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
32 KiB
JSON
61 lines
32 KiB
JSON
{"id": 13246, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "CalConflictBench for long-horizon calendar conflict resolution"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "PEARL: RL-trained self-evolving time-management assistant"}], "pattern": "vary-both", "fixes": ["calendar conflict task"], "varies": ["benchmark", "RL training of agent"], "benchmarks": [{"name": "CalConflictBench", "purpose": "long-horizon calendar conflict resolution", "is_new": true}], "results": [{"benchmark": "CalConflictBench", "system": "Qwen-3-30B-Think", "base_model": null, "setting": "baseline", "metric": "average error rate", "score": 35, "unit": "%"}], "workflow": ["Present conflicts round by round", "Infer user preferences", "Train with RL"], "summary": "Introduces CalConflictBench and PEARL, an RL-trained self-evolving assistant for calendar conflict resolution."}
|
|
{"id": 12101, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Agentic feedback-driven NL2SQL via gold-query rewriting"}, {"area": "code-swe", "role": "train-agent", "why": "Distills lightweight schema-aware rewriter from teacher"}], "pattern": "vary-scaffold", "fixes": ["database", "SQL task"], "varies": ["query rewriting pipeline", "distilled rewriter"], "benchmarks": [], "results": [], "workflow": ["Rewrite question to gold query", "Generate SQL", "Refine with execution feedback"], "summary": "QBridge rewrites questions into SQL-aligned gold queries with a distilled rewriter and agentic refinement for NL2SQL."}
|
|
{"id": 14058, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Open-access publishing platform for human and AI scientists via MCP"}], "pattern": "n/a", "fixes": ["general-purpose models"], "varies": ["publishing platform design"], "benchmarks": [], "results": [], "workflow": ["Host open preprints", "AI-augmented review", "Collect reader feedback"], "summary": "AiraXiv is a deployed open preprint platform with AI-augmented review and MCP access for AI scientists."}
|
|
{"id": 1560, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent system with planner, critics and visual tool agents"}, {"area": "tool-use", "role": "train-agent", "why": "Fine-tunes MLLM into pixel-level grounding model for tool agents"}], "pattern": "vary-scaffold", "fixes": ["MLLM reasoner"], "varies": ["agent system", "grounding tool"], "benchmarks": [], "results": [], "workflow": ["Dispatch to planner", "Ground with CV tools", "Critics verify reasoning"], "summary": "PixelCraft is a multi-agent system with a fine-tuned grounding tool for visual reasoning on charts and diagrams."}
|
|
{"id": 7038, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "RECENT refactors skills as code for small-model skill grounding"}], "pattern": "vary-scaffold", "fixes": ["small language model", "embodied environment"], "varies": ["skill grounding via code refactoring"], "benchmarks": [], "results": [], "workflow": ["Represent skills as code", "Decouple semantics from binding", "Refactor for embodiment"], "summary": "RECENT lets small language models ground reusable embodied skills by refactoring executable skill code."}
|
|
{"id": 18545, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "MoNaCo benchmark of time-consuming multi-document questions"}], "pattern": "fix-agent-vary-env", "fixes": ["frontier LLMs"], "varies": ["question set"], "benchmarks": [{"name": "MoNaCo", "purpose": "natural time-consuming information-seeking questions", "is_new": true}], "results": [{"benchmark": "MoNaCo", "system": "frontier LLMs (best)", "base_model": null, "setting": "baseline", "metric": "F1", "score": 61.2, "unit": "%"}], "workflow": [], "summary": "MoNaCo is a 1,315-question benchmark of natural information-seeking questions requiring many intermediate steps."}
|
|
{"id": 4317, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Semantic-scale guidance hierarchy for precise diffusion image editing"}], "pattern": "n/a", "fixes": ["diffusion models"], "varies": ["guidance editing method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows diffusion guidance magnitude encodes edit semantic scale and uses it for precise text-guided editing."}
|
|
{"id": 15211, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Dynamic reasoning context with sliding window and compression for SWE agents"}], "pattern": "vary-scaffold", "fixes": ["SWE task"], "varies": ["reasoning context management"], "benchmarks": [], "results": [], "workflow": ["Keep recent detailed reasoning", "Compress older reasoning", "Act on task"], "summary": "SWE-AGILE manages reasoning history with a sliding window plus compression for efficient multi-turn SWE agents."}
|
|
{"id": 13322, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "RL-guided test-time tuning for OOD harmful text detection"}], "pattern": "n/a", "fixes": ["detection task"], "varies": ["adaptive tuning method"], "benchmarks": [], "results": [], "workflow": ["Tune with consistency loss", "RL-guide adaptation", "Prevent overfitting"], "summary": "RLAT adapts harmful text detectors at test time using RL-guided tuning to handle distribution shift."}
|
|
{"id": 14745, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "SWE-QA repository-level code question answering benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["language models"], "varies": ["repo-level QA tasks"], "benchmarks": [{"name": "SWE-QA", "purpose": "repository-level code QA, 720 pairs", "is_new": true}], "results": [], "workflow": [], "summary": "SWE-QA is a 720-pair benchmark for repository-level code question answering across files and dependencies."}
|
|
{"id": 6858, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "RL post-training of CodeAct agent for FHIR question answering"}, {"area": "tool-use", "role": "train-agent", "why": "Multi-turn tool-calling agent trained with custom harness"}], "pattern": "fix-env-vary-agent", "fixes": ["FHIR-AgentBench"], "varies": ["RL-trained agent", "reward"], "benchmarks": [{"name": "FHIR-AgentBench", "purpose": "hospital-data QA evaluation", "is_new": false}], "results": [], "workflow": ["Run CodeAct agent", "LLM judge reward", "RL update"], "summary": "Post-trains a multi-turn CodeAct agent with RL on FHIR clinical question answering."}
|
|
{"id": 14947, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Stage-wise distillation of explainable hateful video detection into small multimodal models"}], "pattern": "n/a", "fixes": ["detection task"], "varies": ["distillation method"], "benchmarks": [], "results": [], "workflow": [], "summary": "LEAF distills explainable hateful video detection from large to small multimodal models via self-grounding CoT."}
|
|
{"id": 12296, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Trains global planner via SFT then RL with executor capability gain reward"}], "pattern": "fix-env-vary-agent", "fixes": ["executor agent", "long-horizon tasks"], "varies": ["planner training"], "benchmarks": [], "results": [], "workflow": ["Synthesize plans, filter", "SFT cold start", "RL with capability gain reward"], "summary": "EAGLET trains a plug-and-play global planner with consensus-filtered SFT and RL to help executor agents."}
|
|
{"id": 11890, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "RL fine-tuning of multimodal conversational agents in learned latent action space"}], "pattern": "fix-env-vary-agent", "fixes": ["conversational tasks"], "varies": ["action space", "RL fine-tuning"], "benchmarks": [], "results": [], "workflow": ["Learn latent action codebook", "Use paired and unpaired data", "RL in latent space"], "summary": "Fine-tunes multimodal conversational agents with RL in a compact latent action space with improved codebook coverage."}
|
|
{"id": 8039, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Trains Todo-14B with preference optimization to synthesize planning architectures"}, {"area": "planning-decision", "role": "scaffold", "why": "PlanFactory unified design space of planning structures"}], "pattern": "vary-scaffold", "fixes": ["tasks"], "varies": ["planning architecture", "planner model"], "benchmarks": [], "results": [], "workflow": ["Build PlanFactory space", "Collect planning trajectories", "Train with IGPO"], "summary": "TodoEvolve trains a meta-planner that synthesizes and revises task-specific agent planning architectures over PlanFactory."}
|
|
{"id": 15601, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Evolving hypergraph unifies agent collaboration topology and shared memory"}], "pattern": "vary-scaffold", "fixes": ["LLM agents"], "varies": ["collaboration topology"], "benchmarks": [], "results": [], "workflow": ["Hypergraph binds agents and memory", "Controller edits hypergraph", "Execute collaboration"], "summary": "EvoHyper uses an evolving hypergraph to unify multi-agent communication structure and shared memory, cutting redundancy."}
|
|
{"id": 13757, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "AgentGym2 de-idealized real-world agent benchmark with tool discovery"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["realistic noisy tasks"], "benchmarks": [{"name": "AgentGym2", "purpose": "de-idealized end-to-end agent evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "AgentGym2 evaluates agents on noisy real-world tasks requiring tool discovery, composition and end-to-end execution."}
|
|
{"id": 15848, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "TraineeBench dynamic environment for scheduling, exploration and continuous learning"}], "pattern": "fix-agent-vary-env", "fixes": ["MLLM agents"], "varies": ["dynamic workplace tasks"], "benchmarks": [{"name": "TraineeBench", "purpose": "workplace scheduling, exploration, learning", "is_new": true}], "results": [], "workflow": [], "summary": "TraineeBench is a dynamic evaluation environment for trainee-like agents scheduling, exploring and learning from experience."}
|
|
{"id": 7035, "in_agent_scope": false, "types": [{"area": "reason/math", "role": "method", "why": "Multi-turn off-policy RL and multi-agent tree search for LLM step-provers"}], "pattern": "n/a", "fixes": ["theorem proving task"], "varies": ["RL training", "inference search"], "benchmarks": [], "results": [], "workflow": ["Expert iteration with filtering", "Periodic retraining", "Planner multi-agent search"], "summary": "BFS-Prover-V2 scales step-level theorem proving with multi-turn off-policy RL and planner-enhanced multi-agent search."}
|
|
{"id": 15841, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "application", "why": "Uncertainty-guided distillation for student misconception classification"}], "pattern": "n/a", "fixes": ["classification task"], "varies": ["sample selection for distillation"], "benchmarks": [], "results": [], "workflow": [], "summary": "Two-stage distillation selects critical samples by cognitive uncertainty to classify student misconceptions with small models."}
|
|
{"id": 14687, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Training-free scene and motion-aware token pruning for VideoLLMs"}], "pattern": "n/a", "fixes": ["VideoLLMs"], "varies": ["token pruning"], "benchmarks": [], "results": [], "workflow": [], "summary": "MoPrune prunes video tokens using scene segmentation and motion saliency to accelerate VideoLLMs without training."}
|
|
{"id": 17451, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "application", "why": "Regulation-aware NLI model for detecting product misbranding"}], "pattern": "n/a", "fixes": ["task"], "varies": ["model"], "benchmarks": [{"name": "FDA-Misbrand", "purpose": "evaluating misbranding detection", "is_new": false}], "results": [], "workflow": [], "summary": "RegNLI frames misbranding detection as inference between product claims and FDA statutes with gated contrastive alignment."}
|
|
{"id": 14781, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Gome MLE agent mapping gradient-style optimization onto reasoning and memory"}], "pattern": "vary-scaffold", "fixes": ["LLMs", "MLE-Bench"], "varies": ["agent optimization architecture"], "benchmarks": [{"name": "MLE-Bench", "purpose": "ML engineering agent evaluation", "is_new": false}], "results": [{"benchmark": "MLE-Bench", "system": "Gome", "base_model": null, "setting": "proposed", "metric": "any-medal rate", "score": 35.1, "unit": "%"}], "workflow": ["Diagnostic reasoning as gradient", "Success memory as momentum", "Multi-trace execution"], "summary": "Gome is an MLE agent using diagnostic reasoning as gradients instead of tree search."}
|
|
{"id": 15260, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Self-evolving multi-agent RAG for medical reasoning"}, {"area": "multi-agent", "role": "scaffold", "why": "Decouples interpretation, exploration and adjudication across agents"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-agent RAG workflow"], "benchmarks": [], "results": [], "workflow": ["Interpret question", "Iterative retrieval with sufficiency feedback", "Adjudicate evidence"], "summary": "SEMA-RAG splits medical RAG into role-specialized agents with multi-round exploration and self-evolution."}
|
|
{"id": 14769, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Risk-aware topology evolution to block malicious injection spread in MAS"}], "pattern": "fix-agent-vary-env", "fixes": ["attack threat model"], "varies": ["communication topology"], "benchmarks": [], "results": [], "workflow": ["Monitor with ST-GNN", "Score node and edge risk", "Prune and isolate"], "summary": "TopoSHIELD monitors multi-agent interactions with a graph network and prunes risky edges to block injection diffusion."}
|
|
{"id": 16323, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "MAMA protocol seeds PII and extracts it via attacker agent"}, {"area": "agent-safety", "role": "analysis", "why": "Compares memory leakage across six communication topologies"}], "pattern": "vary-conditions", "fixes": ["multi-agent LLM systems"], "varies": ["graph topology"], "benchmarks": [{"name": "MAMA", "purpose": "topology-conditioned memory leakage", "is_new": true}], "results": [], "workflow": ["Seed PII in memory", "Attacker interacts 10 rounds", "Measure leakage"], "summary": "MAMA measures how multi-agent graph topology affects private memory leakage under attacker extraction."}
|
|
{"id": 4003, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Recurrent memory transformer for offline RL in POMDPs"}], "pattern": "n/a", "fixes": ["offline RL benchmarks"], "varies": ["architecture"], "benchmarks": [{"name": "ViZDoom-Two-Colors", "purpose": "memory-intensive evaluation", "is_new": false}, {"name": "Memory Maze", "purpose": "memory-intensive evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "RATE adds a recurrent memory mechanism to action transformers for offline RL in partially observable environments."}
|
|
{"id": 13857, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Adaptive reasoning trees guided by minimal free resolution theory"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["reasoning tree construction"], "benchmarks": [], "results": [], "workflow": [], "summary": "SyRA builds adaptive reasoning trees with MFR-guided branching and residual backtracking for error correction."}
|
|
{"id": 12989, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "LLM Monte Carlo estimation of affective trajectories and ambiguity"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["probabilistic affect modeling"], "benchmarks": [], "results": [], "workflow": [], "summary": "LLM-MC-Affect models emotion as latent distributions via stochastic LLM decoding to analyze interpersonal dynamics."}
|
|
{"id": 15434, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Temporal-hierarchical memory tree consolidating conversations without fine-tuning"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory organization"], "benchmarks": [], "results": [], "workflow": ["Build temporal memory tree", "Consolidate to persona", "Complexity-aware recall"], "summary": "TiMem organizes conversational memory in a temporal tree with consolidation and complexity-aware recall."}
|
|
{"id": 12436, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "eval-method", "why": "MATU quantifies MAS uncertainty via tensor decomposition"}], "pattern": "vary-evaluation", "fixes": ["multi-agent systems"], "varies": ["uncertainty metric"], "benchmarks": [], "results": [], "workflow": [], "summary": "MATU measures uncertainty in multi-agent LLM systems using tensor decomposition over communication paths and topologies."}
|
|
{"id": 10992, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Training-free latent-space collaboration among LLM agents"}], "pattern": "vary-scaffold", "fixes": ["frozen LLMs"], "varies": ["communication medium"], "benchmarks": [], "results": [], "workflow": ["Generate latent thoughts", "Share latent working memory", "Decode final answer"], "summary": "LatentMAS lets LLM agents collaborate via hidden-state latent thoughts and shared latent working memory without training."}
|
|
{"id": 16650, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "GUI Grounding Sensitivity Benchmark for desktop interfaces"}, {"area": "web-gui", "role": "analysis", "why": "Tests grounder sensitivity to varied descriptions of same element"}], "pattern": "vary-conditions", "fixes": ["grounding models"], "varies": ["instruction phrasing"], "benchmarks": [{"name": "GUI Grounding Sensitivity Benchmark", "purpose": "sensitivity to multiple descriptions of same element", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmarks desktop GUI grounding models on consistency across diverse valid descriptions of the same UI element."}
|
|
{"id": 14322, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Compares context strategies and tool use for multi-format document QA"}], "pattern": "vary-conditions", "fixes": ["six frontier models"], "varies": ["context strategy", "tool access"], "benchmarks": [], "results": [{"benchmark": "multi-format document QA", "system": "RAG+Tools", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 46, "unit": "%"}, {"benchmark": "multi-format document QA", "system": "RAG-only", "base_model": null, "setting": "baseline", "metric": "accuracy", "score": 6, "unit": "%"}], "workflow": [], "summary": "Evaluates full-context, RAG and tool-augmented strategies on Word, Excel and PowerPoint, finding tools help substantially."}
|
|
{"id": 16257, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Hierarchical visual agent with zoom tool and joint image-text context"}], "pattern": "vary-scaffold", "fixes": ["MLLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["Manager plans", "Sub-agents gather evidence", "Update compact context"], "summary": "HierVA uses a manager and sub-agents with a zoom tool to manage joint visual-text context for chart reasoning."}
|
|
{"id": 11874, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Training-free diversity-aware masking RAG for faithful fact correction"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["RAG correction framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Mask-to-Correct is an inference-only RAG framework using masking and retriever diversity to correct claims faithfully."}
|
|
{"id": 11930, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Automated pipeline synthesizing MCP instruction-call pairs and trajectories"}, {"area": "tool-use", "role": "train-agent", "why": "Trains models on synthesized MCP data"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["MCP training data"], "benchmarks": [], "results": [], "workflow": ["Discover servers and tools", "Synthesize instructions and calls", "Train models"], "summary": "MCP-Flow discovers MCP servers and synthesizes 68,733 instruction-call pairs and 6,439 trajectories to train tool-use models."}
|
|
{"id": 4641, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Response-conditioned DAG routing using Shapley-style contribution estimates"}], "pattern": "vary-scaffold", "fixes": ["LLM agents"], "varies": ["communication topology"], "benchmarks": [], "results": [], "workflow": ["Agents answer independently", "Estimate peer contributions", "Build DAG routing"], "summary": "Adapts multi-agent communication on the fly by routing messages from high-contribution agents without extra training."}
|
|
{"id": 13520, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "defense", "why": "Intent-guided safety reasoning mitigating Safety Context Amnesia in multimodal reasoning models"}], "pattern": "n/a", "fixes": ["threat model"], "varies": ["safety reasoning method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Identifies Safety Context Amnesia in multimodal reasoning models and mitigates it with intent-guided safety reasoning."}
|
|
{"id": 13883, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Dual-head model predicting implicit preferences for multi-objective alignment"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["alignment method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Self-Guided Alignment predicts implicit user preferences to remove manual specification in multi-objective generation."}
|
|
{"id": 7655, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "environment", "why": "MEnvAgent automates multilingual executable environment construction"}, {"area": "code-swe", "role": "benchmark", "why": "MEnvBench for environment construction"}], "pattern": "fix-agent-vary-env", "fixes": ["SWE agents"], "varies": ["executable environments"], "benchmarks": [{"name": "MEnvBench", "purpose": "evaluating environment construction", "is_new": true}], "results": [], "workflow": ["Plan construction", "Execute and verify", "Reuse patched environments"], "summary": "MEnvAgent builds verifiable SWE task environments across languages with multi-agent planning, execution, verification and reuse."}
|
|
{"id": 9786, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "ActHook watermarks agent trajectory datasets with hook actions"}], "pattern": "vary-data", "fixes": ["task outcome"], "varies": ["watermark in trajectories"], "benchmarks": [], "results": [], "workflow": ["Embed hook actions", "Activate by secret key", "Verify ownership"], "summary": "ActHook embeds key-activated hook actions into agent trajectories to enable dataset copyright tracing."}
|
|
{"id": 5506, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Learned adaptive frame sampling with CLIP and uniform agents for long video"}], "pattern": "n/a", "fixes": ["VLM"], "varies": ["frame sampling policy"], "benchmarks": [], "results": [], "workflow": [], "summary": "VideoBrain trains VLMs to adaptively acquire frames using complementary retrieval and uniform sampling agents."}
|
|
{"id": 9085, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Uncertainty-guided token and turn-level exploration control for multi-turn agentic RL"}], "pattern": "fix-env-vary-agent", "fixes": ["interactive tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "T2PO controls exploration at token and turn level to stabilize multi-turn agentic RL."}
|
|
{"id": 1153, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "Benchmark for agent memory via incremental multi-turn interactions"}], "pattern": "fix-agent-vary-env", "fixes": ["memory agents"], "varies": ["memory tasks"], "benchmarks": [{"name": "memory agent benchmark", "purpose": "retrieval, test-time learning, long-range understanding, forgetting", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmarks memory agents on accurate retrieval, test-time learning, long-range understanding and selective forgetting."}
|
|
{"id": 14281, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "application", "why": "Data-efficient recipe for grounded banking QA 12B model"}], "pattern": "n/a", "fixes": ["training framework"], "varies": ["domain data pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "FinRAG-12B trains a grounded, citation-aware banking QA model with judge-filtered data and curriculum learning."}
|
|
{"id": 15425, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Continual tool documentation adaptation as toolsets evolve"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Self-evolving agents via documentation updates"}], "pattern": "vary-scaffold", "fixes": ["frozen LLM"], "varies": ["tool documentation"], "benchmarks": [], "results": [], "workflow": ["Relation-guided exploration", "Update documentation", "Retain prior competence"], "summary": "ContDa lets tool agents self-evolve by adapting tool documentation under changing toolsets, balancing stability and adaptation."}
|
|
{"id": 12142, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Verify-before-commit protocol against tool-stream injection"}], "pattern": "fix-agent-vary-env", "fixes": ["injection threat model"], "varies": ["defense protocol"], "benchmarks": [], "results": [], "workflow": ["Generate speculative hypotheses", "Verify against intent", "Commit action"], "summary": "VIGIL defends agents from indirect tool-stream prompt injection via speculative hypotheses and intent-grounded verification."}
|
|
{"id": 11107, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Latent-aware guidance search for unmasking in diffusion LMs"}], "pattern": "n/a", "fixes": ["diffusion LM"], "varies": ["decoding strategy"], "benchmarks": [], "results": [], "workflow": [], "summary": "LUGS guides diffusion LM unmasking with hidden-state latent scores beyond confidence and entropy heuristics."}
|
|
{"id": 2685, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Signal-quality-guided multimodal MoE for cognitive load prediction"}], "pattern": "n/a", "fixes": ["task"], "varies": ["model"], "benchmarks": [], "results": [], "workflow": [], "summary": "CogMoE uses quality-aware gating over EEG, ECG, EDA and gaze for robust cognitive load prediction."}
|
|
{"id": 7066, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Graph-structured thinking flows with controller for research ideation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["ideation workflow"], "benchmarks": [], "results": [], "workflow": ["Controller selects operators", "Compose thinking flow", "Generate and evaluate ideas"], "summary": "MindFlow dynamically composes modular thinking operators via a probabilistic supernet to generate research ideas."}
|
|
{"id": 13771, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Knowledge-graph-guided LLM agent for HS code classification"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["KG-guided reasoning"], "benchmarks": [], "results": [], "workflow": ["Traverse HS knowledge graph", "Reason stepwise", "Validate hierarchy"], "summary": "HSGraphAgent classifies HS codes via stepwise regulation-aware reasoning over a harmonized system knowledge graph."}
|
|
{"id": 13265, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "benchmark", "why": "MMSciCode expert-annotated scientific research coding benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["foundation models"], "varies": ["coding tasks"], "benchmarks": [{"name": "MMSciCode", "purpose": "scientific code generation, 624 problems", "is_new": true}], "results": [], "workflow": [], "summary": "MMSciCode is a 624-problem multilingual multi-discipline benchmark of expert-annotated scientific research coding."}
|
|
{"id": 1504, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Parallel agents with distinct tool-use strategies sharing and refining answers"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["tool-use ensemble"], "benchmarks": [], "results": [], "workflow": ["Run diverse tool agents", "Share responses", "Refine and aggregate"], "summary": "TUMIX ensembles agents using different tool-use strategies that iteratively refine answers for test-time scaling."}
|
|
{"id": 9307, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Group-fair one-shot federated learning with prototype-guided experts"}], "pattern": "n/a", "fixes": ["task"], "varies": ["fairness-aware FL method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Fair-FedMOE reduces subgroup disparity in one-shot federated medical imaging via fairness-aware expert routing."}
|
|
{"id": 813, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Curriculum-guided RL trains low-level mobile executor"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Self-evolving declarative knowledge via Summarize-Reflect-Locate-Revise loop"}], "pattern": "vary-both", "fixes": ["mobile control tasks"], "varies": ["hierarchical agent", "knowledge evolution"], "benchmarks": [], "results": [], "workflow": ["Bootstrap from one demo", "SRLR loop refines knowledge", "RL-train executor"], "summary": "K2-Agent co-evolves declarative planning knowledge and RL-trained procedural execution for hierarchical mobile device control."}
|
|
{"id": 6225, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Judgment Operators enforce allow/edit/escalate/deny at decision time"}, {"area": "multi-agent", "role": "scaffold", "why": "Composition-invariant governance across multi-agent action spaces"}], "pattern": "vary-scaffold", "fixes": ["multi-agent systems"], "varies": ["intervention mechanism"], "benchmarks": [], "results": [], "workflow": [], "summary": "Judgment Operators provide decision-time four-way intervention to enforce constraints with minimal repair across composed agents."}
|
|
{"id": 4173, "in_agent_scope": false, "types": [{"area": "arch/attention", "role": "method", "why": "Hilbert-curve token reordering for block-sparse local attention"}], "pattern": "n/a", "fixes": ["image models"], "varies": ["attention window construction"], "benchmarks": [], "results": [], "workflow": [], "summary": "Reorders image tokens along a Hilbert curve to make local attention windows contiguous and block-sparse efficient."}
|
|
{"id": 12821, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Supervised coreference model plus LLM checker-splitter refinement"}], "pattern": "n/a", "fixes": ["coreference task"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "ImCoref-CeS combines an improved supervised coreference model with LLM-based checking and splitting."}
|
|
{"id": 5390, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "AgentHijack benchmark of nine common environment corruptions"}], "pattern": "fix-agent-vary-env", "fixes": ["computer-use agents"], "varies": ["environment corruptions"], "benchmarks": [{"name": "AgentHijack", "purpose": "computer-use agent robustness under corruptions", "is_new": true}], "results": [], "workflow": [], "summary": "AgentHijack evaluates computer-use agent robustness under nine configurable non-adversarial environment corruptions."}
|