studies verbalized confidence in search agents"}, {"area": "search-research", "role": "scaffold", "why": "Confidence-guided test-time scaling"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["inference procedure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Examines agent verbalized confidence and uses it for test-time scaling of web agents."}
|
|
{"id": 16930, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Test-driven sub-agent workflow for repository patching"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow"], "benchmarks": [], "results": [], "workflow": ["propose patch", "debug", "revise"], "summary": "Agentic workflow decomposing patch proposal, debugging, revision to resolve human-written tests."}
|
|
{"id": 1095, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Empirical study of reinforcement finetuning for generative agents"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["RFT settings"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies reinforcement finetuning applied to Transformer-based generative agents."}
|
|
{"id": 5150, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "eval-method", "why": "Proactive reward system for GUI agents via reasoner-evaluator probing"}], "pattern": "vary-evaluation", "fixes": ["GUI agents"], "varies": ["reward method"], "benchmarks": [], "results": [], "workflow": ["schedule probes", "evaluate state", "reward"], "summary": "Reasoner schedules state-probing tasks for evaluator agents to give GUI agent rewards."}
|
|
{"id": 2678, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "defense", "why": "Test-time defense for CLIP using adversarial directional bias"}], "pattern": "vary-conditions", "fixes": ["threat model"], "varies": ["defense"], "benchmarks": [], "results": [], "workflow": [], "summary": "Test-time defense exploiting dominant feature shift of adversarial images in CLIP."}
|
|
{"id": 14412, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "analysis", "why": "Analysis of Music AVQA data and methods motivating specialized designs"}], "pattern": "vary-conditions", "fixes": ["methods"], "varies": ["designs"], "benchmarks": [], "results": [], "workflow": [], "summary": "Systematic analysis of Music AVQA showing specialized multimodal designs are needed."}
|