Files
ai-taxonomy-literature-anal…/data/labels/o30.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
32 KiB
JSON

{"id": 14002, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "environment", "why": "Unified financial toolchain and brokerage simulation environment for LLM trading agents"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["trading environment/tools"], "benchmarks": [], "results": [], "workflow": ["Standardize data tools", "Structured news retrieval", "Stateful brokerage simulation"], "summary": "Unified trading environment standardizing market data, news retrieval and brokerage simulation for LLM financial agents."}
{"id": 13723, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Multi-turn agentic RL with dual memory for molecular optimization"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Static exemplar and evolving skill memory for sample efficiency"}], "pattern": "fix-env-vary-agent", "fixes": ["oracle budget", "molecular tasks"], "varies": ["memory system", "RL training"], "benchmarks": [], "results": [], "workflow": ["Retrieve exemplars", "Evolve skill memory", "Multi-turn RL optimization"], "summary": "Memory-augmented multi-turn agentic RL for sample-efficient molecular optimization with static exemplar and evolving skill memories."}
{"id": 17341, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "application", "why": "Interactive demo and toolkit for semantic translation memory retrieval"}], "pattern": "n/a", "fixes": ["TM data"], "varies": ["retrieval backends/encoders"], "benchmarks": [], "results": [], "workflow": ["Encode segments", "Vector database search", "Inspect matches in UI"], "summary": "Open-source demo and evaluation toolkit comparing semantic encoders and fuzzy baselines for translation memory retrieval."}
{"id": 5733, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Multi-agent framework exploiting information locality for IP-level Verilog generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["multi-agent RTL generation pipeline"], "benchmarks": [], "results": [], "workflow": ["Localize spec information", "Generate RTL modules", "Debug via verification"], "summary": "Multi-agent framework for IP-level Verilog generation handling long specs, long code and verification debugging."}
{"id": 13034, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "VeriGUI trains GUI agent with verification and recovery via SFT and RL"}, {"area": "web-gui", "role": "data", "why": "Synthetic failure trajectories for robust recovery learning"}], "pattern": "fix-env-vary-agent", "fixes": ["noisy GUI environments"], "varies": ["verification framework", "training pipeline"], "benchmarks": [], "results": [], "workflow": ["TVAE reasoning", "Robust SFT with synthetic failures", "RL refinement"], "summary": "VeriGUI uses a thinking-verification-action-expectation framework and robust SFT with synthetic failures to make GUI agents verify action effects."}
{"id": 15451, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Graph memory with spreading activation retrieval for LLM agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory architecture"], "benchmarks": [{"name": "LoCoMo", "purpose": "evaluate long-term memory", "is_new": false}], "results": [], "workflow": ["Build memory graph", "Spreading activation", "Triple hybrid retrieval"], "summary": "Synapse is a graph-based agent memory using spreading activation, lateral inhibition and hybrid retrieval, evaluated on LoCoMo."}
{"id": 12209, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Q-DAPS estimates question difficulty via entropy of answer plausibility"}], "pattern": "vary-evaluation", "fixes": ["QA datasets"], "varies": ["difficulty estimation method"], "benchmarks": [{"name": "TriviaQA", "purpose": "evaluate difficulty estimation", "is_new": false}, {"name": "NQ", "purpose": "evaluate difficulty estimation", "is_new": false}, {"name": "MuSiQue", "purpose": "evaluate difficulty estimation", "is_new": false}, {"name": "QASC", "purpose": "evaluate difficulty estimation", "is_new": false}], "results": [], "workflow": [], "summary": "Q-DAPS estimates LLM question difficulty from entropy of plausibility scores over candidate answers, outperforming baselines on four QA datasets."}
{"id": 10173, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "application", "why": "Serving system with invocation-distance memory management for large multi-agent simulation"}], "pattern": "n/a", "fixes": ["multi-agent workloads"], "varies": ["memory management/serving policy"], "benchmarks": [], "results": [], "workflow": ["Estimate invocation distance", "Prefetch agent state", "Priority-based eviction"], "summary": "ScaleSim is an LLM serving system that uses invocation distance for prefetching and eviction in large-scale multi-agent simulations."}
{"id": 592, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "MisinfoTask dataset of tasks for MAS misinformation robustness"}, {"area": "agent-safety", "role": "defense", "why": "ARGUS training-free goal-aware misinformation rectification"}], "pattern": "vary-both", "fixes": ["MAS attacks"], "varies": ["defense mechanism", "task set"], "benchmarks": [{"name": "MisinfoTask", "purpose": "evaluate MAS misinformation robustness", "is_new": true}], "results": [], "workflow": ["Identify misinformation", "Goal-aware reasoning", "Rectify information flow"], "summary": "Introduces MisinfoTask dataset and ARGUS, a training-free goal-aware defense that rectifies misinformation in multi-agent systems."}
{"id": 3990, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "train-agent", "why": "Language-conditioned humanoid control via diffusion policy on motion latents"}], "pattern": "n/a", "fixes": ["humanoid robot"], "varies": ["policy conditioning on motion latents"], "benchmarks": [], "results": [], "workflow": ["Language to motion latent", "Diffusion policy denoising", "Execute actions"], "summary": "RoboGhost conditions a diffusion humanoid policy directly on language-grounded motion latents, avoiding motion decoding and retargeting."}
{"id": 1863, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Semantic-guided hierarchical codebook for unified image tokenizer"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["tokenizer design"], "benchmarks": [], "results": [], "workflow": ["Pretrain semantic codebook", "Add pixel sub-codebooks"], "summary": "SemHiTok builds pixel sub-codebooks on a pretrained semantic codebook for a unified image tokenizer serving understanding and generation."}
{"id": 8835, "in_agent_scope": false, "types": [{"area": "reason/math", "role": "method", "why": "Neuro-symbolic WZ-guided Lean 4 proof sketches with LLM prover"}, {"area": "reason/math", "role": "train-agent", "why": "WZ-Prover via verified bootstrapping and DAPO refinement"}], "pattern": "vary-both", "fixes": ["Lean 4 verifier"], "varies": ["proof planning", "prover training"], "benchmarks": [{"name": "LCI-Test", "purpose": "combinatorial identity proving", "is_new": true}], "results": [{"benchmark": "LCI-Test", "system": "WZ-LLM", "base_model": null, "setting": "proposed", "metric": "proof success rate", "score": 34, "unit": "%"}], "workflow": ["WZ proof plan", "Lean sketch", "Prover discharges subgoals"], "summary": "WZ-LLM turns Wilf-Zeilberger plans into Lean 4 sketches and trains a prover, reaching 34% on LCI-Test."}
{"id": 17170, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "MIRAGE benchmark for agent robustness under imperfect guidance"}, {"area": "memory-selfevolve", "role": "analysis", "why": "Measures agent failures under incomplete or outdated guidance"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["guidance quality"], "benchmarks": [{"name": "MIRAGE", "purpose": "agent robustness to imperfect guidance", "is_new": true}], "results": [], "workflow": [], "summary": "MIRAGE benchmark measures LLM agent robustness when manuals, demonstrations or retrieved knowledge are imperfect."}
{"id": 17712, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "benchmark", "why": "HumaniCA datasets for detecting users' ascription of humanness to agents"}], "pattern": "n/a", "fixes": ["detection models"], "varies": ["datasets"], "benchmarks": [{"name": "HumaniCA", "purpose": "detect humanness ascription", "is_new": true}], "results": [], "workflow": [], "summary": "HumaniCA provides three annotated datasets of user turns and linguistic indicators of humanness ascription to conversational agents."}
{"id": 5629, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "TIPEX controllable framework for replica and structural inference parallelism"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["parallelism strategy"], "benchmarks": [], "results": [], "workflow": ["Replica parallelism", "Task decomposition", "Coordinate execution"], "summary": "Models inference-time parallelism in multi-agent systems at replica and structural levels and proposes TIPEX execution framework."}
{"id": 887, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Graph-augmented LLM with RL-guided subgraph reasoning for precision medicine"}], "pattern": "n/a", "fixes": ["multi-omic data"], "varies": ["model and reward design"], "benchmarks": [], "results": [], "workflow": ["Encode text-numeric graph", "RL-guided subgraph reasoning"], "summary": "GALAX integrates multi-omics, graph topology and text with RL-guided subgraph reasoning for explainable precision medicine."}
{"id": 2561, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "SLDAgent evolution-based agent co-optimizing scaling law form and parameters"}, {"area": "science-data", "role": "benchmark", "why": "Seven scaling law discovery tasks from 5,000 experiments"}], "pattern": "vary-both", "fixes": ["LLM"], "varies": ["agent design", "tasks"], "benchmarks": [{"name": "Scaling law discovery tasks", "purpose": "evaluate law discovery agents", "is_new": true}], "results": [], "workflow": ["Propose law formulas", "Evolve model and parameters"], "summary": "Curates scaling-law discovery tasks and introduces SLDAgent, an evolution-based agent finding laws that extrapolate better than human ones."}
{"id": 13945, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "MARCH multi-agent hierarchy for CT report generation"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["multi-agent pipeline"], "benchmarks": [{"name": "RadGenome-ChestCT", "purpose": "evaluate CT report generation", "is_new": false}], "results": [], "workflow": ["Resident drafts report", "Fellows revise with retrieval", "Attending consensus"], "summary": "MARCH emulates radiology hierarchy with resident, fellow and attending agents for iterative consensus CT report generation."}
{"id": 10635, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "HieraMAS optimizes intra-node LLM mixtures and inter-node topology"}, {"area": "multi-agent", "role": "train-agent", "why": "Two-stage optimization algorithm addressing credit assignment"}], "pattern": "vary-scaffold", "fixes": ["tasks"], "varies": ["LLM mixtures", "communication topology"], "benchmarks": [], "results": [], "workflow": ["Form supernodes", "Propose-synthesis", "Optimize topology"], "summary": "HieraMAS builds supernodes of heterogeneous LLMs and optimizes topology with a two-stage credit-assignment algorithm."}
{"id": 1293, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "DAComp benchmark of data engineering and analysis tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["data agents"], "varies": ["tasks"], "benchmarks": [{"name": "DAComp", "purpose": "data engineering and analysis agents", "is_new": true}], "results": [], "workflow": [], "summary": "DAComp is a 236-task benchmark for data agents spanning SQL pipeline engineering and open-ended analysis with LLM-judge scoring."}
{"id": 10894, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Decentralized disentangled task-role representation for offline multi-agent meta-RL"}], "pattern": "n/a", "fixes": ["MARL benchmarks"], "varies": ["representation learning method"], "benchmarks": [], "results": [], "workflow": [], "summary": "D2TR learns decentralized task and role representations via distillation for offline multi-agent meta RL generalization."}
{"id": 734, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "FinSearchComp financial search and reasoning agent benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["search agents"], "varies": ["tasks"], "benchmarks": [{"name": "FinSearchComp", "purpose": "financial search and reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "FinSearchComp is an open agent benchmark with three tasks for realistic financial search and reasoning."}
{"id": 346, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "AGILE interactive jigsaw RL to improve VLM perception"}, {"area": "embodied-nav", "role": "environment", "why": "Jigsaw interactive environment with code actions and visual feedback"}], "pattern": "fix-env-vary-agent", "fixes": ["jigsaw environment"], "varies": ["VLM training"], "benchmarks": [], "results": [], "workflow": ["Generate code action", "Environment visual feedback", "Train VLM"], "summary": "AGILE trains VLMs on interactive jigsaw solving with code actions and visual feedback to improve perception and reasoning."}
{"id": 6246, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Executable memory knowledge graph with value-guided MCTS planning"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Knowledge-graph memory of compressed routines"}], "pattern": "vary-scaffold", "fixes": ["GUI environment"], "varies": ["memory and planning"], "benchmarks": [], "results": [], "workflow": ["Mine action groups", "Build KG memory", "Q-guided MCTS"], "summary": "EAM builds a knowledge graph memory of GUI routines and plans via value-guided MCTS retrieval and execution."}
{"id": 15435, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Knowledge-graph enhanced multi-agent cost-aware sequential diagnosis"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["KG and agents"], "benchmarks": [], "results": [], "workflow": ["Build diagnosis KG", "Perception agent", "Reasoning and decision agents"], "summary": "GraphDx uses medical diagnosis knowledge graphs and perception, reasoning and decision agents for cost-aware sequential diagnosis."}
{"id": 12308, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "Mr Dre suite for multi-turn report revision"}, {"area": "search-research", "role": "analysis", "why": "Shows deep research agents regress during revision"}], "pattern": "fix-agent-vary-env", "fixes": ["five DRAs"], "varies": ["revision turns"], "benchmarks": [{"name": "Mr Dre", "purpose": "multi-turn report revision evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Mr Dre evaluates deep research agents on multi-turn report revision, revealing regression on previously covered content."}
{"id": 16509, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "R1-style RL with format and action rewards for web agents"}], "pattern": "fix-env-vary-agent", "fixes": ["web tasks"], "varies": ["RL training"], "benchmarks": [], "results": [], "workflow": ["Structured reward", "R1-style RL", "Single-step reasoning"], "summary": "WorkForceAgent-R1 trains a web agent with rule-based RL to improve reasoning for enterprise web navigation."}
{"id": 16349, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "Tests whether LLM agents' trading-style switching matches financial theory"}], "pattern": "vary-conditions", "fixes": ["stock market simulation"], "varies": ["behavioral drivers"], "benchmarks": [], "results": [], "workflow": [], "summary": "Assesses behavioral consistency of LLM trading agents' strategy switching with loss aversion, herding and other drivers in simulation."}
{"id": 2572, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "TRAJECT-Bench trajectory-aware tool-use benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "TRAJECT-Bench", "purpose": "trajectory-level tool use", "is_new": true}], "results": [], "workflow": [], "summary": "TRAJECT-Bench evaluates tool selection, parameterization and ordering across trajectories varying in breadth and depth."}
{"id": 5262, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Toolathlon long-horizon multi-application benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "Toolathlon", "purpose": "diverse long-horizon tool tasks", "is_new": true}], "results": [], "workflow": [], "summary": "Toolathlon spans 32 applications and 604 tools for realistic long-horizon agent evaluation with execution-based checks."}
{"id": 15494, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "method", "why": "LLM translates language into ASP with self-correction for nonmonotonic reasoning"}], "pattern": "n/a", "fixes": ["no per-task engineering"], "varies": ["framework"], "benchmarks": [], "results": [], "workflow": ["Translate to ASP", "Run solver", "Self-correct"], "summary": "LLM+ASP translates natural language to Answer Set Programs with self-correction for task-agnostic nonmonotonic reasoning."}
{"id": 11549, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Test-time adaptation for robust cross-modal retrieval under query noise"}], "pattern": "n/a", "fixes": ["VLP model"], "varies": ["adaptation method"], "benchmarks": [{"name": "COCO-C", "purpose": "noisy retrieval evaluation", "is_new": false}, {"name": "Flickr-C", "purpose": "noisy retrieval evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "ReEx refines noisy queries with LLM repairs and adapts using exclusion-guided contrastive learning for robust retrieval."}
{"id": 18599, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "ID-MoCQA Indonesian multi-hop cultural QA dataset"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["questions"], "benchmarks": [{"name": "ID-MoCQA", "purpose": "cultural multi-hop reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "ID-MoCQA is a multi-hop Indonesian cultural QA dataset built by transforming single-hop questions into clue-based chains."}
{"id": 11061, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Studies tool-use collapse in visual CoT agents"}, {"area": "tool-use", "role": "train-agent", "why": "Encourages diverse tool use via entropy-regularized exploration"}], "pattern": "vary-conditions", "fixes": ["visual tasks"], "varies": ["tool-use incentives"], "benchmarks": [], "results": [], "workflow": [], "summary": "Identifies tool-use collapse in visual agents and argues tool diversity matters more than frequency."}
{"id": 12415, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "SynthAgent refines synthetic tasks and trajectories for new websites"}, {"area": "web-gui", "role": "train-agent", "why": "Fine-tunes web agents on refined synthetic data"}], "pattern": "vary-data", "fixes": ["web agent training"], "varies": ["synthetic data quality"], "benchmarks": [], "results": [], "workflow": ["Explore web elements", "Synthesize tasks", "Refine tasks and trajectories"], "summary": "SynthAgent synthesizes and dual-refines tasks and trajectories to adapt web agents to new websites."}
{"id": 15554, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Personalized mobile agent with preference learning and hierarchical memory"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Hierarchical preference memory storing user habits"}], "pattern": "vary-scaffold", "fixes": ["mobile tasks"], "varies": ["personalization modules"], "benchmarks": [], "results": [], "workflow": ["Preference learning", "Personal reward model", "Hierarchical memory"], "summary": "Me-Agent learns user habits at prompt and memory levels for personalized mobile agent interaction."}
{"id": 2956, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "TARE prompt evolution with sharpness-aware robustness"}], "pattern": "n/a", "fixes": ["black-box LLM"], "varies": ["prompt optimization"], "benchmarks": [], "results": [], "workflow": [], "summary": "TARE formalizes textual sharpness and evolves prompts robust to paraphrase in black-box settings."}
{"id": 7902, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "data", "why": "BARRED synthetic data generation with multi-agent debate verification"}, {"area": "safety/jailbreak", "role": "defense", "why": "Trained custom policy guardrail classifiers"}], "pattern": "vary-data", "fixes": ["small language models"], "varies": ["synthetic training data"], "benchmarks": [], "results": [], "workflow": ["Decompose domain dimensions", "Generate examples", "Debate to verify labels"], "summary": "BARRED generates faithful synthetic data via debate for finetuning small custom policy guardrails."}
{"id": 14506, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "MIThinker thinking model trained with SFT and RL for counseling"}, {"area": "domain-agents", "role": "data", "why": "AugR1-MI reverse-engineers counselor thoughts"}], "pattern": "fix-env-vary-agent", "fixes": ["counseling agents"], "varies": ["thinking model"], "benchmarks": [], "results": [], "workflow": ["Reverse-engineer thoughts", "SFT", "RL"], "summary": "MIThinker is a plug-in thinking model trained with SFT and RL on reverse-engineered thoughts for motivational interviewing."}
{"id": 14939, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Retrospective critic gives dense turn-level rewards for search agent RL"}, {"area": "planning-decision", "role": "train-agent", "why": "Fine-grained credit assignment"}], "pattern": "fix-env-vary-agent", "fixes": ["multi-hop QA benchmarks"], "varies": ["reward credit assignment"], "benchmarks": [], "results": [], "workflow": ["Critic evaluates turns", "Dense rewards", "Policy RL"], "summary": "CriticSearch uses a frozen critic with privileged trajectory information to give dense turn-level rewards for search agents."}
{"id": 13474, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "CostBomb computational cost attacks on web agents"}], "pattern": "fix-agent-vary-env", "fixes": ["WebAgents"], "varies": ["adversarial prompts"], "benchmarks": [], "results": [], "workflow": ["Generate adversarial prompts", "RL selector", "Inject into pages"], "summary": "CostBomb generates and selects adversarial prompts that inflate WebAgents' reasoning cost."}
{"id": 3105, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "GLC learns efficient interpretable multi-agent communication via language grounding"}], "pattern": "n/a", "fixes": ["MARL environments"], "varies": ["communication protocol"], "benchmarks": [], "results": [], "workflow": ["Autoencode symbols", "Align to concepts", "Contrastive training"], "summary": "GLC learns discrete communication symbols aligned to LLM-generated human concepts with contrastive consistency."}
{"id": 12280, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "LAS dynamically routes queries to workflows via gate and scheduler"}], "pattern": "vary-scaffold", "fixes": ["workflows"], "varies": ["workflow scheduling"], "benchmarks": [], "results": [], "workflow": ["Gate evaluates output", "Scheduler routes workflow"], "summary": "LLM-as-Scheduler cascades a gate and LLM scheduler to choose workflows per query, cutting tokens and latency."}
{"id": 16797, "in_agent_scope": false, "types": [{"area": "multiling/multi", "role": "analysis", "why": "Compares LLM perplexities with human Hinglish verb preferences"}], "pattern": "vary-conditions", "fixes": ["human judgments"], "varies": ["LLM families/sizes"], "benchmarks": [], "results": [], "workflow": [], "summary": "Finds LLMs do not reliably match native speaker preferences in Hindi-English verb code-mixing."}
{"id": 15777, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "VANE novelty-shaped exploration for RLVR"}], "pattern": "n/a", "fixes": ["Qwen2.5-Math-7B"], "varies": ["exploration reward shaping"], "benchmarks": [], "results": [], "workflow": ["Measure outcome novelty", "Measure process novelty", "Value-align shaping"], "summary": "VANE shapes RLVR rewards with outcome and process novelty plus value alignment to avoid mode collapse."}
{"id": 14282, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent deliberation at escalation boundaries in LLM cascades"}], "pattern": "vary-scaffold", "fixes": ["cascade models"], "varies": ["deliberation at tiers"], "benchmarks": [], "results": [], "workflow": ["Confidence routing", "Agent ensemble debate", "Escalate if unresolved"], "summary": "CascadeDebate inserts selective multi-agent deliberation at cascade tier boundaries to cut costly escalations."}
{"id": 3301, "in_agent_scope": true, "types": [{"area": "search-research", "role": "data", "why": "AgentFrontier ZPD-guided data synthesis for agent training"}, {"area": "search-research", "role": "benchmark", "why": "ZPD Exam self-evolving benchmark"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [{"name": "ZPD Exam", "purpose": "evaluate frontier agent reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "AgentFrontier synthesizes data at an LLM's zone of proximal development and produces the ZPD Exam benchmark."}
{"id": 8186, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "AdaMEM hybrid trajectory and strategy memory for test-time adaptation"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["Retrieve trajectories", "Generate strategy memory", "Guide decisions"], "summary": "AdaMEM adapts agents at test time with long-term trajectory memory and dynamic short-term strategy memory."}
{"id": 4088, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "CoAct-1 multi-agent system mixing GUI operator and coding programmer"}], "pattern": "vary-scaffold", "fixes": ["computer-use tasks"], "varies": ["coding actions"], "benchmarks": [], "results": [], "workflow": ["Orchestrator delegates", "GUI Operator", "Programmer writes scripts"], "summary": "CoAct-1 delegates subtasks to a GUI operator or programmer agent executing scripts for efficient computer use."}
{"id": 5892, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "eval-method", "why": "BSCT benchmark for PES smoothness of interatomic potentials"}], "pattern": "vary-evaluation", "fixes": ["MLIPs"], "varies": ["smoothness metric"], "benchmarks": [{"name": "BSCT", "purpose": "PES smoothness evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "BSCT efficiently probes potential energy surface smoothness, correlates with MD stability and guides MLIP design."}
{"id": 11724, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "SGP-CoT self-guided CoT pruning for efficient reasoning"}], "pattern": "n/a", "fixes": ["LRMs"], "varies": ["pruning method"], "benchmarks": [], "results": [], "workflow": ["Segment reasoning", "Assess necessity", "Prune"], "summary": "SGP-CoT uses the model's likelihood landscape to prune redundant chain-of-thought segments."}
{"id": 14015, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "ClinQueryAgent translates population health questions into queries with privacy"}, {"area": "tool-use", "role": "application", "why": "Sub-agent retrieval for database querying"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": [], "summary": "ClinQueryAgent is a deployed conversational agent turning health questions into database queries without patient data leaving the environment."}
{"id": 7813, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Executable exclusion contracts for KGQA with scoped negation"}], "pattern": "n/a", "fixes": ["KG"], "varies": ["negation handling"], "benchmarks": [], "results": [], "workflow": ["Specify exclusions", "Compile masks", "Execute multi-hop"], "summary": "EEC and MatLogic compile scope-bound exclusions into masks for multi-hop KG query answering."}
{"id": 14702, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "ShapleyFlow attributes gains of component upgrades in agentic workflows"}, {"area": "planning-decision", "role": "benchmark", "why": "CapaBench of 1,500+ tasks"}], "pattern": "vary-conditions", "fixes": ["workflow structure"], "varies": ["component LLM allocation"], "benchmarks": [{"name": "CapaBench", "purpose": "evaluate modular workflow components", "is_new": true}], "results": [], "workflow": [], "summary": "ShapleyFlow uses Shapley values to attribute workflow gains to component upgrades, with CapaBench introduced."}
{"id": 6536, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "EvolveR self-distills experience into principles and updates policy"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["experience-driven training"], "benchmarks": [], "results": [], "workflow": ["Distill principles", "Deduplicate and score", "Retrieve and reinforce"], "summary": "EvolveR distills past trajectories into strategic principles retrieved online and used to improve the agent."}
{"id": 9855, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "HistBench of 414 historical reasoning questions"}, {"area": "domain-agents", "role": "application", "why": "HistAgent history-specialized agent"}], "pattern": "vary-both", "fixes": ["agents"], "varies": ["tasks", "domain agent"], "benchmarks": [{"name": "HistBench", "purpose": "historical reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "HistBench evaluates historical reasoning and HistAgent is a domain agent for it."}
{"id": 14076, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Self-evolution via executable subagent code accumulation"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["experience representation"], "benchmarks": [], "results": [], "workflow": ["Solve task", "Save subagent code", "Refine and reuse"], "summary": "AgentFactory saves successful solutions as refined executable Python subagents reused across tasks."}
{"id": 117, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "benchmark", "why": "MedAraBench Arabic medical multiple-choice QA dataset"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["dataset"], "benchmarks": [{"name": "MedAraBench", "purpose": "Arabic medical QA", "is_new": true}], "results": [], "workflow": [], "summary": "MedAraBench is a large Arabic medical QA dataset digitized from academic materials, validated by experts and LLM judge."}
{"id": 13518, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "BlindGuard unsupervised malicious agent detection in MAS"}], "pattern": "fix-agent-vary-env", "fixes": ["unknown attacks"], "varies": ["defense method"], "benchmarks": [], "results": [], "workflow": ["Encode interactions", "Corruption-guided training", "Detect agents"], "summary": "BlindGuard unsupervised defense detects malicious agents using hierarchical encoder without attack labels."}
{"id": 350, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Isolates long-horizon execution failures by providing plan and knowledge"}], "pattern": "vary-conditions", "fixes": ["provided plan"], "varies": ["task length", "model size"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows small per-step gains compound and failures in long tasks are execution errors with degrading per-step accuracy."}