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
31 KiB
JSON
61 lines
31 KiB
JSON
{"id": 7366, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "theory", "why": "Proves optimal Bayesian stopping policy for self-consistency sampling with L=3 aggregation"}], "pattern": "vary-scaffold", "fixes": ["LLM", "sampling"], "varies": ["stopping policy"], "benchmarks": [], "results": [], "workflow": ["Sample answers", "Track top counts", "Stop when consistent"], "summary": "Bayesian optimal stopping rule for self-consistency sampling that saves samples with asymptotic optimality guarantees."}
|
|
{"id": 11079, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "Framework defining refusal criteria and evaluation under benign and adversarial conditions for security agents"}, {"area": "agent-safety", "role": "benchmark", "why": "Assesses LLM agents on web offensive-security refusal scenarios"}], "pattern": "vary-conditions", "fixes": ["agents under test"], "varies": ["refusal criteria", "benign vs adversarial prompts"], "benchmarks": [], "results": [], "workflow": [], "summary": "First framework for refusal boundaries of cybersecurity agents, evaluating adherence under benign and adversarial conditions."}
|
|
{"id": 18300, "in_agent_scope": false, "types": [{"area": "train/distill", "role": "method", "why": "Multi-agent teacher distillation into small LLMs for Vietnamese sentiment analysis"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["distillation method"], "benchmarks": [], "results": [], "workflow": ["Detect", "Process", "Generate rationale", "Aggregate"], "summary": "Multi-agent knowledge distillation framework where teacher agents transfer rationales to 1-1.5B students for sentiment analysis."}
|
|
{"id": 7624, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "2000-instance multi-language benchmark of agentic coding from GitHub PRs"}], "pattern": "fix-agent-vary-env", "fixes": ["agent frameworks", "LLMs"], "varies": ["task types", "languages", "scenarios"], "benchmarks": [{"name": "SWE-Compass", "purpose": "unified agentic coding evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Comprehensive agentic coding benchmark across 8 task types and 10 languages evaluating ten LLMs under SWE-Agent and Claude Code."}
|
|
{"id": 2047, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Self-supervised RL on K-step GUI transition inverse dynamics for VLM GUI agents"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training task", "RL"], "benchmarks": [{"name": "AndroidControl", "purpose": "GUI task automation", "is_new": false}, {"name": "GUI Odyssey", "purpose": "GUI task automation", "is_new": false}, {"name": "ScreenSpot-v2", "purpose": "GUI grounding", "is_new": false}, {"name": "ScreenSpot-Pro", "purpose": "GUI grounding", "is_new": false}], "results": [], "workflow": ["Build transition data", "Filter data", "Rule-based RL"], "summary": "Self-supervised RL framework training VLM GUI agents to predict actions between GUI states without instruction annotations."}
|
|
{"id": 8264, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Analyzes length bias in MCQ scoring and proposes Bayesian accuracy"}], "pattern": "vary-evaluation", "fixes": ["models"], "varies": ["scoring rule"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analysis of length bias in likelihood-based multiple-choice accuracy and a Bayesian accuracy scoring rule replacing it."}
|
|
{"id": 13686, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Inference-time value-of-information rule deciding when agents ask clarifying questions"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["ask-vs-act policy"], "benchmarks": [{"name": "20 Questions", "purpose": "clarification domain", "is_new": false}, {"name": "flight booking", "purpose": "clarification domain", "is_new": false}, {"name": "e-commerce", "purpose": "clarification domain", "is_new": false}], "results": [], "workflow": ["Estimate utility gain", "Compare to user cost", "Ask or act"], "summary": "Decision-theoretic VoI framework letting LLM agents weigh information gain against user cost when asking questions."}
|
|
{"id": 9689, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Driving world model predicting latent BEV features plus text reasoning for end-to-end driving"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["world model method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Driving world model with parallel latent BEV future prediction and adaptive text reasoning for long-tail driving."}
|
|
{"id": 14796, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Agentic workflow producing offline SQL+Jinja templates for query-focused table summarization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow"], "benchmarks": [], "results": [], "workflow": ["Generate SQL", "Generate template", "Render summary"], "summary": "Agentic workflow generating reusable offline SQL and Jinja2 templates to summarize tables privately and efficiently."}
|
|
{"id": 10736, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Offline agent learning for hardware testbench generation with execution feedback"}, {"area": "code-swe", "role": "data", "why": "Execution-validated curation and policy-aware synthesis of training data"}], "pattern": "fix-env-vary-agent", "fixes": ["verification simulator"], "varies": ["training data", "sampling"], "benchmarks": [], "results": [], "workflow": ["Curate validated data", "Synthesize policy-aware data", "Prioritize worst states"], "summary": "Offline learning framework with execution-validated data curation lets a compact 4B model generate high-coverage testbenches."}
|
|
{"id": 16671, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Bayesian inference sampling for training-free workflow generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow generation algorithm"], "benchmarks": [], "results": [], "workflow": ["Look-ahead rollouts", "Importance weighting", "Refine pool"], "summary": "Training-free Bayesian workflow generation via importance-weighted look-ahead rollouts, improving accuracy up to 9 points."}
|
|
{"id": 15815, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Dynamic discourse tree to manage non-linear dialogue history"}, {"area": "memory-selfevolve", "role": "benchmark", "why": "Introduces Non-linear Task Multi-turn dialogue benchmark"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["context structure"], "benchmarks": [{"name": "Non-linear Task Multi-turn Dialogue", "purpose": "non-linear dialogue evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Context-Agent models dialogue history as a dynamic tree and introduces a benchmark for non-linear multi-turn tasks."}
|
|
{"id": 14304, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "method", "why": "Two-stage MQM-guided diagnose-then-repair post-editing for domain MT"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["post-editing pipeline"], "benchmarks": [], "results": [], "workflow": ["MQM diagnosis", "Minimal edit repair"], "summary": "Evaluator-guided post-editing: an LLM gives span-level MQM diagnoses and another applies minimal edits, improving COMET."}
|
|
{"id": 15130, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Manager-centric multi-agent framework with memory and multi-branch reasoning for insight discovery"}, {"area": "multi-agent", "role": "scaffold", "why": "Planner, Executor, Manager coordination"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["Manager coordinates", "Planner subgoals", "Executor acts"], "summary": "Multi-agent framework for automated data insight discovery with manager, dual-layer memory and multi-role planning."}
|
|
{"id": 11116, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Self-distillation for task decomposition then diversity-aware RL for tool-using reasoning models"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training recipe"], "benchmarks": [{"name": "BFCLv3", "purpose": "tool-use accuracy", "is_new": false}], "results": [{"benchmark": "BFCLv3", "system": "D-CORE-8B", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 77.7, "unit": "%"}], "workflow": ["Self-distill decomposition", "Diversity-aware RL"], "summary": "Two-stage training teaching large reasoning models task decomposition for complex tool use."}
|
|
{"id": 12902, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Iterative triple construction-integration retrieval for multi-hop QA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval procedure"], "benchmarks": [], "results": [], "workflow": ["Construct triples", "Integrate", "Next-hop query"], "summary": "Training-free iterative RAG preserving multiple evidence chains and adaptive evidence granularity for multi-hop QA."}
|
|
{"id": 8886, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Multi-step tool tasks with shortcut opportunities measuring reward hacking"}, {"area": "agent-safety", "role": "analysis", "why": "Compares exploit rates across 13 models and RL post-training"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["models", "chain length"], "benchmarks": [{"name": "Reward Hacking Benchmark", "purpose": "exploit rate of agents", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark measuring reward-hacking exploits of tool-using LLM agents across 13 frontier models."}
|
|
{"id": 4420, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Implicit step rewards via jointly trained implicit PRM for agentic RL credit assignment"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["credit assignment reward"], "benchmarks": [], "results": [], "workflow": ["Train implicit PRM", "Derive step rewards", "Update policy"], "summary": "iStar assigns step-level credit in agentic RL with an implicit process reward model trained alongside the policy."}
|
|
{"id": 14761, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "scaffold", "why": "Multi-agent cultural debate to mitigate Western-centric bias"}, {"area": "safety/fairness", "role": "benchmark", "why": "CEBiasBench Chinese-English bias benchmark"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["debate mitigation"], "benchmarks": [{"name": "CEBiasBench", "purpose": "cultural bias measurement", "is_new": true}], "results": [], "workflow": [], "summary": "Bilingual cultural-bias benchmark and multi-agent cultural debate mitigating Western-centric bias."}
|
|
{"id": 14826, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Layered agent network optimized via textual backpropagation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["team structure"], "benchmarks": [], "results": [], "workflow": ["Decompose tasks", "Build layered teams", "Textual backprop"], "summary": "Agentic Neural Network treats agent teams as layers and refines them through forward decomposition and textual backward feedback."}
|
|
{"id": 2183, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "PPO with verifiable emotion rewards from simulated users for empathetic dialogue"}], "pattern": "fix-env-vary-agent", "fixes": ["simulated users"], "varies": ["RL reward"], "benchmarks": [{"name": "Sentient-Benchmark", "purpose": "empathy score", "is_new": false}], "results": [{"benchmark": "Sentient-Benchmark", "system": "RLVER", "base_model": "Qwen2.5-7B-Instruct", "setting": "finetuned", "metric": "score", "score": 79.2, "unit": "other"}, {"benchmark": "Sentient-Benchmark", "system": "Qwen2.5-7B-Instruct", "base_model": "Qwen2.5-7B-Instruct", "setting": "baseline", "metric": "score", "score": 13.3, "unit": "other"}], "workflow": ["Simulated user dialogue", "Emotion score reward", "PPO"], "summary": "RL framework using simulated-user emotion scores as verifiable rewards to improve LLM empathy."}
|
|
{"id": 10355, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Evidence-driven agentic MLLM framework for explainable AI-generated video detection"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["detection framework"], "benchmarks": [], "results": [], "workflow": ["Plan strategy", "Evidence reasoning", "Graph deliberation"], "summary": "Hermes agentic framework plans detection strategy and verifies evidence for trustworthy AIGV detection."}
|
|
{"id": 1192, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Agentic continual pre-training for deep research agent foundation model"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["pretraining stage"], "benchmarks": [], "results": [], "workflow": ["Agentic CPT", "Post-training"], "summary": "Proposes Agentic CPT to build agentic foundation models; AgentFounder-30B evaluated on 10 benchmarks."}
|
|
{"id": 2632, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Automated scalable generation of verifiable multi-tool agentic tasks for SFT/RL"}, {"area": "search-research", "role": "environment", "why": "Task generation workflow"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training tasks"], "benchmarks": [{"name": "GAIA", "purpose": "cited existing agent benchmark", "is_new": false}, {"name": "BrowseComp", "purpose": "cited existing agent benchmark", "is_new": false}], "results": [], "workflow": ["Atomic tasks", "Extend depth/width", "Validate", "Sample trajectories"], "summary": "TaskCraft generates verifiable agentic tasks via depth and width extension, improving multi-hop reasoning when used for training."}
|
|
{"id": 11185, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "LLM generates basis terms pruned by influence scores for symbolic regression"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search feedback"], "benchmarks": [], "results": [], "workflow": ["LLM proposes terms", "Influence scoring", "Prune"], "summary": "LLM-driven symbolic regression using granular influence-score feedback to prune candidate equation terms."}
|
|
{"id": 13018, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Multi-agent pipeline synthesizing function-calling training data"}], "pattern": "vary-data", "fixes": ["8B model", "training recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": ["Take reliable tools", "Generate dialogues", "Multi-stage evaluation"], "summary": "GenesisFunc generates diverse, verified function-calling dialogues with multi-agent framework to fine-tune an 8B LLM."}
|
|
{"id": 4566, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Six-game benchmark with modular harness to probe LLM/VLM capabilities"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["games", "harness modules"], "benchmarks": [{"name": "LMGame-Bench", "purpose": "game-playing evaluation of 13 models", "is_new": true}], "results": [], "workflow": [], "summary": "Gym-style game benchmark with toggleable perception, memory, reasoning modules evaluating 13 models."}
|
|
{"id": 6202, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free hierarchical scene graph generation with geometry-guided relations"}], "pattern": "n/a", "fixes": ["VLM"], "varies": ["inference framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Training-free open-world scene graph generation using hierarchical entities and geometry-guided relation filtering."}
|
|
{"id": 8013, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "1,150-task enterprise benchmark for stateful planning and tool use"}, {"area": "planning-decision", "role": "environment", "why": "Containerized enterprise sandbox with database tables"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["enterprise tasks", "sandbox"], "benchmarks": [{"name": "EnterpriseOps-Gym", "purpose": "enterprise agentic planning", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark environment with expert tasks and a containerized sandbox evaluating enterprise agentic planning."}
|
|
{"id": 16130, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "File-centric state externalization keeping agent context bounded"}], "pattern": "vary-scaffold", "fixes": ["20B model"], "varies": ["context management"], "benchmarks": [{"name": "DeepResearch", "purpose": "long-horizon evaluation", "is_new": false}, {"name": "80-paper literature review", "purpose": "long-horizon coverage", "is_new": false}], "results": [], "workflow": ["Externalize state", "Snapshot workspace", "Recent action window"], "summary": "Training-free framework externalizing state to files so context stays bounded on long-horizon tasks."}
|
|
{"id": 8567, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "In-context coordination transformer adapting to unfamiliar partners in multi-agent RL"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Coordination Transformers use in-context learning to adapt to diverse partner behaviors."}
|
|
{"id": 11515, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "VeriCode taxonomy of verifiable code instructions for instruction-following evaluation"}, {"area": "reason/code", "role": "eval-method", "why": "Aligns code evaluation with human preference beyond pass@k"}], "pattern": "vary-evaluation", "fixes": ["models"], "varies": ["evaluation metric"], "benchmarks": [{"name": "VeriCode", "purpose": "code instruction following", "is_new": true}], "results": [], "workflow": [], "summary": "Code evaluation of verifiable non-functional instruction following aligned with human preference."}
|
|
{"id": 9760, "in_agent_scope": false, "types": [{"area": "efficiency/compress", "role": "method", "why": "Layer-sensitivity-guided quantization for one-step diffusion video super-resolution"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["quantization method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Quantization with dynamic range adaptive quantizer and sensitivity-guided layers for diffusion video super-resolution."}
|
|
{"id": 10679, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Agentic RL combining behavior enhancement for proactive agents balancing performance and engagement"}], "pattern": "fix-env-vary-agent", "fixes": ["user environment"], "varies": ["RL framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "BAO agentic RL framework improving proactive LLM agents' task performance versus user engagement trade-off."}
|
|
{"id": 11689, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Multi-scale probabilistically consistent action generation for robot imitation learning"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Coarse-to-fine multi-scale autoregressive action generation for long-horizon robotic imitation."}
|
|
{"id": 14052, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Benchmark and evaluation suite for academic webpage generation"}, {"area": "web-gui", "role": "application", "why": "PWAgent system for paper-to-website generation"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["tasks"], "benchmarks": [{"name": "Paper2Web", "purpose": "academic webpage generation", "is_new": true}], "results": [], "workflow": [], "summary": "Paper2Web benchmark and multi-dimensional evaluation for academic webpage generation, plus an agent for it."}
|
|
{"id": 8291, "in_agent_scope": true, "types": [{"area": "search-research", "role": "eval-method", "why": "Two-layer expert-grounded claim-level evaluation for open-ended professional tasks"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["evaluation protocol"], "benchmarks": [], "results": [], "workflow": [], "summary": "JADE combines predefined expert evaluation skills with claim-level assessment to judge open-ended agent reports."}
|
|
{"id": 5599, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "LLM-driven stochastic generative optimization with priority queue for prompts and agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["optimizer"], "benchmarks": [], "results": [], "workflow": ["Priority queue", "Local aggregation", "Evaluate candidates"], "summary": "POLCA framework for LLM-based optimization of prompts and agents under noisy feedback."}
|
|
{"id": 16318, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Dynamic agent panel pruning via Nash selection and doubly-robust credits for VLM MAS"}], "pattern": "vary-scaffold", "fixes": ["VLMs"], "varies": ["panel selection"], "benchmarks": [], "results": [], "workflow": ["Learn agent values", "Online Nash pruning"], "summary": "Cost-efficient VLM multi-agent system pruning agents dynamically using causal value learning and Nash selection."}
|
|
{"id": 7114, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Dense screen-parsing dataset of 771K screenshots for CUA grounding"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [{"name": "ScreenParse", "purpose": "complete screen parsing supervision", "is_new": true}], "results": [], "workflow": [], "summary": "Large dataset with dense UI element annotations generated by automated Webshot pipeline for efficient screen parsing."}
|
|
{"id": 17693, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "analysis", "why": "Linguistic coordination analysis in novel dialogue"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [{"name": "PDNC", "purpose": "novel dialogue corpus", "is_new": false}], "results": [], "workflow": [], "summary": "Shows linguistic convergence in fictional dialogues of Austen and Forster using LIWC and network measures."}
|
|
{"id": 6558, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Activation-based detection of malicious agents in multi-agent systems"}], "pattern": "vary-conditions", "fixes": ["threat model"], "varies": ["detector"], "benchmarks": [], "results": [], "workflow": [], "summary": "AcMAS detects stealthy malicious behavior in LLM multi-agent systems from internal activation states."}
|
|
{"id": 10483, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Benchmark of real multi-file refactorings with tests and static checks"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["refactoring tasks"], "benchmarks": [{"name": "CodeTaste", "purpose": "refactoring ability", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark mined from open-source refactorings testing whether coding agents execute and choose human-like refactorings."}
|
|
{"id": 13132, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Multilingual persuasion detection benchmark of human vs LLM text"}, {"area": "safety/security", "role": "analysis", "why": "Compares detectability of AI vs human persuasion"}], "pattern": "fix-agent-vary-env", "fixes": ["detectors"], "varies": ["text source", "languages"], "benchmarks": [{"name": "Persuaficial", "purpose": "persuasion detection", "is_new": true}], "results": [], "workflow": [], "summary": "Six-language benchmark showing detectability differences between human and LLM-generated persuasion."}
|
|
{"id": 14927, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Intent-indexed agent memory retrieval matching contextual intent"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory indexing"], "benchmarks": [], "results": [], "workflow": ["Index with intent", "Match current intent", "Filter snippets"], "summary": "STITCH indexes trajectory steps by contextual intent to retrieve context-matched memories for long-horizon agents."}
|
|
{"id": 6751, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Agentic thinking-with-videos model jointly learning grounding and QA"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["model training"], "benchmarks": [], "results": [], "workflow": ["Localize", "Clip", "Refine", "Answer"], "summary": "Unified agentic video model with on-demand clipping and refined temporal localization for long-video understanding."}
|
|
{"id": 15065, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "State-aware fine-grained memory at reasoning-thought level"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory retrieval"], "benchmarks": [], "results": [], "workflow": [], "summary": "SAMem retrieves state-specific reasoning thoughts to improve LLM agent multi-step decisions."}
|
|
{"id": 18396, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "survey", "why": "Survey of 21 linguistic annotation tools with interactive platform"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Survey of 21 manual linguistic annotation tools documenting 50+ features in an open interactive dataset."}
|
|
{"id": 18440, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "analysis", "why": "Reading-time analysis of syntactic information maintenance in Japanese"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows syntactically guided maintenance costs, predicted heads and incomplete dependencies, in Japanese reading times."}
|
|
{"id": 12667, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Model-aware iterative data evolution and label verification"}, {"area": "tool-use", "role": "train-agent", "why": "Closed data-training loop for tool calling"}], "pattern": "vary-data", "fixes": ["environment"], "varies": ["data and model co-evolution"], "benchmarks": [], "results": [], "workflow": ["Probe capabilities", "Verify labels", "Evolve data", "Train"], "summary": "LoopTool couples data synthesis and training, probing weaknesses and verifying labels to improve tool calling."}
|
|
{"id": 17427, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "scaffold", "why": "Selective debate initiation and evidence-weighted self-consistency for efficient debate"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["debate procedure"], "benchmarks": [], "results": [], "workflow": ["Detect disagreement", "Selective debate", "Weighted aggregation"], "summary": "Lightweight debate-on-demand framework triggering debate only on disagreement and aggregating by evidence weight."}
|
|
{"id": 13262, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Agentic RL training for multi-turn text-to-SQL with execution and verification"}], "pattern": "fix-env-vary-agent", "fixes": ["database environment"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["Propose SQL", "Execute", "Verify", "Refine"], "summary": "MTSQL-R1 trains agent in propose-execute-verify-refine loop for long-horizon multi-turn Text-to-SQL."}
|
|
{"id": 638, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "train-agent", "why": "R1-style RL fine-tuning of traffic simulation next-token models"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["training method"], "benchmarks": [], "results": [], "workflow": [], "summary": "SMART-R1 reinforcement fine-tuning aligns multi-agent traffic simulation behavior with human preferences and metrics."}
|
|
{"id": 15634, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "analysis", "why": "Refinement Gap metric and scaling of parallel self-refinement"}, {"area": "reason/cot", "role": "train-agent", "why": "Trains models to refine parallel reasoning"}], "pattern": "vary-conditions", "fixes": ["methods"], "varies": ["model size"], "benchmarks": [], "results": [], "workflow": [], "summary": "Introduces Refinement Gap metric, shows scaling trends of parallel self-refinement, and trains models to refine."}
|
|
{"id": 18030, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "data", "why": "Swedish parliamentary motions corpus 1867-2024"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [{"name": "Swedish Parliamentary Motions Corpus", "purpose": "new research corpus", "is_new": true}], "results": [], "workflow": [], "summary": "Annotated corpus of Swedish parliamentary motions over 150 years with author metadata."}
|
|
{"id": 10728, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "260-task data visualization agent benchmark with spreadsheet, evolution and interaction domains"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "DV-World", "purpose": "data visualization agents", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark for data visualization agents across native spreadsheet, adaptation, and proactive intent alignment."}
|
|
{"id": 12439, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "defense", "why": "Lightweight KB-aligned OOD query detection gate for RAG"}], "pattern": "vary-conditions", "fixes": ["RAG"], "varies": ["detector"], "benchmarks": [], "results": [], "workflow": [], "summary": "PCA-based KB-aligned OOD detection gating queries for safer RAG across 16 domains."}
|
|
{"id": 5182, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Cold-start plus RL for tool-invoking multimodal model"}, {"area": "tool-use", "role": "data", "why": "Curated training data where tool use helps"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training pipeline", "data"], "benchmarks": [], "results": [], "workflow": ["Cold-start SFT", "RL refinement"], "summary": "DeepEyesV2 builds agentic multimodal model via data curation and two-stage cold-start plus RL training."}
|
|
{"id": 13568, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Multimodal cultural QA benchmark of 15 African languages"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["languages", "modality"], "benchmarks": [{"name": "Afri-MCQA", "purpose": "African cultural multimodal QA", "is_new": true}], "results": [], "workflow": [], "summary": "First multilingual cultural QA benchmark with parallel text and speech across 15 African languages."}
|
|
{"id": 12407, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "scaffold", "why": "Training-free KG reasoning with relational blueprints and failure-aware refinement"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["KG reasoning procedure"], "benchmarks": [], "results": [], "workflow": [], "summary": "CoG training-free framework for LLM reasoning over knowledge graphs with blueprint guidance and refinement."}
|