Files
ai-taxonomy-literature-anal…/data/labels/o27.jsonl
T
zzyang d6fc7c1dfa Add types-of-work classification, benchmarks and score pages
- Work type = area x role (fixed vs varied), vocabulary in scripts/worktypes.json
- LLM-read labels for 2,678 agent-related papers (data/labels), rule labels for the rest
- Per-type synthesis notes (why-same, sub-variants, benchmarks, workflow)
- Pages: /work-types matrix and detail (timeline, benchmarks, workflow, score charts), /benchmarks
- build_worktypes.py adds work-type tables to data/papers.db

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bhs9LU8eMixk6g1kb1d3j7
2026-10-21 11:55:46 +00:00

61 lines
26 KiB
JSON

{"id": 11672, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Activation-guided perturbation subspaces for zeroth-order LLM fine-tuning"}], "pattern": "vary-data", "fixes": ["benchmarks"], "varies": ["ZO perturbation scheme"], "benchmarks": [], "results": [], "workflow": [], "summary": "Zeroth-order fine-tuning using activation-informed subspaces for memory-efficient LLM tuning."}
{"id": 10038, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "Game-based benchmark for multi-agent reasoning and theory of mind"}], "pattern": "fix-env-vary-agent", "fixes": ["LLMs under test"], "varies": ["game tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "Decrypto, an interactive game-based benchmark for multi-agent reasoning and theory of mind in LLMs."}
{"id": 13753, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Small planner injects reasoning strategies into LLM chain of thought via RL"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["strategy injection planner"], "benchmarks": [], "results": [], "workflow": [], "summary": "Reinforced strategy injection with a small planner lets any LLM exhibit reasoning-model behaviour."}
{"id": 15853, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "LLM-generated synthetic data bootstraps interpretable Tsetlin Machine text classifiers"}], "pattern": "vary-data", "fixes": ["classifier"], "varies": ["synthetic data curriculum"], "benchmarks": [], "results": [], "workflow": [], "summary": "Transfers LLM semantic knowledge into Tsetlin Machines for interpretable text classification."}
{"id": 2644, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Hybrid on/off-policy RL using memory for exploration in LLM agents"}, {"area": "planning-decision", "role": "train-agent", "why": "RL exploration for novel state discovery"}], "pattern": "fix-env-vary-agent", "fixes": ["ScienceWorld, WebShop"], "varies": ["RL algorithm", "memory use"], "benchmarks": [], "results": [], "workflow": [], "summary": "EMPO2 hybrid RL framework with memory-driven exploration improving over GRPO for LLM agents."}
{"id": 17496, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "analysis", "why": "Analyzes proactivity and multimodal features in task-guidance dialogue corpus"}], "pattern": "vary-conditions", "fixes": ["dialogue corpus"], "varies": ["multimodal features"], "benchmarks": [], "results": [], "workflow": [], "summary": "Corpus analysis of proactive behaviour in multimodal task-guidance dialogues."}
{"id": 11714, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Training-free hierarchical multi-agent framework generating scientific posters via Poster Tree"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["poster generation pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "PosterForest: training-free multi-agent scientific poster generation with a hierarchical Poster Tree representation."}
{"id": 16653, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Knowledge-grounded user simulator with business-logic personas for agent evaluation"}, {"area": "social-sim", "role": "eval-method", "why": "Simulated users replace human assessment for multi-turn agents"}], "pattern": "vary-evaluation", "fixes": ["agents under test"], "varies": ["user simulator"], "benchmarks": [], "results": [], "workflow": [], "summary": "SAGE user simulator grounded in business knowledge for realistic multi-turn agent evaluation."}
{"id": 7774, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Gaze-guided attention and cross-modal fusion to reduce VLM hallucination"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["attention guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "Mitigates VLM hallucination by correcting visual attention sinks and cross-modal fusion balance."}
{"id": 13012, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "environment", "why": "Automatically synthesizes functional interconnected web environments at scale for GUI agent training"}], "pattern": "fix-agent-vary-env", "fixes": ["GUI agent"], "varies": ["web environments"], "benchmarks": [], "results": [], "workflow": [], "summary": "InfiniteWeb generates scalable functional websites for training GUI agents."}
{"id": 7390, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "Benchmark of 100 narrative environments testing agents' narrative commitment against user interventions"}], "pattern": "fix-env-vary-agent", "fixes": ["LLM agents under test"], "varies": ["narrative environments"], "benchmarks": [], "results": [], "workflow": [], "summary": "NCP-Bench evaluates whether LLM agents keep narrative consistency under unconstrained user interventions."}
{"id": 9284, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Selective memory sharing across parallel agent teams to cut redundant computation"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "Learned sharing policy for memory"}], "pattern": "vary-scaffold", "fixes": ["agent models"], "varies": ["memory sharing"], "benchmarks": [], "results": [], "workflow": [], "summary": "Learning to Share reduces redundant computation across parallel agent teams via selective memory."}
{"id": 9853, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Multi-agent closed-loop framework for composite-instruction expressive TTS"}], "pattern": "vary-scaffold", "fixes": ["TTS model"], "varies": ["control framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "AgentSteerTTS: multi-agent closed-loop control of composite instructions in text-to-speech."}
{"id": 5735, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Trains orchestrator to design multi-agent systems holistically"}, {"area": "multi-agent", "role": "benchmark", "why": "Controlled benchmarks to study when MAS beats single agents"}], "pattern": "vary-both", "fixes": ["tasks"], "varies": ["orchestration design"], "benchmarks": [], "results": [], "workflow": [], "summary": "MAS-Orchestra trains holistic MAS orchestration and offers controlled benchmarks comparing MAS with single agents."}
{"id": 14814, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Dynamic simulation from social corpus for persona-level role-playing"}, {"area": "social-sim", "role": "benchmark", "why": "Evaluates persona-level role-playing in LLMs"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["social scenarios"], "benchmarks": [], "results": [], "workflow": [], "summary": "PersonaArena dynamic simulation for evaluating and improving persona-level role-playing."}
{"id": 15079, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "End-to-end RL for search agent maintaining compact memory"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Compact iterative memory keeps context stable"}], "pattern": "fix-env-vary-agent", "fixes": ["search benchmarks"], "varies": ["memory design", "RL training"], "benchmarks": [], "results": [], "workflow": [], "summary": "MemSearcher keeps compact memory across turns and is trained end-to-end with RL."}
{"id": 11656, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Harmonic memory representation balancing abstraction and specificity with cue anchors"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory representation"], "benchmarks": [], "results": [], "workflow": [], "summary": "Memora: memory representation balancing abstraction and specificity for agent retrieval."}
{"id": 18138, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Emotion-conditioned representations with learnable psycholinguistic mapping for personality prediction"}], "pattern": "vary-data", "fixes": ["datasets"], "varies": ["model design"], "benchmarks": [], "results": [], "workflow": [], "summary": "Big Five personality prediction using emotion-conditioned representations."}
{"id": 13408, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Multi-graph agentic memory with semantic, temporal, causal, entity graphs"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "MAGMA multi-graph memory architecture with query-intent-aligned retrieval."}
{"id": 13600, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "eval-method", "why": "Generative projective testing as psychometric tool for LLM agents"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["psychometric protocol"], "benchmarks": [], "results": [], "workflow": [], "summary": "GenPT adapts projective tests to measure LLM psychological traits more reliably than self-report."}
{"id": 12901, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Dual-layer memory and multi-round debate for long-horizon GUI automation"}], "pattern": "vary-scaffold", "fixes": ["MLLM"], "varies": ["memory", "debate verification"], "benchmarks": [], "results": [], "workflow": [], "summary": "MaDS combines dual-layer memory and debate to improve long-horizon GUI agents."}
{"id": 6462, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "Expert-guided query-based model collaboration for clinical text augmentation"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["augmentation method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Clinical text augmentation integrating expert knowledge via model collaboration."}
{"id": 3066, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Graph-based router generating routing workflows across heterogeneous agents"}, {"area": "planning-decision", "role": "train-agent", "why": "Learned workflow generation"}], "pattern": "vary-scaffold", "fixes": ["LLM pool"], "varies": ["routing"], "benchmarks": [], "results": [], "workflow": [], "summary": "GraphPlanner heterogeneous-graph agentic router for LLM routing workflows."}
{"id": 2423, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "theory", "why": "Models LLM-controlled sequential social learning with information-mediating planner"}], "pattern": "n/a", "fixes": [], "varies": ["planner information structure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Formal model of controlled sequential social learning where an LLM planner shapes information."}
{"id": 10655, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Training-free test-time policy optimization using non-parametric experience memory"}], "pattern": "vary-scaffold", "fixes": ["frozen weights"], "varies": ["memory-based advantage estimation"], "benchmarks": [], "results": [], "workflow": [], "summary": "JitRL enables continual LLM agent adaptation without gradient updates via retrieved experience."}
{"id": 218, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "VLMs visually guide subproblem selection for large-scale TSP"}], "pattern": "vary-scaffold", "fixes": ["benchmarks"], "varies": ["solver framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "ViTSP uses VLMs to guide large-scale travelling salesman problem solving."}
{"id": 14029, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "Library evaluating entire multi-agent systems rather than models"}, {"area": "multi-agent", "role": "benchmark", "why": "System-level evaluation across topology and orchestration"}], "pattern": "vary-evaluation", "fixes": ["tasks"], "varies": ["system components"], "benchmarks": [], "results": [], "workflow": [], "summary": "MASEval treats whole agentic system as unit of analysis for multi-agent evaluation."}
{"id": 13736, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Verifiable synthetic data pipeline for ambiguous, changing, infeasible user requests"}, {"area": "tool-use", "role": "train-agent", "why": "Trains robust tool-calling agents"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": [], "summary": "Trajectory2Task synthesizes verifiable tasks to train tool-calling agents on complex user intents."}
{"id": 13062, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "analysis", "why": "Dual-view explainability of LLM-based VLN agent failures"}], "pattern": "vary-conditions", "fixes": ["VLN agents"], "varies": ["explanation method"], "benchmarks": [], "results": [], "workflow": [], "summary": "DEFT explains LLM navigation agent failures by jointly analyzing decision timing and visual evidence."}
{"id": 13068, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Training-free calibrated speculative decoding recovering falsely rejected tokens"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["verification rule"], "benchmarks": [], "results": [], "workflow": [], "summary": "CSD recovers valid draft tokens for faster speculative decoding."}
{"id": 13933, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Structured hierarchical memory preserving event bindings for long-term conversational agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory structure"], "benchmarks": [], "results": [], "workflow": [], "summary": "StructMem structure-enriched hierarchical memory for temporal and multi-hop reasoning."}
{"id": 5614, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "Standardized framework unifying data science agent benchmarks with data-grounded tasks"}, {"area": "science-data", "role": "environment", "why": "Execution environments for evaluating and training data science agents"}], "pattern": "fix-env-vary-agent", "fixes": ["agents"], "varies": ["tasks", "environments"], "benchmarks": [], "results": [], "workflow": [], "summary": "DSGym standardizes evaluation and training of data science agents in self-contained execution environments."}
{"id": 15680, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "Long-term memory benchmark covering remembering, reasoning, recommending with updates"}], "pattern": "fix-env-vary-agent", "fixes": ["memory systems"], "varies": ["tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "Memora benchmark for long-term memory of personalized agents across weeks-long conversations."}
{"id": 6485, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "VLM-guided pose-control agent with frozen perception for unsupervised domain adaptation"}], "pattern": "vary-scaffold", "fixes": ["perception modules"], "varies": ["pose control agent"], "benchmarks": [], "results": [], "workflow": [], "summary": "Sea2 adapts deployment via active pose-control agent instead of finetuning perception models."}
{"id": 6115, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Compiles task descriptions into executable code to cut computer-use agent latency"}], "pattern": "vary-scaffold", "fixes": ["model"], "varies": ["planning/execution strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "Agent JIT compilation turns tasks into parallelizable code for faster web agents."}
{"id": 15431, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Psychologically-guided dynamic self-exploration reasoning for personality detection"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["reasoning framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "PsyPath lets LLMs generate and answer psychology-guided questions for personality detection."}
{"id": 16862, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Teams of LLM agents exploit real-world zero-day vulnerabilities"}, {"area": "multi-agent", "role": "scaffold", "why": "Planning agent with task-specific subagents"}], "pattern": "vary-scaffold", "fixes": ["vulnerabilities"], "varies": ["agent team structure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows teams of LLM agents can exploit zero-day vulnerabilities better than single agents."}
{"id": 16550, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Dataset of 2,941 debatable questions with multi-perspective answers"}, {"area": "eval/judge", "role": "eval-method", "why": "Perspective Diversity and Dispute Awareness metrics"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["metrics", "questions"], "benchmarks": [], "results": [], "workflow": [], "summary": "DebateQA dataset and metrics for evaluating answers to debatable questions."}
{"id": 15555, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "method", "why": "Dynamic PMI contrastive decoding with input transformations to reduce hallucination"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["decoding"], "benchmarks": [], "results": [], "workflow": [], "summary": "Unified contrastive decoding framework using dynamic PMI to reduce hallucination."}
{"id": 13527, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "train-agent", "why": "Multi-agent RL for hallucination checking in RAG"}], "pattern": "vary-scaffold", "fixes": ["RAG"], "varies": ["verifier training"], "benchmarks": [], "results": [], "workflow": [], "summary": "MARCH uses reinforced multi-agent self-check to detect hallucination while avoiding confirmation bias."}
{"id": 6185, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Optimizes repo-specific documentation as playbooks for coding agents"}], "pattern": "vary-scaffold", "fixes": ["coding agent"], "varies": ["documentation"], "benchmarks": [], "results": [], "workflow": [], "summary": "Black-box optimization of agent-oriented documentation judged by downstream code correctness."}
{"id": 933, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "Multi-source information seeking benchmark with MCP tools"}, {"area": "tool-use", "role": "benchmark", "why": "Tests integration of domain tools with web search"}], "pattern": "fix-env-vary-agent", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "InfoMosaic-Bench evaluates multi-source information seeking in tool-augmented agents."}
{"id": 17469, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "Multi-task RL with CoT for paralinguistic understanding in speech LLMs"}], "pattern": "vary-data", "fixes": ["benchmarks"], "varies": ["RL training"], "benchmarks": [], "results": [], "workflow": [], "summary": "Multi-task RL aligns speech LLMs to paralinguistic cues."}
{"id": 11670, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Multi-agent LLM-driven evolutionary search for inverse microstructure design"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "AutoMS multi-agent neuro-symbolic evolutionary search for cross-physics microstructure design."}
{"id": 18657, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "analysis", "why": "Tests whether formal and functional linguistic mechanisms are dissociated in LLMs"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["tasks/mechanisms"], "benchmarks": [], "results": [], "workflow": [], "summary": "Mechanistic study of dissociation between formal and functional linguistic abilities in LLMs."}
{"id": 7047, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "Compares multi-agent communication protocols on success, latency, overhead, robustness"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["protocols"], "benchmarks": [], "results": [], "workflow": [], "summary": "ProtocolBench benchmarks agent communication protocols along four axes."}
{"id": 17244, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Cognitive multi-agent in-context exploration with declarative and procedural history"}], "pattern": "vary-scaffold", "fixes": ["model"], "varies": ["exploration framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "HintNavigator multi-agent in-context exploration for dynamic web navigation."}
{"id": 1948, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Multi-agent framework for interactive embodied theatrical drama"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "HAMLET multi-agent framework for live interactive theatre with scene interaction."}
{"id": 6598, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "First benchmark for agentic kernel generation on emerging hardware ISAs"}], "pattern": "fix-env-vary-agent", "fixes": ["agents"], "varies": ["hardware tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "KernelCraft benchmarks LLM agents generating low-level kernels for new accelerators."}
{"id": 10600, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "KV cache sharing across multi-LoRA agents"}], "pattern": "vary-scaffold", "fixes": ["backbone"], "varies": ["cache sharing"], "benchmarks": [], "results": [], "workflow": [], "summary": "LRAgent shares KV cache among multi-LoRA agents to cut memory and compute."}
{"id": 14541, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Template-guided RL with MCTS-built template library for reasoning"}], "pattern": "vary-data", "fixes": ["benchmarks"], "varies": ["RL guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "TemplateRL augments GRPO-style RL with structured template guidance."}
{"id": 14024, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Monitoring and visualization of scientific agent execution traces as graphs"}], "pattern": "n/a", "fixes": [], "varies": ["trace visualization"], "benchmarks": [], "results": [], "workflow": [], "summary": "Graph of Trace records and visualizes scientific agent workflows for human review."}
{"id": 18443, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "analysis", "why": "Age-based plasticity to reduce language drift in neural emergent communication"}], "pattern": "vary-conditions", "fixes": ["population setup"], "varies": ["plasticity schedule"], "benchmarks": [], "results": [], "workflow": [], "summary": "Developmental trajectories improve explore-exploit dynamics in emergent communication."}
{"id": 3232, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Collaborative long-horizon vision-language navigation benchmark with multiple agents"}], "pattern": "fix-env-vary-agent", "fixes": ["agents"], "varies": ["collaborative tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "CoNavBench evaluates collaborative multi-agent VLN."}
{"id": 14116, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Test-time strategies for efficient agentic RAG"}], "pattern": "vary-scaffold", "fixes": ["Search-R1"], "varies": ["test-time strategies"], "benchmarks": [], "results": [], "workflow": [], "summary": "Test-time strategies reduce redundant retrieval and improve accuracy in agentic RAG."}
{"id": 5941, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "analysis", "why": "Energy-based analysis of negative-label OOD detection in VLMs"}], "pattern": "vary-conditions", "fixes": ["VLM"], "varies": ["negative labels"], "benchmarks": [], "results": [], "workflow": [], "summary": "Energy-based framework analyzing negative label guided OOD detection."}
{"id": 636, "in_agent_scope": false, "types": [{"area": "train/data", "role": "method", "why": "Semantic voting pseudo-labels for self-improvement on unverifiable tasks"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["pseudo-label"], "benchmarks": [], "results": [], "workflow": [], "summary": "Semantic voting gives self-evaluation-free self-improvement for open-ended tasks."}
{"id": 10837, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Joint instruction and weight test-time adaptation for multi-turn interactions"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["words and weights"], "benchmarks": [], "results": [], "workflow": [], "summary": "ROSA2 co-adapts prompts and weights for multi-turn interaction."}
{"id": 10568, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Energy-guided test-time sampling from RL-optimal policy without training"}], "pattern": "vary-scaffold", "fixes": ["model"], "varies": ["decoding"], "benchmarks": [], "results": [], "workflow": [], "summary": "ETS training-free sampling approximates RL-aligned policy."}
{"id": 12712, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "train-agent", "why": "Bernoulli-gated RL for prior-guided reasoning under conflicting retrieval"}], "pattern": "vary-scaffold", "fixes": ["RAG"], "varies": ["reasoning strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "BrPr elicits parametric priors to improve RAG robustness under conflicts."}