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
30 KiB
JSON
61 lines
30 KiB
JSON
{"id": 8272, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Long-horizon Red Alert RTS benchmark testing rule robustness, Elo strength, steerability of LLMs"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs under test"], "varies": ["RTS rules", "opponents", "language interventions"], "benchmarks": [{"name": "Red Alert RTS benchmark", "purpose": "diagnose long-horizon decision making", "is_new": true}], "results": [], "workflow": [], "summary": "Diagnostic long-horizon RTS benchmark with hierarchical macro-intent interface evaluating robustness, strength and human steerability of LLMs."}
|
|
{"id": 4061, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "RL-ZVP extracts learning signal from zero-variance prompts via entropy-guided advantage shaping"}], "pattern": "n/a", "fixes": ["RLVR setup", "benchmarks"], "varies": ["advantage shaping for zero-variance prompts"], "benchmarks": [], "results": [], "workflow": ["Detect zero-variance prompts", "Shape advantage by token entropy", "Policy update"], "summary": "RL-ZVP algorithm uses zero-variance prompts in GRPO-style RLVR with entropy-guided advantage shaping to improve reasoning."}
|
|
{"id": 7092, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "method", "why": "Dual-process neuro-symbolic framework combining LLM Socratic guidance with symbolic solvers for logical reasoning"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["reasoning framework"], "benchmarks": [], "results": [], "workflow": ["LLM translates problem", "Symbolic solver derives", "Socratic feedback refines"], "summary": "LogicSAGE combines LLM intuition and symbolic solvers with Socratic guidance for logical reasoning."}
|
|
{"id": 14001, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "End-to-end system generating editable SVG scientific illustrations from text with reference styling"}], "pattern": "n/a", "fixes": ["general LLM/image models"], "varies": ["figure generation pipeline"], "benchmarks": [], "results": [], "workflow": ["Understand long text", "Style from reference image", "Native SVG editing"], "summary": "AutoFigure-Edit system generates editable scientific illustrations from text with reference-guided styling."}
|
|
{"id": 16331, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "AutoPage multi-agent system builds project webpages from papers with human collaboration"}, {"area": "multi-agent", "role": "application", "why": "Hierarchical coarse-to-fine multi-agent pipeline for paper-to-page creation"}], "pattern": "n/a", "fixes": ["base LLMs"], "varies": ["paper-to-page pipeline"], "benchmarks": [], "results": [], "workflow": ["Narrative planning", "Content generation", "Page assembly with human checkpoints"], "summary": "AutoPage is a human-in-the-loop hierarchical multi-agent system turning papers into interactive project webpages."}
|
|
{"id": 18281, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "method", "why": "Translation-guided ASR with parallel gated cross-attention for Taiwanese Hokkien"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["ASR architecture"], "benchmarks": [], "results": [], "workflow": ["Extract translation embeddings", "Gated cross-attention fusion", "Recognize speech"], "summary": "TG-ASR integrates multilingual translation embeddings via gated cross-attention for low-resource Taiwanese Hokkien speech recognition."}
|
|
{"id": 17892, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "benchmark", "why": "ObfusQA obfuscated factual QA with three obfuscation levels testing LLM robustness"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs under test"], "varies": ["question obfuscation type"], "benchmarks": [{"name": "ObfusQA", "purpose": "robustness to obfuscated factual questions", "is_new": true}], "results": [], "workflow": [], "summary": "ObfusQAte framework and ObfusQA dataset test LLM robustness on obfuscated factual questions."}
|
|
{"id": 16402, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "data", "why": "LLM-based context and reference-guided source-side augmentation for movie subtitle translation"}], "pattern": "vary-data", "fixes": ["translation fine-tuning recipe"], "varies": ["augmented training data"], "benchmarks": [], "results": [], "workflow": ["Extract context info", "Rewrite subtitles with LLM", "Fine-tune translator"], "summary": "Data augmentation rewrites source subtitles using context and references to improve movie-specific translation with few bilingual pairs."}
|
|
{"id": 14875, "in_agent_scope": false, "types": [{"area": "safety/privacy", "role": "method", "why": "Mechanistic circuit-guided difficulty metric for unlearning"}], "pattern": "n/a", "fixes": [], "varies": ["unlearning difficulty metric"], "benchmarks": [], "results": [], "workflow": [], "summary": "Mechanistic perspective and circuit-guided metric for measuring difficulty of unlearning (abstract missing)."}
|
|
{"id": 3458, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Inference-time logit arithmetic transferring small RFT reasoner behavior to large LVLMs without training"}], "pattern": "n/a", "fixes": ["large model weights"], "varies": ["decoding distribution"], "benchmarks": [], "results": [], "workflow": ["Compute small reasoner and base outputs", "Subtract distributions", "Guide large model decoding"], "summary": "ProxyThinker subtracts base from RFT small-model output distributions to elicit slow-thinking visual reasoning in large LVLMs."}
|
|
{"id": 17911, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "ESG-QA dataset of 87k QA-context triplets benchmarking open LLMs on ESG"}], "pattern": "fix-agent-vary-env", "fixes": ["Llama-3, Gemma-3, Qwen-3"], "varies": ["ESG pillars", "RAG vs abstractive"], "benchmarks": [{"name": "ESG-QA", "purpose": "ESG question answering evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "ESG-QA is a large LLM-generated, filtered QA dataset for ESG, benchmarking three open LLM families."}
|
|
{"id": 15586, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Test-time rubric-guided iterative verification lets deep research agents self-evolve without training"}], "pattern": "vary-scaffold", "fixes": ["policy model"], "varies": ["verification and refinement procedure"], "benchmarks": [], "results": [], "workflow": ["Generate answer", "Verify with failure-taxonomy rubrics", "Refine iteratively"], "summary": "Inference-time scaling of verification: rubric-guided iterative self-verification improves deep research agents without training."}
|
|
{"id": 9557, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "method", "why": "Parallel decoding with special tokens for low-latency function calling"}], "pattern": "n/a", "fixes": ["function-calling task"], "varies": ["decoding method"], "benchmarks": [], "results": [], "workflow": ["Compress structure with special tokens", "Decode arguments in parallel"], "summary": "RealtimeTool uses special tokens and parallel decoding to cut function-calling latency for real-time applications."}
|
|
{"id": 15674, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Theory-of-Mind facilitator infers goals via Bayesian inverse planning and gives feedback to promote prosocial actions"}], "pattern": "vary-scaffold", "fixes": ["agents"], "varies": ["facilitator feedback"], "benchmarks": [], "results": [], "workflow": ["Infer goals by inverse planning", "Select feedback maximizing utility", "Communicate to agents"], "summary": "ProToM facilitates multi-agent prosociality via ToM-informed feedback derived from Bayesian goal inference."}
|
|
{"id": 417, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "method", "why": "Textual equilibrium propagation addresses exploding and vanishing textual gradients in deep compound AI optimization"}], "pattern": "n/a", "fixes": ["LLM modules"], "varies": ["feedback propagation scheme"], "benchmarks": [], "results": [], "workflow": ["Local textual feedback", "Equilibrium-style propagation", "Update prompts"], "summary": "Textual Equilibrium Propagation optimizes deep compound AI systems while avoiding exploding and vanishing textual gradients."}
|
|
{"id": 14991, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Trains agentic RAG with reflection tagging to self-evaluate retrieved content and optimize strategy"}], "pattern": "fix-env-vary-agent", "fixes": ["RAG setting"], "varies": ["training with self-evaluation"], "benchmarks": [], "results": [], "workflow": ["Reflection tagging of retrieved content", "Policy learning from critique"], "summary": "Reflective RAG trains agentic RAG models to critique retrieved content and use it to guide later retrieval and generation."}
|
|
{"id": 13969, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Parallel exploration text-to-SQL agent using atomic test-case SQLs for coverage"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["exploration strategy"], "benchmarks": [{"name": "Spider 2.0", "purpose": "text-to-SQL evaluation", "is_new": false}], "results": [], "workflow": ["Generate atomic test SQLs", "Execute in parallel", "Final SQL when covered"], "summary": "PExA reframes text-to-SQL as test coverage with parallel atomic SQL exploration, evaluated on Spider 2.0."}
|
|
{"id": 14720, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Hierarchical multi-agent system with manager agents for paper-to-code reproduction"}, {"area": "code-swe", "role": "eval-method", "why": "Introduces refined Paper2Code-Extra evaluation protocol"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["agent coordination structure"], "benchmarks": [{"name": "Paper2Code", "purpose": "paper reproduction evaluation", "is_new": false}, {"name": "Paper2Code-Extra", "purpose": "refined evaluation", "is_new": true}], "results": [], "workflow": ["Manager agents coordinate", "Specialised agents per stage", "Execute code"], "summary": "HiRAS is a hierarchical multi-agent framework for paper reproduction plus refined P2C-Ex evaluation."}
|
|
{"id": 9832, "in_agent_scope": false, "types": [{"area": "train/sft", "role": "theory", "why": "Finite-sample guarantees for iterative self-improvement with easy-to-hard curricula"}], "pattern": "n/a", "fixes": [], "varies": ["curriculum", "assumptions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Theory of iterative self-improvement as MLE on reward-filtered data with finite-sample guarantees and curriculum benefits."}
|
|
{"id": 1912, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "data", "why": "Selects data by human uncertainty and auto-labels for VLM SFT in VQA"}, {"area": "arch/multimodal", "role": "analysis", "why": "Evaluates VLMs across human uncertainty levels"}], "pattern": "vary-data", "fixes": ["SFT recipe"], "varies": ["selected data", "HU levels"], "benchmarks": [], "results": [], "workflow": ["Assess human uncertainty", "Select data", "Auto-label and fine-tune"], "summary": "Analyzes human uncertainty effect on VLM SFT and proposes uncertainty-aware data selection and auto-labeling."}
|
|
{"id": 1822, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "SPEAR self-imitation RL with progressive exploration for agentic LLMs"}], "pattern": "fix-env-vary-agent", "fixes": ["agent tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Curriculum exploration", "Self-imitation replay of successes", "Policy update"], "summary": "SPEAR is a self-imitation RL recipe balancing exploration and exploitation for long-horizon tool-use LLM agents."}
|
|
{"id": 7890, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Distills group-relative semantic advantages into language experience prior injected in context for video anomaly reasoning"}], "pattern": "n/a", "fixes": ["MLLM weights"], "varies": ["experience prior"], "benchmarks": [], "results": [], "workflow": ["Sample reasoning trajectories", "Distill linguistic advantages", "Inject prior in context"], "summary": "LRPO adapts MLLMs for video anomaly reasoning via linguistic experience priors without parameter updates."}
|
|
{"id": 13463, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Guided Topology Diffusion generates task-adaptive communication topologies with graph diffusion"}], "pattern": "n/a", "fixes": ["agent LLMs"], "varies": ["communication topology"], "benchmarks": [], "results": [], "workflow": ["Condition on task", "Diffuse graph topology", "Run agents"], "summary": "GTD dynamically generates multi-agent communication topologies via conditional graph diffusion balancing performance, cost and robustness."}
|
|
{"id": 2357, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Training-free framework where LLM guidance improves small model contextual reasoning"}], "pattern": "n/a", "fixes": ["SLM weights"], "varies": ["guidance procedure"], "benchmarks": [], "results": [], "workflow": ["LLM produces navigation", "SLM reasons following it"], "summary": "Navigation is a training-free approach using LLM guidance to enhance small language model contextual reasoning."}
|
|
{"id": 9520, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Self-supervised RL trains MLLMs for online mental-state inference without annotations"}], "pattern": "fix-env-vary-agent", "fixes": ["ToM tasks"], "varies": ["training reward"], "benchmarks": [], "results": [], "workflow": ["Generate mental state hypotheses", "Reward via behavior likelihood", "RL update"], "summary": "MindZero trains MLLMs with self-supervised RL for efficient online Theory-of-Mind reasoning without annotations."}
|
|
{"id": 16262, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "Social Welfare Function benchmark simulating LLM dictator allocating resources, efficiency vs fairness"}], "pattern": "fix-agent-vary-env", "fixes": ["20 LLMs"], "varies": ["allocation scenarios"], "benchmarks": [{"name": "SWF Benchmark", "purpose": "efficiency versus fairness in LLM allocation", "is_new": true}], "results": [], "workflow": [], "summary": "SWF benchmark simulates an LLM dictator distributing tasks, exposing efficiency-fairness trade-offs across 20 LLMs."}
|
|
{"id": 15353, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Graph-GRPO group relative policy optimization for topology learning"}], "pattern": "fix-env-vary-agent", "fixes": ["MAS tasks"], "varies": ["topology policy RL algorithm"], "benchmarks": [], "results": [], "workflow": ["Sample multiple topologies", "Group-relative advantage", "Update policy"], "summary": "Graph-GRPO stabilizes multi-agent topology learning with group-relative advantages to reduce variance and credit assignment issues."}
|
|
{"id": 12036, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "AgencyBench: 138 long-horizon real-world tasks over 6 capabilities with rubric evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["scenarios", "capabilities"], "benchmarks": [{"name": "AgencyBench", "purpose": "long-horizon real-world agent evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "AgencyBench is a benchmark of 32 scenarios and 138 tasks needing 1M-token contexts for autonomous agents."}
|
|
{"id": 17199, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "Meta-agent generates adaptive adversarial persona tests and scores with LLM judge"}], "pattern": "vary-evaluation", "fixes": ["agents under test"], "varies": ["test generation protocol"], "benchmarks": [], "results": [], "workflow": ["Analyze code and literature", "Generate persona tests", "Judge and adapt"], "summary": "Agent-Testing Agent automatically tests conversational agents, surfacing diverse failures faster than human annotators."}
|
|
{"id": 14298, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent tool-augmented grounded Islamic QA system"}], "pattern": "n/a", "fixes": ["general LLMs"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": ["Route query", "Retrieve scripture or compute", "Grounded answer"], "summary": "Fanar-Sadiq is a bilingual multi-agent, tool-augmented Islamic QA system grounded in Quran and Hadith."}
|
|
{"id": 13060, "in_agent_scope": true, "types": [{"area": "search-research", "role": "method", "why": "TECQA builds temporal evidence chains for LLM temporal KGQA"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["evidence retrieval"], "benchmarks": [], "results": [], "workflow": ["Construct evidence chain", "LLM reasons over chain"], "summary": "TECQA constructs reliable temporal evidence chains for LLM-based temporal knowledge graph QA."}
|
|
{"id": 14574, "in_agent_scope": false, "types": [{"area": "knowledge/editing", "role": "method", "why": "Queue-based self-correction for sequential model editing"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["editing method"], "benchmarks": [], "results": [], "workflow": ["Queue edits", "Self-correct structure"], "summary": "QueueEDIT improves sequential model editing while preserving general capabilities of LLMs."}
|
|
{"id": 12651, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Co-evolves capability expansion and experience distillation in frozen-LLM agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["self-evolution mechanism"], "benchmarks": [], "results": [], "workflow": ["Distill experience", "Create tools or agents", "Co-evolve"], "summary": "Mem^2Evolve jointly evolves agent assets and experience for self-evolving agents."}
|
|
{"id": 4507, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Multi-turn RL optimizing self-verification with tool-based evaluation for code"}], "pattern": "fix-env-vary-agent", "fixes": ["code tasks"], "varies": ["RL with verification"], "benchmarks": [], "results": [], "workflow": ["Generate code", "Self-verify with tools", "Multi-turn RL"], "summary": "ReVeal trains code agents with multi-turn RL that explicitly optimizes self-verification using tool feedback."}
|
|
{"id": 13307, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "SOAP-guided multi-view text modeling for ESI prediction"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["modeling framework"], "benchmarks": [], "results": [], "workflow": ["SOAP-structured views", "Predict ESI"], "summary": "SOAPTriage models triage text through SOAP views to predict Emergency Severity Index."}
|
|
{"id": 14398, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Spectral selection and uncertainty fusion for multimodal rumor detection"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["detection method"], "benchmarks": [], "results": [], "workflow": ["DCT features", "Entropy-guided fusion"], "summary": "Multimodal rumor detection with multi-scale spectral selection and entropy-guided uncertainty fusion."}
|
|
{"id": 16433, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "ToolQP iterative query planning for multi-step tool retrieval"}], "pattern": "vary-scaffold", "fixes": ["retriever"], "varies": ["retrieval procedure"], "benchmarks": [], "results": [], "workflow": ["Decompose instruction", "Generate sub-queries", "Interact with retriever"], "summary": "ToolQP frames tool retrieval as iterative query planning over sub-tasks."}
|
|
{"id": 3636, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "REMem two-phase episodic memory construction and reasoning for language agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory design"], "benchmarks": [], "results": [], "workflow": ["Offline episodic indexing", "Online reasoning over events"], "summary": "REMem builds and reasons over episodic memory for language agents."}
|
|
{"id": 3321, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Annotation-free agentic RL for medical VLM visual reasoning with entropy-guided regrounding"}], "pattern": "fix-env-vary-agent", "fixes": ["medical tasks"], "varies": ["RL method"], "benchmarks": [], "results": [], "workflow": ["Entropy-guided regrounding", "Consensus credit assignment", "RL"], "summary": "MedVR trains medical VLMs with annotation-free agentic RL for visually grounded reasoning."}
|
|
{"id": 17016, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "analysis", "why": "Studies cross-language variation in LLM answers to value-laden MCQs"}, {"area": "safety/fairness", "role": "benchmark", "why": "Releases MEVS multilingual value survey corpus"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["question language"], "benchmarks": [{"name": "MEVS", "purpose": "multilingual value-laden MCQ", "is_new": true}], "results": [], "workflow": [], "summary": "Releases multilingual European Value Survey corpus and studies language-induced variation in LLM values."}
|
|
{"id": 4975, "in_agent_scope": false, "types": [{"area": "efficiency/compress", "role": "method", "why": "KLT-guided SVD with bias-corrected vector quantization for MoE LLMs"}], "pattern": "n/a", "fixes": ["MoE models"], "varies": ["quantization method"], "benchmarks": [], "results": [], "workflow": ["KLT-guided SVD", "Bias-corrected VQ"], "summary": "KBVQ-MoE enables ultra-low-bit vector quantization of MoE LLMs."}
|
|
{"id": 14658, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Concept-guided expert routing for VQA"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["routing method"], "benchmarks": [], "results": [], "workflow": ["Concept-guided routing", "Reweight experts by options"], "summary": "CoGR-MoE stabilizes and diversifies MoE routing in VQA using answer-option semantics."}
|
|
{"id": 12554, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "scaffold", "why": "Fine-grained beam search with process reward model for function calling inference scaling"}, {"area": "tool-use", "role": "data", "why": "Builds first intra-call supervision dataset"}], "pattern": "vary-scaffold", "fixes": ["function-calling LLM"], "varies": ["inference search", "reward model"], "benchmarks": [], "results": [], "workflow": ["Mask functions and roll out", "Annotate steps", "Beam search with PRM"], "summary": "ToolPRM scores intra-call decisions to scale inference for structured function calling."}
|
|
{"id": 3267, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Transformer with external layer memory for long-horizon partially observable RL"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["Per-layer memory", "Update/rewrite"], "summary": "ELMUR adds structured external layer memory to transformers for long-horizon robotic RL."}
|
|
{"id": 8645, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "CausalGame interactive games benchmark of causal thinking with confounders and noise"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["game scenarios"], "benchmarks": [{"name": "CausalGame", "purpose": "agent causal thinking", "is_new": true}], "results": [], "workflow": [], "summary": "CausalGame benchmarks causal thinking of LLM agents via interactive games with hidden confounders."}
|
|
{"id": 5189, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Interact-RAG gives agents corpus interaction primitives beyond black-box retrieval"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval interface"], "benchmarks": [], "results": [], "workflow": ["Corpus Interaction Engine", "Fine-grained actions"], "summary": "Interact-RAG lets agents actively manipulate retrieval via action primitives."}
|
|
{"id": 14502, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Lifelong agent builds and refines user profiles from dialogue for personalization"}, {"area": "social-sim", "role": "method", "why": "Proactive personalization with cold-start handling"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["profile inference"], "benchmarks": [], "results": [], "workflow": ["Decompose dialogue", "Infer preferences", "Refine profile"], "summary": "PersonalAgent continually infers user preferences to maintain a profile for personalization."}
|
|
{"id": 13775, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "method", "why": "Frames image editing as decompose-then-act in structured interactive environment"}], "pattern": "n/a", "fixes": ["editing models"], "varies": ["editing paradigm"], "benchmarks": [], "results": [], "workflow": ["Decompose", "Act in structured environment"], "summary": "I2E recasts text-guided image editing as actions in a structured environment."}
|
|
{"id": 17368, "in_agent_scope": false, "types": [{"area": "arch/tokenization", "role": "analysis", "why": "Systematic comparison of MWE-aware tokenizers across four languages"}], "pattern": "vary-conditions", "fixes": [], "varies": ["tokenizer architecture", "languages"], "benchmarks": [], "results": [], "workflow": [], "summary": "Compares UDPipe, Elephant and MaChAmp for multiword-aware tokenization in English, German, Italian, Dutch."}
|
|
{"id": 8117, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Hierarchical plan-execute RL for multi-turn LLM agents"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["policy hierarchy"], "benchmarks": [], "results": [], "workflow": ["Plan subgoals", "Execute", "Hierarchical RL"], "summary": "HiPER uses hierarchical plan-execute RL to improve long-horizon sparse-reward LLM agents."}
|
|
{"id": 12842, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Agentic oversight via dialectic debate with weak-to-strong judging"}], "pattern": "n/a", "fixes": ["models"], "varies": ["oversight protocol"], "benchmarks": [], "results": [], "workflow": ["Two experts defend answers", "Blind judge decides"], "summary": "Agentic Oversight uses dialectic argumentation between experts and a blind judge for scalable oversight."}
|
|
{"id": 395, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "application", "why": "Agentic framework for secure FHE code generation"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": ["Prompt formalizer", "Retrieval", "Secure code generation"], "summary": "FHE-Coder is an agentic framework generating secure FHE code from natural language."}
|
|
{"id": 5287, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "SFT warm-up then RL with decoupled entropy constraints for tool-use scaling"}], "pattern": "fix-env-vary-agent", "fixes": ["tool tasks"], "varies": ["training paradigm"], "benchmarks": [], "results": [], "workflow": ["Warm-up SFT", "RL with entropy constraints"], "summary": "AutoTool scales tool-use reasoning via SFT and RL with decoupled entropy constraints, reducing overthinking."}
|
|
{"id": 11381, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Cross-trajectory exploration conditioned on compressed state of prior trajectories"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["search procedure"], "benchmarks": [], "results": [], "workflow": ["Run trajectory", "Summarize state", "Condition next"], "summary": "Re-TRAC recursively compresses trajectories for globally informed deep search."}
|
|
{"id": 16033, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Two-level hierarchical LLM-guided memory for conversational agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory retrieval"], "benchmarks": [], "results": [], "workflow": ["Event-turn hierarchy", "LLM-guided retrieval"], "summary": "HiGMem retrieves compact evidence using hierarchical LLM-guided memory."}
|
|
{"id": 11690, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Heterogeneous medical multi-agent framework with specialists and clinician intervention"}], "pattern": "n/a", "fixes": ["generalist LLMs"], "varies": ["agent orchestration"], "benchmarks": [], "results": [], "workflow": ["Fuse evidence", "Trigger clinician by uncertainty"], "summary": "HetMedAgent orchestrates generalists, specialists and clinicians for medical AI."}
|
|
{"id": 5908, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Optimizes context compression of observations and history for agents without weight updates"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["compression guidelines"], "benchmarks": [], "results": [], "workflow": ["Compress observations and history", "Optimize guidelines"], "summary": "ACON compresses agent context via optimized guidelines for long-horizon agents."}
|
|
{"id": 10204, "in_agent_scope": false, "types": [{"area": "safety/jailbreak", "role": "defense", "why": "Online self-play MARL co-evolving attacker and defender for safety alignment"}], "pattern": "n/a", "fixes": ["reward model"], "varies": ["self-play training"], "benchmarks": [], "results": [], "workflow": ["Single policy attacks", "Same policy defends", "Reward adjudicates"], "summary": "Self-RedTeam co-evolves attacker and defender in online self-play for safer LLMs."}
|
|
{"id": 7730, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Task-agnostic plugin memory graph module for LLM agents"}], "pattern": "vary-scaffold", "fixes": ["LLM agent"], "varies": ["memory module"], "benchmarks": [], "results": [], "workflow": ["Structure episodic memories", "Build knowledge graph", "Retrieve"], "summary": "PlugMem is a plug-in knowledge-centric memory graph attachable to any LLM agent."}
|
|
{"id": 9392, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "theory", "why": "Derives non-vacuous trust region bounds for long-horizon off-policy mismatch"}, {"area": "align/rlvr", "role": "method", "why": "Trust region masking algorithm"}], "pattern": "n/a", "fixes": [], "varies": ["trust region bound", "masking"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows classical trust region bounds scale O(T^2) and derives tighter bounds with masking for LLM RL."}
|