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
33 KiB
JSON
61 lines
33 KiB
JSON
{"id": 3467, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Cognitively inspired three-stage memory system reducing time and compute overhead for LLM memory"}], "pattern": "vary-scaffold", "fixes": ["LLM backbone"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["Sensory memory filters and groups", "Topic-aware short-term consolidation", "Long-term memory update"], "summary": "LightMem is a lightweight three-stage memory system for LLMs balancing performance and efficiency."}
|
|
{"id": 10829, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "eval-method", "why": "Empowerment-based information-theoretic metric EELMA to evaluate LM agents without hand-made benchmarks"}], "pattern": "vary-evaluation", "fixes": ["LM agents under test"], "varies": ["evaluation metric"], "benchmarks": [{"name": "textual games", "purpose": "environments for empowerment estimation", "is_new": false}, {"name": "web and tool-use environments", "purpose": "environments for empowerment estimation", "is_new": false}], "results": [], "workflow": ["Collect multi-turn interactions", "Estimate effective empowerment", "Correlate with task performance"], "summary": "EELMA estimates agent empowerment from text interactions, correlating with task performance as a scalable evaluation."}
|
|
{"id": 17653, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "RST discourse-tree RAG with query-complexity classifier for long-document query-based summarization"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["retrieval method"], "benchmarks": [{"name": "SQuALITY", "purpose": "evaluate query-based summarization", "is_new": false}], "results": [], "workflow": ["Build discourse tree", "Navigate by rhetorical relations", "Select retrieval strategy by complexity"], "summary": "DiscoRAG uses discourse trees and query complexity routing for RAG-based query summarization of long narratives."}
|
|
{"id": 16692, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Two-stage SFT plus execution-guided RL SQL generator with self-correction"}, {"area": "code-swe", "role": "scaffold", "why": "Vector-based schema retriever for efficient schema linking"}], "pattern": "vary-both", "fixes": ["benchmarks"], "varies": ["schema retriever", "SQL generator training"], "benchmarks": [], "results": [], "workflow": ["Retrieve schema via embeddings", "SFT SQL generator", "Execution-guided RL"], "summary": "LitE-SQL is a lightweight text-to-SQL framework with embedding schema linking and execution-guided RL self-correction."}
|
|
{"id": 16230, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "CoopValue scenarios assessing LLM value preferences in cooperative multi-agent settings"}], "pattern": "vary-conditions", "fixes": ["LLMs under test"], "varies": ["cooperation type", "value pair"], "benchmarks": [{"name": "CoopValue", "purpose": "LLM value preferences in cooperation", "is_new": true}], "results": [], "workflow": [], "summary": "CoopValue is a multi-agent benchmark of 1,778 scenarios probing LLM Schwartz-value preferences across cooperation types."}
|
|
{"id": 15543, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "scaffold", "why": "Error notebooks of security and function failures guide inference-time secure code generation"}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["inference guidance"], "benchmarks": [], "results": [], "workflow": ["Record security/function errors", "Retrieve notebook entries", "Contrastive guidance at inference"], "summary": "SafeNote uses security and function error notebooks to steer LLMs toward secure and correct code."}
|
|
{"id": 1653, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "theory", "why": "Theoretical analysis of autoregressive diffusion-loss image generation and condition error decay"}, {"area": "arch/diffusion", "role": "method", "why": "Optimal-transport condition refinement approach"}], "pattern": "n/a", "fixes": [], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Theoretical comparison of autoregressive diffusion and diffusion models plus an OT-based condition refinement approach."}
|
|
{"id": 6787, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Geometry-guided preprocessing converting transparent regions to opaque images for depth estimation"}], "pattern": "n/a", "fixes": ["downstream models"], "varies": ["preprocessing"], "benchmarks": [], "results": [], "workflow": ["Disentangle masks and transparency", "Recover surface normals", "Synthesize opaque RGB"], "summary": "GHOST preprocesses transparent objects into opaque structured images to improve depth and 3D reconstruction without retraining."}
|
|
{"id": 15410, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Autonomous LLM agent framework with tool-use environment for open-ended climate analysis"}], "pattern": "vary-scaffold", "fixes": ["LLM backbone"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "ClimAgent is a general-purpose autonomous agent framework executing diverse climate science research tasks."}
|
|
{"id": 13255, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "Multi-source benchmark for data analytics agents across relational, CSV and NoSQL data"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["data sources"], "benchmarks": [{"name": "UniDataBench", "purpose": "evaluate data analytics agents on diverse sources", "is_new": true}], "results": [], "workflow": [], "summary": "UniDataBench evaluates data analytics agents across structured and unstructured sources built from industry reports."}
|
|
{"id": 12660, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Retrieval-based recovery of missing modalities for audio-visual question answering"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["missing modality handling"], "benchmarks": [], "results": [], "workflow": ["Cross-modal retrieval via embeddings", "Purify retrieved semantics", "Answer AVQA"], "summary": "R2ScP replaces generative imputation with retrieval-based recovery for incomplete audio-visual QA."}
|
|
{"id": 17907, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Manual Swedish factual-knowledge QA benchmark with English translations"}], "pattern": "fix-agent-vary-env", "fixes": ["models under test"], "varies": ["benchmark"], "benchmarks": [{"name": "Sweden-related factual QA benchmark", "purpose": "measure Swedish factual recall", "is_new": true}], "results": [], "workflow": [], "summary": "A manually written benchmark on Sweden-related facts measuring factual recall and cross-lingual consistency."}
|
|
{"id": 15784, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Learns optimal message representations for agent communication via EMDP-based optimization"}], "pattern": "vary-scaffold", "fixes": ["tasks"], "varies": ["communication representation"], "benchmarks": [], "results": [], "workflow": [], "summary": "OPTiMACS dynamically learns message representations for LLM agent collaboration, modeled as an Expanding MDP."}
|
|
{"id": 2055, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "RL from algorithm feedback trains GNN to guide SAT solver branching with GRPO"}], "pattern": "n/a", "fixes": ["SAT solvers"], "varies": ["guidance policy"], "benchmarks": [], "results": [], "workflow": ["GNN assigns variable weights", "Inject into branching heuristics", "Train with solver cost reward"], "summary": "RLAF trains GNNs with GRPO using solver cost as reward to guide SAT branching heuristics."}
|
|
{"id": 17001, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Cooperative LLM agents retrieve from causal-temporal knowledge graph to predict post-earnings shocks"}], "pattern": "vary-scaffold", "fixes": ["LLM, no task training"], "varies": ["agent team"], "benchmarks": [], "results": [], "workflow": ["Build causal-temporal KG", "Specialized agents retrieve evidence", "Predict price shock"], "summary": "CARAG uses cooperative agents over a causal-temporal knowledge graph to forecast post-earnings price shocks."}
|
|
{"id": 17252, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "Adversarial multi-agent debate framework with advocates, judge, jury for LLM evaluation"}], "pattern": "vary-evaluation", "fixes": ["LLMs under evaluation"], "varies": ["judging protocol"], "benchmarks": [], "results": [], "workflow": ["Advocates defend answers", "Judge deliberates", "Budgeted stopping and decision"], "summary": "D3 is a cost-aware multi-agent debate judging framework with probabilistic analysis of reliability."}
|
|
{"id": 10189, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "data", "why": "On-policy expert corrections mix student and expert rollouts to address covariate shift"}, {"area": "code-swe", "role": "train-agent", "why": "Imitation learning of multi-turn SWE agents"}], "pattern": "vary-data", "fixes": ["SFT recipe"], "varies": ["trajectory data generation"], "benchmarks": [], "results": [], "workflow": ["Student starts rollout", "Switch to expert mid-trajectory", "Fine-tune on mixed data"], "summary": "On-policy expert corrections generate partially on-policy data to improve imitation learning for SWE agents."}
|
|
{"id": 6358, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Multi-objective RL with Pareto ranking for tool-using LLMs balancing accuracy and efficiency"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Hypervolume-guided scalarization", "Pareto-ranking advantage"], "summary": "ParetoPO aligns tool-integrated agents under competing objectives using Pareto-ranking policy optimization."}
|
|
{"id": 7123, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Multi-turn RL trains SQL generation agent with execution feedback in ReAct loop"}, {"area": "code-swe", "role": "scaffold", "why": "Three-role multi-agent workflow with validation"}], "pattern": "fix-env-vary-agent", "fixes": ["databases"], "varies": ["agent training"], "benchmarks": [], "results": [], "workflow": ["Schema grounding", "RL-trained query generation", "Solution validation"], "summary": "MARS-SQL makes a text-to-SQL multi-agent workflow trainable with multi-turn RL on live database execution."}
|
|
{"id": 13139, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Self-audited verification of candidate beliefs before agent action commitment"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["verification procedure"], "benchmarks": [], "results": [], "workflow": ["Generate persona-based beliefs", "Verify faithfulness", "Commit action"], "summary": "SAVeR enforces verification over internal belief states before action to improve faithful agent reasoning."}
|
|
{"id": 15395, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "scaffold", "why": "Knapsack context selection and NLI-guided MCTS for faithful RAG generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["RAG pipeline"], "benchmarks": [], "results": [], "workflow": ["MMKP context selection", "NLI-guided MCTS", "Validate faithfulness"], "summary": "Self-Correcting RAG formalizes context selection as knapsack and uses NLI-guided MCTS for faithfulness."}
|
|
{"id": 5145, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Two-stage DAG-reasoning and RL-trained clarifying question asking in RAG"}], "pattern": "fix-env-vary-agent", "fixes": ["RAG setting"], "varies": ["question-asking strategy"], "benchmarks": [], "results": [], "workflow": ["DAG reasoning over retrieved rules", "Traversal-based clarification", "RL training"], "summary": "GPS trains LLMs to proactively ask clarifying questions in RAG using DAG reasoning structure."}
|
|
{"id": 3544, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Visual autoregressive next-scale prediction framework for instruction-guided image editing"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["generation paradigm"], "benchmarks": [], "results": [], "workflow": [], "summary": "VAREdit casts image editing as next-scale prediction with visual autoregressive modeling."}
|
|
{"id": 14683, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Training-free multi-agent debate with critical agent for risk-aware embodied planning"}, {"area": "embodied-nav", "role": "benchmark", "why": "SafeAware-VH adversarial safe-instruction benchmark"}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["debate architecture"], "benchmarks": [{"name": "SafeAware-VH", "purpose": "probe over-rejection of benign instructions", "is_new": true}], "results": [], "workflow": ["Agents debate plans", "Critical agent evaluates via Toulmin", "Select safe plan"], "summary": "MADRA uses multi-agent debate for safer embodied planning and introduces SafeAware-VH."}
|
|
{"id": 6796, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Cost-aware multi-turn RL for task-oriented dialogue service agents"}, {"area": "social-sim", "role": "environment", "why": "User-centric interaction framework as persona-driven training gym"}], "pattern": "fix-env-vary-agent", "fixes": ["user simulator"], "varies": ["policy optimization"], "benchmarks": [], "results": [], "workflow": ["Persona-driven user gym", "CMPO hybrid advantage", "PID-Lagrangian cost control"], "summary": "InteractCS-RL trains service agents with cost-aware multi-turn RL in a persona-driven interaction gym."}
|
|
{"id": 13028, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Metacognitive reflection distills principles and procedures into optimized instructions in one cycle"}], "pattern": "vary-scaffold", "fixes": ["LLM weights"], "varies": ["self-improvement loop"], "benchmarks": [], "results": [], "workflow": ["Principle-based reflection", "Procedural reflection", "Synthesize optimized instructions"], "summary": "MARS achieves efficient agent self-evolution via principle and procedural reflection in a single cycle."}
|
|
{"id": 11920, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "MCP-based benchmark of 118 medical calculator scenarios with EHR interaction"}, {"area": "domain-agents", "role": "benchmark", "why": "Clinical calculator tasks with process-level evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["23 LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "MedMCP-Calc", "purpose": "evaluate LLMs on medical calculator scenarios", "is_new": true}], "results": [], "workflow": [], "summary": "MedMCP-Calc benchmarks LLMs on realistic multi-stage medical calculator scenarios via MCP, with 23 models evaluated."}
|
|
{"id": 17167, "in_agent_scope": false, "types": [{"area": "eval/robust", "role": "analysis", "why": "Linguistic-feature explanation of AI-text detector generalization across prompts, models, domains"}], "pattern": "vary-conditions", "fixes": ["detectors"], "varies": ["generation conditions"], "benchmarks": [{"name": "AI-text detection benchmark", "purpose": "cross-prompt/model/dataset generalization", "is_new": true}], "results": [], "workflow": [], "summary": "Systematic study explaining AI-text detector generalization failures with linguistic feature shifts."}
|
|
{"id": 17346, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Multi-agent system validating and remediating web-sourced data from web agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": [], "summary": "The AI Committee is a model-agnostic multi-agent system that validates and fixes web-sourced datasets."}
|
|
{"id": 8348, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Failure-aware agentic controller with closed-loop physics verification for STEM crystal reconstruction"}, {"area": "science-data", "role": "benchmark", "why": "STEM2Mat-Bench dataset of annotated STEM samples"}], "pattern": "vary-scaffold", "fixes": ["modules"], "varies": ["controller"], "benchmarks": [{"name": "STEM2Mat-Bench", "purpose": "evaluate crystal structure reconstruction", "is_new": true}], "results": [], "workflow": ["Denoise and retrieve templates", "Reconstruct atomic structure", "MLIP validation, rollback and retry"], "summary": "AutoMat is an agentic controller with physics verification for STEM crystal reconstruction, with STEM2Mat-Bench."}
|
|
{"id": 15745, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "eval-method", "why": "Capability-oriented testing and failure attribution for VLN agents via adaptive test generation"}], "pattern": "vary-evaluation", "fixes": ["VLN agents"], "varies": ["testing method"], "benchmarks": [], "results": [], "workflow": ["Seed selection and mutation", "Capability oracles", "Attribute failures and feed back"], "summary": "Testing approach that discovers and attributes VLN agent failures to specific capabilities."}
|
|
{"id": 17425, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "data", "why": "LLM agents synthesize realistic digital footprints from user profiles for fine-tuning"}], "pattern": "vary-data", "fixes": ["fine-tuning recipe"], "varies": ["synthetic data"], "benchmarks": [], "results": [], "workflow": ["Structured user profile", "Generate event sequences", "Produce digital artifacts"], "summary": "PersonaTrace synthesizes diverse digital footprints with LLM agents, improving models fine-tuned on them."}
|
|
{"id": 12987, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Entropy-reduction based sparse and dense rewards to optimize tool-use behavior"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["reward design"], "benchmarks": [], "results": [], "workflow": [], "summary": "Uses entropy reduction as supervision, with outcome and process rewards, to improve efficiency and quality of tool calls."}
|
|
{"id": 15331, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Memory-augmented social simulation with norms and forgetting for deep research in social sciences"}, {"area": "search-research", "role": "application", "why": "Deep research agent for social science paper writing"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["simulation pipeline"], "benchmarks": [], "results": [], "workflow": ["Goal-path planning with norms", "Memory cold-start dataset", "Ebbinghaus-style forgetting"], "summary": "MASS grounds deep research in memory-augmented social simulations to improve social science research generation."}
|
|
{"id": 2860, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "analysis", "why": "Evaluates LLM agents detecting underspecification, asking clarifications, and using interaction in code tasks"}], "pattern": "vary-conditions", "fixes": ["methods"], "varies": ["models", "instruction specificity"], "benchmarks": [], "results": [], "workflow": [], "summary": "Study of LLM agents' ability to detect underspecified instructions and use clarification in interactive code generation."}
|
|
{"id": 1934, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "AstaBench suite with reproducible tools for rigorous science-agent evaluation"}, {"area": "search-research", "role": "benchmark", "why": "Literature and research tasks with baseline agents"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "AstaBench", "purpose": "holistic science research agent evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "AstaBench is a scientific research benchmark suite with controlled tools, cost accounting and baseline agents."}
|
|
{"id": 8941, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Fine-tuned auditor agent roleplays malicious users to stress-test CLI agents on illegal tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["CLI agents"], "varies": ["adversarial user interaction"], "benchmarks": [{"name": "ANCHOR", "purpose": "audit CLI agents on court-case illegal tasks", "is_new": true}], "results": [], "workflow": ["Auditor trained on dark personality data", "Decompose and reframe requests", "Measure agent compliance"], "summary": "ANCHOR red-teams CLI agents with a trained malicious-user auditor, finding compliance reaches 100% under persistence."}
|
|
{"id": 11220, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Multi-answer RL objective trains LMs to output multiple hypotheses with confidence"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["RL objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "Multi-answer RL trains LMs for distributional reasoning, producing several plausible answers with confidences."}
|
|
{"id": 12462, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Biomedical QA agent with query refinement, reflective retrieval and cited responses"}, {"area": "search-research", "role": "scaffold", "why": "Self-critic MeSH query refinement and batch reflective retrieval"}], "pattern": "vary-scaffold", "fixes": ["GPT-4o"], "varies": ["retrieval workflow"], "benchmarks": [{"name": "PubMedQA", "purpose": "biomedical QA accuracy", "is_new": false}], "results": [{"benchmark": "PubMedQA", "system": "PubMed Reasoner", "base_model": "GPT-4o", "setting": "proposed", "metric": "accuracy", "score": 78.32, "unit": "%"}], "workflow": ["Self-critic query refinement", "Reflective batch retrieval", "Cited answer generation"], "summary": "PubMed Reasoner is a biomedical QA agent refining queries and retrieving reflectively, reaching 78.32% on PubMedQA."}
|
|
{"id": 13762, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "method", "why": "LLM pseudo-labeling with code-switched prompts and contrastive distillation for zero-shot cross-lingual ASTE"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["transfer method"], "benchmarks": [], "results": [], "workflow": ["Code-switched structured prompting", "Consistency filtering", "Contrastive distillation"], "summary": "CLAOCS-TX transfers aspect-sentiment triplet extraction across languages using LLM pseudo-labels and distillation."}
|
|
{"id": 9201, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "VLM agent framework generating editable vector scientific illustrations"}], "pattern": "vary-scaffold", "fixes": ["VLMs"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "LiveFigure is an agentic framework generating editable scientific figures with VLM agents."}
|
|
{"id": 14866, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "Finance and accounting spreadsheet workflow benchmark from real enterprise workspaces"}, {"area": "science-data", "role": "benchmark", "why": "Multi-step data entry, calculation and reporting workflows"}], "pattern": "fix-agent-vary-env", "fixes": ["AI agents under test"], "varies": ["workflows"], "benchmarks": [{"name": "Finch", "purpose": "evaluate agents on finance workflows", "is_new": true}], "results": [], "workflow": [], "summary": "Finch benchmarks AI agents on real enterprise finance and accounting workflows mined from Enron and other sources."}
|
|
{"id": 15106, "in_agent_scope": true, "types": [{"area": "science-data", "role": "train-agent", "why": "Planner with tool environment trained by SFT trajectories and online RL for protein design"}, {"area": "science-data", "role": "application", "why": "Text-guided protein design agent"}], "pattern": "fix-env-vary-agent", "fixes": ["tool environment"], "varies": ["agent training"], "benchmarks": [], "results": [], "workflow": ["LLM plans", "Tool feedback", "Reflective plan revision"], "summary": "ProtoCycle is an RL-trained reflective tool-augmented planner for text-guided protein design."}
|
|
{"id": 12117, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "REFORM discovers RM failures via RM-guided decoding and fine-tunes on them for robustness"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["RM training augmentation"], "benchmarks": [{"name": "Anthropic Helpful Harmless", "purpose": "reward model robustness", "is_new": false}, {"name": "PKU Beavertails", "purpose": "reward model robustness", "is_new": false}], "results": [], "workflow": ["Reward-guided decoding finds failures", "Fine-tune RM on augmentations"], "summary": "REFORM finds reward-model failures with controlled decoding and fine-tunes to improve robustness."}
|
|
{"id": 9794, "in_agent_scope": true, "types": [{"area": "science-data", "role": "theory", "why": "Statistical provability theory with reachability MDP bounds for agentic theorem provers"}], "pattern": "n/a", "fixes": [], "varies": ["assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Formal analysis bounding the provability gap of learned agentic theorem provers modeled as finite-horizon MDPs."}
|
|
{"id": 9231, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "GRPO-trained decision agent selects cluster number for multi-view clustering"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["clustering method"], "benchmarks": [], "results": [], "workflow": [], "summary": "GROK adapts GRPO to autonomously determine cluster number in unknown-K multi-view clustering."}
|
|
{"id": 14268, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Diagnostic QA framework with persistent state and root-cause retrieval aggregation for IT support"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["RAG pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "DQA maintains diagnostic state and aggregates cases by root cause for multi-turn enterprise IT troubleshooting."}
|
|
{"id": 18388, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "analysis", "why": "Annotated analysis of caregiver QA feedback types in child-directed dialogue"}], "pattern": "n/a", "fixes": [], "varies": ["development stage"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analysis and dataset release of caregiver feedback adapting over child language development."}
|
|
{"id": 16316, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Brain-inspired multi-subsystem memory architecture with timeline-based retrieval"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory architecture"], "benchmarks": [{"name": "LoCoMo", "purpose": "long-term memory evaluation", "is_new": false}], "results": [{"benchmark": "LoCoMo", "system": "BMAM", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 78.45, "unit": "%"}], "workflow": ["Specialized memory subsystems", "Six-phase lifecycle", "Fused evidence retrieval"], "summary": "BMAM is a brain-inspired multi-subsystem agent memory achieving 78.45% accuracy on LoCoMo."}
|
|
{"id": 15602, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "train-agent", "why": "Training framework decoupling QA-based reasoning from character description generation"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["training framework"], "benchmarks": [], "results": [], "workflow": ["Reasoning model produces QA trace", "Generator writes description"], "summary": "Decouples structured QA reasoning from generation to improve character descriptions from long narratives."}
|
|
{"id": 7152, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "EvoClaw benchmark of continuous software evolution built with DeepCommit milestone DAG pipeline"}], "pattern": "fix-agent-vary-env", "fixes": ["12 models, 4 frameworks"], "varies": ["tasks"], "benchmarks": [{"name": "EvoClaw", "purpose": "long-term software evolution", "is_new": true}], "results": [], "workflow": [], "summary": "EvoClaw benchmarks agents on continuous software evolution via milestone DAGs reconstructed from commit logs."}
|
|
{"id": 347, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Adaptive agent foundation model with route-then-align training over agentic, reasoning and instant modes"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["model training"], "benchmarks": [], "results": [], "workflow": ["Learn task-aware routing", "Align mode-specific trajectories"], "summary": "A2FM unifies agentic, reasoning and instant modes under one backbone to reduce overthinking and over-calling tools."}
|
|
{"id": 14990, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Graph memory compression with time-decay retrieval for streaming video agents"}], "pattern": "vary-scaffold", "fixes": ["agent model"], "varies": ["memory compression"], "benchmarks": [{"name": "M3-Bench-robot", "purpose": "video memory evaluation", "is_new": false}, {"name": "M3-Bench-web", "purpose": "video memory evaluation", "is_new": false}, {"name": "Video-MME-Long", "purpose": "video memory evaluation", "is_new": false}], "results": [], "workflow": ["Minmax sampling isolated nodes", "Edge-aware pruning", "Time-decay retrieval"], "summary": "StreamMeCo compresses agent memory graphs for streaming video while preserving accuracy."}
|
|
{"id": 9575, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "RL probe identifies reasoning-critical attention heads for KV cache compression"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["compression method"], "benchmarks": [], "results": [], "workflow": ["RL probe over head cache usage", "Full cache for key heads", "Compress others"], "summary": "RLKV uses RL to find reasoning-critical heads and allocates KV cache accordingly."}
|
|
{"id": 3699, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Adaptive factual-guided activation editing to reduce LVLM object hallucination"}], "pattern": "n/a", "fixes": ["LVLM weights"], "varies": ["activation editing"], "benchmarks": [], "results": [], "workflow": [], "summary": "AFTER edits LVLM activations guided by factual text semantics to mitigate object hallucination."}
|
|
{"id": 7040, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Benchmark of human-in-the-loop value in agentic coding with collaboration-necessary templates"}], "pattern": "fix-agent-vary-env", "fixes": ["5 LLMs, 45 participants"], "varies": ["intervention level"], "benchmarks": [{"name": "CentaurEval", "purpose": "human-AI collaboration in coding", "is_new": true}], "results": [], "workflow": [], "summary": "CentaurEval measures human-in-the-loop value in coding with 450 collaboration-necessary tasks."}
|
|
{"id": 13572, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Structure-guided RAG with subgraph-matching retrieval for exact factual queries"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["retrieval method"], "benchmarks": [], "results": [], "workflow": [], "summary": "SG-RAG formalizes the Exact Retrieval Problem and retrieves via embedding-based subgraph matching."}
|
|
{"id": 13497, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "benchmark", "why": "Process-level benchmark of small LM reasoning traces on commonsense QA"}, {"area": "eval/judge", "role": "eval-method", "why": "Reasoning-aware judging with LLM judges"}], "pattern": "vary-evaluation", "fixes": ["SLMs"], "varies": ["evaluation protocol"], "benchmarks": [{"name": "ReTraceQA", "purpose": "reasoning trace validity", "is_new": true}], "results": [], "workflow": [], "summary": "ReTraceQA reveals SLMs often reach correct answers with flawed reasoning, via expert-annotated process evaluation."}
|
|
{"id": 11328, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Large-scale evaluation of 17 LLMs on Hanabi under varying scaffolds and player counts"}], "pattern": "vary-conditions", "fixes": ["Hanabi"], "varies": ["models", "scaffold", "players"], "benchmarks": [{"name": "Hanabi", "purpose": "cooperative reasoning under incomplete information", "is_new": false}], "results": [], "workflow": [], "summary": "Evaluates 17 LLMs as Hanabi agents across scaffolds, finding best models trail human experts."}
|
|
{"id": 16276, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Inference-time contrastive alignment triggered by attention entropy and image correlation indicators"}], "pattern": "n/a", "fixes": ["MLLM weights"], "varies": ["decoding alignment"], "benchmarks": [], "results": [], "workflow": [], "summary": "DICA monitors two indicators during inference and applies contrastive alignment to reduce MLLM hallucination."}
|