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
28 KiB
JSON
61 lines
28 KiB
JSON
{"id": 17741, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Releases Emirati Arabic speech corpus; zero-shot ASR/TTS baselines"}], "pattern": "fix-agent-vary-env", "fixes": ["existing ASR/TTS models"], "varies": ["speech data/dialect"], "benchmarks": [{"name": "Ramsa", "purpose": "Emirati Arabic ASR and TTS baselines", "is_new": true}], "results": [], "workflow": [], "summary": "A 41-hour Emirati Arabic speech corpus with zero-shot ASR and TTS baselines of commercial and open models."}
|
|
{"id": 4605, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Trains VLM driving agent with active perception tools and hybrid thinking"}], "pattern": "fix-env-vary-agent", "fixes": ["driving benchmark"], "varies": ["training recipe", "tool-based active perception"], "benchmarks": [], "results": [], "workflow": ["Detect uncertainty", "Invoke visual tools", "Hybrid text/tool reasoning", "Plan behavior"], "summary": "DriveAgent-R1 is a VLM driving agent that actively invokes visual tools and switches between thinking modes for planning."}
|
|
{"id": 12807, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Frequency-domain adapter integrating text for multi-scale adaptation"}], "pattern": "vary-data", "fixes": ["pretrained model"], "varies": ["adapter design"], "benchmarks": [], "results": [], "workflow": ["Transform to frequency domain", "Text-guided multi-scale adaptation"], "summary": "FreqAdapter, a parameter-efficient adapter using text-guided multi-scale frequency representations."}
|
|
{"id": 15617, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Investigative multi-agent framework for explainable fake video detection"}, {"area": "domain-agents", "role": "application", "why": "Short-video fake news detection system"}], "pattern": "vary-scaffold", "fixes": ["MLLM backbone"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["Multimodal forensics and retrieval", "Investigative reasoning", "Verdict with explanation"], "summary": "CSI is a multi-agent framework with forensics and reasoning units for explainable short-video fake news detection."}
|
|
{"id": 15273, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Active reflection-driven context management for long-horizon search agents"}], "pattern": "vary-scaffold", "fixes": ["model weights", "search tasks"], "varies": ["context management strategy"], "benchmarks": [], "results": [], "workflow": ["Reflect on history", "Actively restructure context", "Continue search"], "summary": "ARC treats context management as active, reflective process to counter context rot in deep search agents."}
|
|
{"id": 14171, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "analysis", "why": "Evaluates LLM acceptability judgments on 150 minimal pairs"}], "pattern": "vary-conditions", "fixes": ["forced-choice paradigm"], "varies": ["models", "phenomena", "prompting"], "benchmarks": [{"name": "BLiMP", "purpose": "phenomenon categories", "is_new": false}, {"name": "Linguistic Inquiry minimal pairs", "purpose": "150 pairs sampled", "is_new": false}], "results": [], "workflow": [], "summary": "Studies LLM forced-choice acceptability judgments on minimal pairs, finding stronger alignment on morphosyntax than semantics."}
|
|
{"id": 5397, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Tiered benchmark for multi-step scientific tool use"}, {"area": "science-data", "role": "environment", "why": "Interactive environment with 1,780 scientific tools"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated LLM agents"], "varies": ["tools", "task tiers"], "benchmarks": [{"name": "SciAgentGym", "purpose": "interactive scientific tool environment", "is_new": true}, {"name": "SciAgentBench", "purpose": "tiered tool-use evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "SciAgentGym environment with 1,780 scientific tools and SciAgentBench tiered suite for evaluating multi-step tool use."}
|
|
{"id": 17212, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Supervised knowledge-graph-guided router for multi-agent collaboration"}, {"area": "domain-agents", "role": "application", "why": "Nutrition question answering"}], "pattern": "vary-scaffold", "fixes": ["LLM agents"], "varies": ["agent routing/collaboration"], "benchmarks": [], "results": [], "workflow": ["Build nutritional graph", "Supervise router", "Route agents"], "summary": "NG-Router uses graph-supervised routing to orchestrate multi-agent collaboration for nutrition QA."}
|
|
{"id": 9914, "in_agent_scope": false, "types": [{"area": "align/values", "role": "theory", "why": "Formal analysis of competition among misaligned agents yielding aligned outcomes"}], "pattern": "n/a", "fixes": [], "varies": ["agent misalignment", "model diversity"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows that competition among differently misaligned AI agents can match a perfectly aligned one if user utility lies in their convex hull."}
|
|
{"id": 3576, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Website-generation agent with multi-level visual feedback"}, {"area": "code-swe", "role": "train-agent", "why": "Step-level RL for website generation"}], "pattern": "vary-both", "fixes": ["website generation task"], "varies": ["feedback signals", "step-level RL"], "benchmarks": [], "results": [], "workflow": ["Generate code", "Screenshot and GUI feedback", "Refine", "Step-level RL"], "summary": "WebGen-Agent uses screenshot and GUI-testing feedback plus step-level RL to iteratively generate websites."}
|
|
{"id": 17635, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "application", "why": "Interactive tool visualizing coreferential ambiguity"}], "pattern": "n/a", "fixes": [], "varies": ["visualization"], "benchmarks": [{"name": "Phrase Detectives", "purpose": "evaluate the tool", "is_new": false}], "results": [], "workflow": [], "summary": "AmbiCoRefVis visualizes multiple global coreference interpretations of ambiguous referring expressions."}
|
|
{"id": 11474, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Cue-Tag-Content graph memory with active reconstruction"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory structure and access"], "benchmarks": [], "results": [], "workflow": ["Build cue-tag-content graph", "Active reconstruction", "Reason on evidence"], "summary": "MRAgent represents memory as an associative graph and actively reconstructs it during inference for long-history reasoning."}
|
|
{"id": 7986, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Preference pairs from mutual information between prompt and response"}], "pattern": "vary-data", "fixes": ["base LLM"], "varies": ["preference data construction"], "benchmarks": [], "results": [], "workflow": ["Generate positive response", "Build contrastive pair", "Preference optimize"], "summary": "MIPO builds preference pairs by maximizing prompt-response mutual information to improve LLMs without extra data."}
|
|
{"id": 4812, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Agentless training recipe as skill prior for SWE-agent adaptation"}], "pattern": "fix-env-vary-agent", "fixes": ["SWE-bench Verified"], "varies": ["training recipe"], "benchmarks": [{"name": "SWE-bench Verified", "purpose": "evaluate SWE LLM", "is_new": false}], "results": [{"benchmark": "SWE-bench Verified", "system": "Kimi-Dev", "base_model": null, "setting": "proposed", "metric": "resolve rate", "score": 60.4, "unit": "%"}], "workflow": ["Agentless training", "Skill priors", "SWE-agent adaptation"], "summary": "Kimi-Dev, an open-source SWE LLM from Agentless training, adapts efficiently to SWE-agent frameworks."}
|
|
{"id": 13000, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "benchmark", "why": "Benchmark for implicit memory in LLMs"}], "pattern": "fix-agent-vary-env", "fixes": ["evaluated models"], "varies": ["memory constructs"], "benchmarks": [{"name": "ImplicitMemBench", "purpose": "implicit memory: procedural, priming", "is_new": true}], "results": [], "workflow": [], "summary": "ImplicitMemBench evaluates unconscious behavioral adaptation via procedural memory, priming and related constructs."}
|
|
{"id": 15323, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Hypergraph-guided VLM framework for procedural planning"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["structured semantic guidance"], "benchmarks": [], "results": [], "workflow": ["Mine hypergraph", "Guide VLM", "Plan procedure"], "summary": "GaLa uses hypergraphs mined from multimodal inputs to guide VLMs on procedural planning in embodied settings."}
|
|
{"id": 7173, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Benchmark for agentic workflows in Triton kernel generation"}], "pattern": "fix-agent-vary-env", "fixes": ["models under test"], "varies": ["workflows", "tasks"], "benchmarks": [{"name": "TritonGym", "purpose": "agentic Triton code generation", "is_new": true}], "results": [], "workflow": [], "summary": "TritonGym benchmarks agentic tool-using workflows for Triton GPU kernel generation with standardized function-call tools."}
|
|
{"id": 5576, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Subtask-level structurally aligned memory for SWE agents"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Memory granularity at subtask level"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory granularity"], "benchmarks": [], "results": [], "workflow": ["Segment into subtasks", "Store memory", "Retrieve per stage"], "summary": "Replaces instance-level memory with subtask-level structurally aligned memory for SWE agents."}
|
|
{"id": 11370, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "VLM language reflections as guidance for RL in manipulation"}], "pattern": "fix-env-vary-agent", "fixes": ["manipulation tasks"], "varies": ["language feedback for RL"], "benchmarks": [], "results": [], "workflow": ["Summarize attempt", "Localize decisive moments", "Guide RL"], "summary": "LAGEA turns VLM reflections on episodes into temporally grounded guidance for RL in robotic manipulation."}
|
|
{"id": 13703, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Contract-checked graph editing for verifier-guided inference-time search"}], "pattern": "vary-scaffold", "fixes": ["LLM", "verifier"], "varies": ["search procedure"], "benchmarks": [], "results": [], "workflow": ["Build reasoning DAG", "Gate edits", "Verify"], "summary": "Represents candidates as typed reasoning DAGs with deterministic structural gates before verification during inference-time search."}
|
|
{"id": 15375, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Context management as callable tool for SWE agents"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Structured context maintenance"}], "pattern": "vary-scaffold", "fixes": ["SWE tasks"], "varies": ["context management"], "benchmarks": [], "results": [], "workflow": ["Agent calls context tool", "Condense context", "Continue"], "summary": "Cat elevates context maintenance to a callable tool in long-horizon SWE agents."}
|
|
{"id": 12971, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Automated design of MAS with heterogeneous LLMs"}], "pattern": "vary-scaffold", "fixes": ["tasks"], "varies": ["topology and LLM assignment"], "benchmarks": [], "results": [], "workflow": [], "summary": "Hetero-Designer automatically designs multi-agent topologies assigning heterogeneous LLMs."}
|
|
{"id": 12319, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Prompting MLLMs to produce textual allocentric 3D representations"}], "pattern": "vary-scaffold", "fixes": ["MLLM"], "varies": ["prompting"], "benchmarks": [], "results": [], "workflow": ["Generate text spatial map", "Reason over it"], "summary": "TRACE prompts MLLMs to build text-based allocentric representations from egocentric video to improve 3D spatial reasoning."}
|
|
{"id": 6099, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "method", "why": "Query circuits tracing information flow for specific inputs"}], "pattern": "n/a", "fixes": [], "varies": ["explanation method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Introduces query circuits that trace in-model information flow explaining why a model answers a specific query."}
|
|
{"id": 3505, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Spatial-prior training for dual-system VLA"}], "pattern": "fix-env-vary-agent", "fixes": ["robot tasks"], "varies": ["spatial grounding pretraining"], "benchmarks": [], "results": [], "workflow": ["Spatial grounding pretraining", "Action alignment"], "summary": "SP-VLA aligns action learning with spatial priors via grounding pretraining and a dual-system architecture."}
|
|
{"id": 1098, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "benchmark", "why": "ParaS2SBench for paralinguistic S2S"}, {"area": "align/pref", "role": "method", "why": "RL on waveform-level content and style"}], "pattern": "vary-both", "fixes": ["spoken LMs"], "varies": ["RL reward"], "benchmarks": [{"name": "ParaS2SBench", "purpose": "paralinguistic-aware S2S evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "ParaS2S benchmarks and aligns speech-to-speech models for paralinguistic awareness with waveform-level RL."}
|
|
{"id": 12977, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Workflow fusion for automated agentic workflow generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow search"], "benchmarks": [], "results": [], "workflow": ["Generate workflows", "Fuse", "Evaluate"], "summary": "FusionFlow fuses independent workflows for structural leaps in automated agentic workflow generation."}
|
|
{"id": 11413, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Physics-guided diffusion to predict protein unfolding trajectories"}], "pattern": "n/a", "fixes": [], "varies": ["generative model"], "benchmarks": [], "results": [], "workflow": [], "summary": "GenUnfold is a physics-guided diffusion model predicting protein mechanical unfolding trajectories."}
|
|
{"id": 10403, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Agentic HTML source editing for rich document generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": ["Edit HTML source", "Render", "Review", "Refine"], "summary": "FormAct uses editing and review agents with tools to generate rich-format documents from scratch."}
|
|
{"id": 9584, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Causal-attribution guardrail against indirect prompt injection"}], "pattern": "fix-env-vary-agent", "fixes": ["IPI threat"], "varies": ["guardrail"], "benchmarks": [], "results": [], "workflow": ["Attribute causal influence", "Detect shift", "Sanitize"], "summary": "CausalArmor detects dominance shifts via causal attribution to trigger sanitization only on likely injections."}
|
|
{"id": 7493, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "LLM agents de-anonymize identities from weak cues"}, {"area": "agent-safety", "role": "benchmark", "why": "InferLink benchmark"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["fingerprint type", "task framing", "attacker knowledge"], "benchmarks": [{"name": "InferLink", "purpose": "de-anonymization risk", "is_new": true}], "results": [], "workflow": [], "summary": "Shows LLM agents can re-identify people from scattered cues, measured with the InferLink benchmark."}
|
|
{"id": 12369, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "User-controllable memory usage dimension"}, {"area": "social-sim", "role": "scaffold", "why": "Long-term personalization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory reliance"], "benchmarks": [], "results": [], "workflow": [], "summary": "Models memory reliance as controllable dimension to balance anchoring and innovation in long-term interaction."}
|
|
{"id": 10616, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Momentum-driven evolution with compact delta context"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Evolutionary context"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["context construction"], "benchmarks": [], "results": [], "workflow": [], "summary": "DeltaEvolve replaces full-code histories with momentum-style deltas in LLM evolutionary discovery."}
|
|
{"id": 14523, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Self-evolving service dialogue agent with simulated users"}, {"area": "domain-agents", "role": "application", "why": "Service dialogue"}], "pattern": "fix-env-vary-agent", "fixes": ["service dialogue"], "varies": ["training via simulated users"], "benchmarks": [], "results": [], "workflow": [], "summary": "SEAD trains service dialogue agents via self-evolution with simulated users without large human annotation."}
|
|
{"id": 13829, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "RL for privacy-aware collaboration between local and proprietary LLMs"}, {"area": "multi-agent", "role": "train-agent", "why": "RL multi-LLM collaboration"}], "pattern": "fix-env-vary-agent", "fixes": ["task"], "varies": ["RL policy"], "benchmarks": [], "results": [], "workflow": [], "summary": "Privacy-R1 trains via RL a collaboration between local and cloud LLMs preserving privacy."}
|
|
{"id": 15190, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Probabilistic blending for selective inference-time alignment"}], "pattern": "vary-scaffold", "fixes": ["base model"], "varies": ["intervention policy"], "benchmarks": [], "results": [], "workflow": [], "summary": "Decides when to intervene with guidance during decoding via probabilistic model blending."}
|
|
{"id": 16961, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent pipeline for video storytelling"}, {"area": "domain-agents", "role": "application", "why": "Long-sequence video generation"}], "pattern": "vary-scaffold", "fixes": ["generators"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "MAViS orchestrates specialized agents using Explore, Examine, Enhance for long video storytelling."}
|
|
{"id": 1451, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Collaborative-semantic tool indexing for generative tool use"}], "pattern": "fix-env-vary-agent", "fixes": ["tool benchmarks"], "varies": ["tool representation"], "benchmarks": [], "results": [], "workflow": [], "summary": "ToolWeaver encodes tools with collaborative semantics instead of unique tokens for scalable generative tool use."}
|
|
{"id": 5567, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Staged knowledge injection SFT then agentic RLVR for remote sensing"}, {"area": "domain-agents", "role": "application", "why": "Remote sensing understanding"}], "pattern": "fix-env-vary-agent", "fixes": ["zoom-in tools"], "varies": ["post-training stages"], "benchmarks": [], "results": [], "workflow": [], "summary": "Finds text knowledge injection before vision matters for agentic RLVR in ultra-high-resolution remote sensing."}
|
|
{"id": 6850, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Token dropping preserving salience and diversity in VLMs"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["token reduction"], "benchmarks": [], "results": [], "workflow": [], "summary": "SPLIT drops vision tokens with salience-guided partitioning for efficient VLMs."}
|
|
{"id": 14305, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Enterprise deep research with controlled info flow and evidence-aware termination"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["architecture"], "benchmarks": [], "results": [], "workflow": ["Outline with reflection", "Dependency-guided execution", "Evidence-based termination"], "summary": "Scalable enterprise deep research architecture with coverage objectives, localized context and evidence-based stopping."}
|
|
{"id": 4953, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Generative latent memory with trigger and weaver"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["memory mechanism"], "benchmarks": [], "results": [], "workflow": [], "summary": "MemGen weaves generative latent memory into reasoning for self-evolving agents."}
|
|
{"id": 4325, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Topic-guided findings generation for radiology reports"}], "pattern": "vary-data", "fixes": ["VLM"], "varies": ["report structure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shifts radiology report generation from narrative flow to topic-guided findings to improve visual grounding."}
|
|
{"id": 10683, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Audio function-calling benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["audio tasks"], "benchmarks": [{"name": "BFCL Audio", "purpose": "audio function calling, 6.2K tasks", "is_new": true}], "results": [], "workflow": [], "summary": "BFCL Audio benchmarks audio function calling with 6.2K expert-verified tasks."}
|
|
{"id": 16664, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent procedural graph extraction with refinement"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Multi-agent framework extracting procedural graphs with structural and logical refinement agents."}
|
|
{"id": 6248, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Expert-guided agent for GPU kernel generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["expert guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "EGG guides kernel-generation agents with domain optimization expertise."}
|
|
{"id": 11969, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Colluding agents manipulate beliefs with truthful fragments"}], "pattern": "fix-agent-vary-env", "fixes": ["victim agent"], "varies": ["adversarial narrative"], "benchmarks": [], "results": [], "workflow": [], "summary": "Generative Montage lets colluding agents steer victim beliefs using only truthful evidence fragments."}
|
|
{"id": 5660, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Federated learning for workflow-structured LLM agent collaboration"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["federated training"], "benchmarks": [], "results": [], "workflow": [], "summary": "FedWave federates training of collaborating LLM agents under workflow dependencies and data locality."}
|
|
{"id": 17435, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "Customer support benchmark for business-adherence"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["policies", "workflows"], "benchmarks": [], "results": [], "workflow": [], "summary": "Benchmarks customer support LLM agents on policy adherence, task dependencies and robustness."}
|
|
{"id": 12202, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Memory-augmented multi-agent feature generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent system"], "benchmarks": [], "results": [], "workflow": [], "summary": "Memory-augmented LLM multi-agent system for automated tabular feature generation."}
|
|
{"id": 8230, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Identifies Lazy Likelihood Displacement causing GRPO collapse"}, {"area": "search-research", "role": "train-agent", "why": "Mitigates collapse in search-agent RL"}], "pattern": "vary-conditions", "fixes": ["Search-R1 setting"], "varies": ["GRPO training"], "benchmarks": [{"name": "Search-R1", "purpose": "tool-integrated RL baseline", "is_new": false}], "results": [], "workflow": [], "summary": "Identifies Lazy Likelihood Displacement as the cause of GRPO collapse in tool-integrated search RL."}
|
|
{"id": 1891, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "method", "why": "VLM-guided physics-based human-object interaction"}], "pattern": "vary-scaffold", "fixes": ["VLM"], "varies": ["motion policy design"], "benchmarks": [], "results": [], "workflow": [], "summary": "VLM-guided relative movement dynamics enable physics-based long-horizon human-object interaction."}
|
|
{"id": 12599, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Fine-grained RL rewards for memory manager"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["reward"], "benchmarks": [], "results": [], "workflow": [], "summary": "Fine-Mem adds chunk-level step rewards for credit assignment in RL memory management."}
|
|
{"id": 3334, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "Parallel relative policy optimization for chart deep research"}], "pattern": "fix-env-vary-agent", "fixes": ["chart tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "Trains LVLMs for chart deep research with parallel relative policy optimization."}
|
|
{"id": 14778, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Candidate-aware retrieval and reranking for Arabic MCQA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval"], "benchmarks": [], "results": [], "workflow": [], "summary": "Candidate-aware retrieval and reranking RAG for Arabic multiple-choice QA."}
|
|
{"id": 12774, "in_agent_scope": false, "types": [{"area": "reason/math", "role": "method", "why": "Structure-aware iterative autoformalization of combinatorics in Lean 4"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["refinement"], "benchmarks": [], "results": [], "workflow": [], "summary": "SAIR-Comb iteratively refines Lean 4 autoformalization of combinatorics."}
|
|
{"id": 17322, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "method", "why": "Tool for parallel corpora augmentation and refinement"}], "pattern": "n/a", "fixes": [], "varies": ["data tool"], "benchmarks": [], "results": [], "workflow": [], "summary": "AlignFix refines and augments noisy parallel corpora for machine translation."}
|
|
{"id": 3359, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "analysis", "why": "Analysis of safety-reasoning trade-off in MLRMs"}, {"area": "align/pref", "role": "method", "why": "Policy-guided safety tuning"}], "pattern": "vary-conditions", "fixes": ["MLRMs"], "varies": ["safety tuning"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analyzes safety degradation in multimodal reasoning models and proposes policy-guided safety tuning."}
|
|
{"id": 16070, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Entropy-guided citation selection for medical RAG"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["citation selection"], "benchmarks": [], "results": [], "workflow": [], "summary": "Selects low-entropy citations to resist retrieval noise in medical RAG."}
|
|
{"id": 9062, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "HyperTrack dataset and GUIEvalKit"}, {"area": "web-gui", "role": "analysis", "why": "Data scaling and reasoning study"}], "pattern": "vary-both", "fixes": ["VLM agents"], "varies": ["data scale", "fine-tuning type"], "benchmarks": [{"name": "HyperTrack", "purpose": "16000 tasks, 650 Chinese apps", "is_new": true}, {"name": "GUIEvalKit", "purpose": "unified offline GUI benchmarking", "is_new": true}], "results": [], "workflow": [], "summary": "Studies data scaling and reasoning for mobile GUI VLM agents with the HyperTrack dataset."}
|