Files
ai-taxonomy-literature-anal…/data/labels/o41.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
29 KiB
JSON

{"id": 16275, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Unifies query diversification with early retriever and answerability verification for ambiguous queries."}], "pattern": "vary-scaffold", "fixes": ["backbone LLMs", "ASQA benchmark"], "varies": ["disambiguation pipeline"], "benchmarks": [{"name": "ASQA", "purpose": "grounding-aware F1 evaluation", "is_new": false}], "results": [], "workflow": ["generate interpretations", "retriever feedback", "answerability feedback"], "summary": "VerDICT merges interpretation diversification and verification for ambiguous query RAG, reducing cascading errors."}
{"id": 18463, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "method", "why": "Inference-time scaling framework for solving linguistic olympiad puzzles without fine-tuning."}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["inference-time scaling method"], "benchmarks": [{"name": "International Linguistics Olympiad", "purpose": "linguistic puzzle evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "Domain-specific inference-time scaling framework improves LLMs on linguistics olympiad puzzles without fine-tuning."}
{"id": 6457, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "RL-trained agentic MLLM for video misinformation detection with selective external verification."}], "pattern": "fix-env-vary-agent", "fixes": ["video misinformation task"], "varies": ["agent training with RL"], "benchmarks": [], "results": [], "workflow": ["assess task ambiguity", "call external tools", "iterative reasoning"], "summary": "FactGuard is an RL-trained agentic MLLM framework that iteratively reasons and uses external verification for video misinformation detection."}
{"id": 11889, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Residual Semantic Steering makes VLA models robust to linguistic perturbations."}], "pattern": "vary-scaffold", "fixes": ["VLA backbone"], "varies": ["language guidance mechanism"], "benchmarks": [], "results": [], "workflow": [], "summary": "Residual Semantic Steering disentangles affordance from semantics to stabilize VLA models against instruction rephrasing."}
{"id": 12069, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "Benchmark testing whether agents actively apply long-term memory to select tools and parameters."}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks requiring memory use"], "benchmarks": [{"name": "Mem2ActBench", "purpose": "memory-driven tool action evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Mem2ActBench evaluates proactive use of long-term memory for tool-based actions by LLM agents."}
{"id": 3128, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "application", "why": "LLMs as human proxy to generate chunk importance weights for video streaming QoE."}], "pattern": "n/a", "fixes": ["streaming system"], "varies": ["weight generation method"], "benchmarks": [], "results": [], "workflow": ["perceive frames locally", "rate chunks", "aggregate weights"], "summary": "HiVid uses LLMs to produce content-aware chunk weights for VOD and live streaming."}
{"id": 2802, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "method", "why": "Hierarchical predictive correction to reduce cascading failures in VLA long-horizon tasks."}, {"area": "planning-decision", "role": "train-agent", "why": "Reflective contrastive alignment training for subgoal correction."}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["planning and correction mechanism"], "benchmarks": [], "results": [], "workflow": [], "summary": "ReCAPA adds hierarchical predictive correction so VLA agents recover from mis-specified subgoals and avoid error cascades."}
{"id": 12017, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Hierarchical RL conditioning on single-step transitions instead of full histories for LLM agents."}], "pattern": "fix-env-vary-agent", "fixes": ["interactive environments"], "varies": ["RL framework, history representation"], "benchmarks": [], "results": [], "workflow": ["decompose into subtasks", "step-level transitions", "hierarchical RL"], "summary": "STEP-HRL trains LLM agents hierarchically using step-level transitions, avoiding long interaction histories."}
{"id": 1927, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Equips LLM agents with classic exploration algorithms for data-efficient RL."}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["exploration algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows existing LLM agent designs explore poorly and adapts classic RL exploration algorithms to LLM agents."}
{"id": 8659, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Agentic framework optimized for small language models with prompt compression and tool calling."}], "pattern": "vary-scaffold", "fixes": ["small LMs"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "EffGen is an open-source agentic framework letting small language models act as capable local agents."}
{"id": 5208, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Adversarial testing of computer-use agents via indirect prompt injection in hybrid web-OS sandbox."}, {"area": "agent-safety", "role": "benchmark", "why": "Hybrid VM and Docker sandbox with adversarial scenarios."}], "pattern": "fix-env-vary-agent", "fixes": ["CUAs under test"], "varies": ["adversarial injections"], "benchmarks": [{"name": "RedTeamCUA", "purpose": "hybrid web-OS adversarial evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "RedTeamCUA is a hybrid sandbox framework for testing computer-use agents against indirect prompt injection."}
{"id": 11905, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Guided GRPO with adaptive ground-truth reasoning guidance to help small models."}], "pattern": "vary-both", "fixes": ["RLVR setup"], "varies": ["guidance configuration"], "benchmarks": [], "results": [], "workflow": [], "summary": "G2RPO-A adaptively injects ground-truth reasoning steps into rollouts to improve RLVR for small language models."}
{"id": 2340, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "DAG-based parallel execution framework for web agents."}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["execution paradigm"], "benchmarks": [], "results": [], "workflow": ["decompose into DAG", "parallel execution", "aggregate results"], "summary": "Flash-Searcher decomposes tasks into a dependency DAG for concurrent tool execution in web agents."}
{"id": 14035, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Interactive agentic system for scRNA-seq analysis with on-device engine and experiment kernel."}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "Gardener is an agentic scRNA-seq analysis system keeping data local and externalizing analysis progress."}
{"id": 12255, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Training-free sparse attention for long-context diffusion LLM inference."}], "pattern": "vary-scaffold", "fixes": ["dLLM weights"], "varies": ["attention sparsification"], "benchmarks": [], "results": [], "workflow": [], "summary": "Focus-dLLM accelerates long-context diffusion LLM inference via confidence-guided training-free sparse attention."}
{"id": 16477, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "data", "why": "Multi-agent teacher trajectories distilled into single-agent recommender."}], "pattern": "vary-data", "fixes": ["recommendation task"], "varies": ["distilled trajectories"], "benchmarks": [], "results": [], "workflow": ["multi-agent teacher", "trajectory distillation", "train single model"], "summary": "STAR internalizes multi-agent reasoning trajectories into one efficient LLM recommender."}
{"id": 13446, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Execution-based verification and self-correction for KBQA semantic parsing."}], "pattern": "vary-scaffold", "fixes": ["KBQA task"], "varies": ["verification method"], "benchmarks": [], "results": [], "workflow": [], "summary": "EVER reframes KBQA generation as execution-verified fine-grained self-correction to fix query errors."}
{"id": 18538, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Controlled study of how peer sycophancy rankings affect multi-agent discussion outcomes."}], "pattern": "vary-conditions", "fixes": ["open-source LLMs"], "varies": ["peer sycophancy information"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies how awareness of peers' sycophancy levels propagates sycophancy in LLM multi-agent discussions."}
{"id": 15535, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Salient-entity-guided decoupling for document-level relation extraction."}], "pattern": "vary-scaffold", "fixes": ["DocRE task"], "varies": ["extraction method"], "benchmarks": [], "results": [], "workflow": [], "summary": "SegDRE uses salient entities to split dense and sparse pairs for document-level relation extraction."}
{"id": 9140, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Self-evolving value-driven memory agent for NPU kernel synthesis without fine-tuning."}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Value-driven memory for drafting and continual refining."}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["memory-based framework"], "benchmarks": [], "results": [], "workflow": ["draft kernel", "store value-scored memory", "refine continually"], "summary": "EvoKernel is a self-evolving agentic framework using value-driven memory to synthesize NPU kernels in data-scarce settings."}
{"id": 14454, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent system generating 3D games in Unreal Engine with RAG-grounded tools."}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["multi-agent pipeline"], "benchmarks": [], "results": [], "workflow": ["retrieve models", "generate scene", "synthesize code", "test game"], "summary": "AutoUE coordinates agents to build Unreal Engine 3D games end-to-end with retrieval-grounded tool use."}
{"id": 15426, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Zero-shot fake news detection with entity-guided retrieval and multi-LLM interaction."}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["retrieval and interaction pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "ZoFia detects fake news zero-shot via entity-guided retrieval and multi-LLM debate."}
{"id": 9016, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Stratified GRPO normalizes advantages per tool-call structure for search agent RL."}], "pattern": "fix-env-vary-agent", "fixes": ["search tasks"], "varies": ["RL advantage estimation"], "benchmarks": [], "results": [], "workflow": [], "summary": "Stratified GRPO removes cross-stratum bias in credit assignment when training RL search agents."}
{"id": 15873, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Diversity-encouraging RL for general reasoning without domain verifiers."}], "pattern": "vary-both", "fixes": ["RLVR setup"], "varies": ["reward and diversity objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "DARL encourages diverse answers during verifier-free RL, reducing overfitting to reference answers."}
{"id": 14039, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent system for streaming media monitoring, analysis and reporting."}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["multi-agent pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "A multi-agent media monitoring system with matching, batched extraction and report agents reducing LLM calls."}
{"id": 15243, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "SFT with balanced training signals and hard-example weighting for function calling."}], "pattern": "fix-env-vary-agent", "fixes": ["function calling benchmarks"], "varies": ["SFT objective and data hardness"], "benchmarks": [], "results": [], "workflow": [], "summary": "BalanceSFT rebalances CoT versus function-call tokens and data hardness to improve function-calling SFT."}
{"id": 14356, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Systematic study of diversity collapse in multi-agent idea generation across model, agent and system levels."}], "pattern": "vary-conditions", "fixes": ["ideation task"], "varies": ["model strength, agent cognition, system dynamics"], "benchmarks": [], "results": [], "workflow": [], "summary": "Empirical study showing multi-agent systems often fail to expand idea diversity, with structural coupling causing collective collapse."}
{"id": 6539, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Uncertainty-driven policy optimization to prevent debate collapse in multi-agent debate."}, {"area": "multi-agent", "role": "eval-method", "why": "Hierarchical uncertainty metric at agent, inter-agent and system levels."}], "pattern": "fix-env-vary-agent", "fixes": ["debate benchmarks"], "varies": ["training objective, uncertainty metric"], "benchmarks": [], "results": [], "workflow": [], "summary": "Proposes hierarchical uncertainty metrics for debate collapse and an uncertainty-driven policy optimization to mitigate it."}
{"id": 1484, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Benchmark of 12 video games via MCP interface for training and evaluating LLM agents."}], "pattern": "fix-env-vary-agent", "fixes": ["LLM agents under test"], "varies": ["games, agentic modules"], "benchmarks": [{"name": "Orak", "purpose": "LLM agent evaluation across 12 games", "is_new": true}], "results": [], "workflow": [], "summary": "Orak provides 12-game benchmark with MCP interface, agentic module studies and fine-tuning datasets for gaming agents."}
{"id": 5965, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Gradient-free test-time guidance for flow models via parallel tempering on source distributions."}], "pattern": "vary-scaffold", "fixes": ["pretrained flow model"], "varies": ["guidance method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Source Parallel Tempering guides pretrained flow and diffusion models toward high-reward regions without gradients."}
{"id": 13487, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Unified policy trained for GUI and embodied navigation."}, {"area": "embodied-nav", "role": "data", "why": "Visual-target trajectory collection pipeline across both domains."}], "pattern": "fix-env-vary-agent", "fixes": ["GUI and navigation tasks"], "varies": ["unified training policy and data"], "benchmarks": [], "results": [], "workflow": [], "summary": "NaviMaster unifies GUI and embodied navigation as MDPs in one agent trained on collected trajectories."}
{"id": 735, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Query-guided spatial-temporal-frequency interaction for music audio-visual QA."}], "pattern": "vary-scaffold", "fixes": ["AVQA task"], "varies": ["model architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "A query-guided spatial-temporal-frequency model for music audio-visual question answering."}
{"id": 9225, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Self-play memory-augmented context optimization for stable multi-turn multi-agent LLM games."}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["inference-time context"], "benchmarks": [], "results": [], "workflow": [], "summary": "MEMO optimizes inference-time context with memory via self-play to reduce variance and improve multi-agent game play."}
{"id": 851, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Agent architecture using soft propositional reasoning for robust LLM analysis."}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analytica estimates soft truth values of outcome propositions to give stable, compositional LLM-driven analysis."}
{"id": 14141, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "data", "why": "Validator-guided hard negative mining for MLM fine-tuning on Sumerian."}], "pattern": "vary-data", "fixes": ["mBERT"], "varies": ["negative selection"], "benchmarks": [], "results": [], "workflow": [], "summary": "A linguistic validator filters candidates and mines hard negatives to improve masked language modeling for Sumerian."}
{"id": 17239, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Localization-guided instructions for LLM repair of ReDoS regexes."}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["repair guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "Combines symbolic localization with LLM instructions to repair regex denial-of-service vulnerabilities."}
{"id": 8553, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Entropy dynamics framework analyzing centralized orchestration in multi-agent systems."}, {"area": "multi-agent", "role": "benchmark", "why": "Inverse Workflow Generation synthesizes verifiable high-complexity benchmarks."}], "pattern": "vary-conditions", "fixes": ["MAS"], "varies": ["orchestration topology"], "benchmarks": [], "results": [], "workflow": [], "summary": "Models orchestration as mean-field entropy dynamics and introduces Inverse Workflow Generation to build benchmarks."}
{"id": 15291, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Benchmark for VLMs detecting and reasoning about high-level failures in embodied tasks."}], "pattern": "fix-env-vary-agent", "fixes": ["VLMs under test"], "varies": ["failure scenarios"], "benchmarks": [{"name": "FAER", "purpose": "failure-aware embodied reasoning evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "FAER benchmarks VLMs on detecting and reasoning about high-level failures in long-horizon embodied tasks."}
{"id": 10334, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL with synthetic semantic information gain reward for retrieval-augmented agentic reasoning."}], "pattern": "fix-env-vary-agent", "fixes": ["retrieval setup"], "varies": ["reward signal"], "benchmarks": [], "results": [], "workflow": [], "summary": "InfoReasoner rewards information seeking via a synthetic semantic information gain reward grounded in uncertainty reduction."}
{"id": 4511, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "AST-guided task decomposition training for small LLMs on NL2SQL."}], "pattern": "fix-env-vary-agent", "fixes": ["NL2SQL task"], "varies": ["training with decomposition"], "benchmarks": [], "results": [], "workflow": [], "summary": "LearNAT trains small public LLMs for NL2SQL using AST-guided task decomposition."}
{"id": 10521, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Studies answer revision dynamics in multi-agent debate using distance measures on ConceptARC."}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["peer answer configurations"], "benchmarks": [{"name": "ConceptARC", "purpose": "distance-based revision analysis", "is_new": false}], "results": [], "workflow": [], "summary": "Analyzes how LLMs revise answers under peer influence in debate, depending on answer distance."}
{"id": 7690, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Benchmark of language agents under controllable, extreme context growth."}], "pattern": "fix-env-vary-agent", "fixes": ["LLM agents under test"], "varies": ["context length"], "benchmarks": [{"name": "LOCA-bench", "purpose": "agent reliability under context growth", "is_new": true}], "results": [], "workflow": [], "summary": "LOCA-bench evaluates language agents as context grows in long-running tasks, measuring context rot."}
{"id": 13965, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Neuro-symbolic agentic RL with POMDP formulation for long-term character companions."}], "pattern": "fix-env-vary-agent", "fixes": ["companion interaction setting"], "varies": ["RL training"], "benchmarks": [], "results": [], "workflow": [], "summary": "NSARL trains original-character companion agents with trajectory-level value learning for consistent persona and memory."}
{"id": 12561, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "analysis", "why": "Empirical study of SWE agents for Linux kernel fault localization."}], "pattern": "vary-conditions", "fixes": ["SWE agents"], "varies": ["kernel fault localization setting"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies how software engineering agents perform at fault localization in the Linux kernel and why they fail."}
{"id": 13298, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "MDL-guided rule learning from failure traces injected into prompts."}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["injected rules"], "benchmarks": [], "results": [], "workflow": ["collect failure traces", "propose rules", "MDL consolidation", "inject in prompt"], "summary": "RIMRULE distills compact rules from failure traces using MDL and injects them to improve tool use."}
{"id": 12090, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Retrieval-augmented multi-agent router and detectors with prompt evolution for vulnerability detection."}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-agent framework, prompts"], "benchmarks": [], "results": [], "workflow": [], "summary": "MulVul routes code to specialized detector agents with retrieval and cross-model prompt evolution."}
{"id": 1040, "in_agent_scope": true, "types": [{"area": "search-research", "role": "analysis", "why": "Analysis of efficiency bottlenecks in search agents."}, {"area": "search-research", "role": "scaffold", "why": "Efficiency enhancements for retrieval and system design."}], "pattern": "vary-conditions", "fixes": ["search agents"], "varies": ["retrieval method, system design"], "benchmarks": [], "results": [], "workflow": [], "summary": "Identifies retrieval and system inefficiencies in LLM search agents and proposes efficiency improvements."}
{"id": 4746, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Knowledge-induced training of web agents using a Factual, Conceptual, Procedural framework."}, {"area": "web-gui", "role": "data", "why": "Knowledge-content datasets for web agent learning."}], "pattern": "fix-env-vary-agent", "fixes": ["web tasks"], "varies": ["knowledge training data"], "benchmarks": [], "results": [], "workflow": [], "summary": "Web-CogReasoner trains web agents first on structured web knowledge then cognitive reasoning."}
{"id": 14788, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "eval-method", "why": "Diagnostic framework separating text and speech routes of persona emotion consistency."}], "pattern": "vary-evaluation", "fixes": ["spoken agents"], "varies": ["evaluation route decomposition"], "benchmarks": [], "results": [], "workflow": [], "summary": "PED diagnoses persona breaks in spoken role-play agents by comparing text and speech affective routes."}
{"id": 15121, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Self-improving human-agent framework for thematic analysis."}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["human-agent workflow"], "benchmarks": [], "results": [], "workflow": [], "summary": "CentaurTA is an iterative human-agent framework improving fine-grained qualitative coding for thematic analysis."}
{"id": 12267, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Codified decision trees as executable behavioral profiles for role-playing agents."}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["character profile representation"], "benchmarks": [], "results": [], "workflow": [], "summary": "CDT induces executable decision trees from storylines to give role-playing agents consistent behavior."}
{"id": 14515, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Multi-agent framework automating modeling and solving of optimization under uncertainty."}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["multi-agent pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "LLM4OuU automates uncertainty model building and solving with specialized agents across industries."}
{"id": 11809, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free difference-guided CLIP embedding editing for few-shot anomaly detection."}], "pattern": "vary-scaffold", "fixes": ["CLIP"], "varies": ["embedding editing"], "benchmarks": [], "results": [], "workflow": [], "summary": "DE-CLIP edits CLIP embeddings using differences to separate normal and anomalous samples without training."}
{"id": 11110, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Examines whether trajectory-SFT agents learn semantics or interface patterns."}, {"area": "tool-use", "role": "eval-method", "why": "PIPE rewrites environment interfaces to diagnose interface reliance."}], "pattern": "fix-agent-vary-env", "fixes": ["task semantics"], "varies": ["environment interface"], "benchmarks": [], "results": [], "workflow": [], "summary": "PIPE minimally rewrites interfaces to show whether trajectory-SFT agents rely on interface memorization."}
{"id": 6902, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Multi-agent episodic context reconstruction instead of memory preprocessing."}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory mechanism"], "benchmarks": [], "results": [], "workflow": [], "summary": "E-mem reconstructs episodic context via multiple agents to preserve narrative integrity in agent memory."}
{"id": 15587, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Schema-based card reranking as constraint satisfaction for financial QA."}], "pattern": "vary-scaffold", "fixes": ["financial QA"], "varies": ["reranking method"], "benchmarks": [], "results": [], "workflow": [], "summary": "FinCARDS reranks financial filing chunks using structured schema fields and constraint satisfaction."}
{"id": 17471, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "data", "why": "Multi-agent pipeline synthesizing financial QA data to fine-tune small models."}], "pattern": "vary-data", "fixes": ["SLM"], "varies": ["synthetic data"], "benchmarks": [], "results": [], "workflow": [], "summary": "A modular multi-agent pipeline synthesizes numerical financial QA data for fine-tuning small language models."}
{"id": 6514, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "attack", "why": "Key collision attack on embedding-based LLM semantic caching."}], "pattern": "vary-conditions", "fixes": ["semantic cache"], "varies": ["adversarial inputs"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows semantic cache keys conflict with collision resistance and demonstrates a key collision attack."}
{"id": 6879, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Large-scale stock market simulation with hierarchical generative agents."}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["simulation framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "MarketSim simulates stock markets using hierarchical generative agents balancing structural and behavioral fidelity."}
{"id": 2777, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "theory", "why": "Minimax regret analysis of aggregating expert decisions under quantal response."}], "pattern": "n/a", "fixes": [], "varies": ["behavior model assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Proves majority voting is the optimal robust aggregator for experts modeled by quantal response."}