Files
ai-taxonomy-literature-anal…/data/labels/o21.jsonl
T
zzyang d6fc7c1dfa Add types-of-work classification, benchmarks and score pages
- Work type = area x role (fixed vs varied), vocabulary in scripts/worktypes.json
- LLM-read labels for 2,678 agent-related papers (data/labels), rule labels for the rest
- Per-type synthesis notes (why-same, sub-variants, benchmarks, workflow)
- Pages: /work-types matrix and detail (timeline, benchmarks, workflow, score charts), /benchmarks
- build_worktypes.py adds work-type tables to data/papers.db

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Bhs9LU8eMixk6g1kb1d3j7
2026-10-21 11:55:46 +00:00

61 lines
29 KiB
JSON

{"id": 12580, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Progressive multimodal search and reasoning trajectory for knowledge-intensive VQA"}], "pattern": "vary-scaffold", "fixes": ["base model"], "varies": ["retrieval and reasoning procedure"], "benchmarks": [], "results": [], "workflow": ["Dual-scope query retrieval", "Synthesize evidence into records", "Revise reasoning trajectory"], "summary": "PMSR progressively builds a reasoning trajectory with dual-scope retrieval and evidence synthesis for knowledge-intensive VQA."}
{"id": 1870, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Test-time updating of routing over world models for embodied agents"}], "pattern": "vary-scaffold", "fixes": ["LM weights", "environments"], "varies": ["test-time routing over world models"], "benchmarks": [], "results": [], "workflow": ["Route over world models", "Update routing at test time", "Recombine/expand models"], "summary": "TMoW updates routing over world models at test time so embodied agents adapt to unseen, evolving domains."}
{"id": 1614, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Entropy-guided adaptive parameter sharing across LoRA layers"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["LoRA sharing method"], "benchmarks": [], "results": [], "workflow": [], "summary": "E2LoRA adaptively shares LoRA parameters across layers guided by gradient-based entropy to improve PEFT efficiency."}
{"id": 12065, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "analysis", "why": "Evaluates whether LLMs work as text-based world models across fidelity, scaling, agent utility"}, {"area": "embodied-nav", "role": "environment", "why": "LLM world models as learned simulators for agents"}], "pattern": "vary-conditions", "fixes": ["methods"], "varies": ["data and model scale", "environments"], "benchmarks": [], "results": [], "workflow": [], "summary": "Investigates LLMs as implicit text-based world models with a three-level evaluation framework and agent utility tests."}
{"id": 844, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Human-demonstration desktop grounding dataset GroundCUA with 3.56M annotations"}, {"area": "web-gui", "role": "train-agent", "why": "GroundNext grounding models at 3B and 7B"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data"], "benchmarks": [{"name": "GroundCUA", "purpose": "new desktop grounding dataset", "is_new": true}], "results": [], "workflow": ["Annotate demonstration screenshots", "Generate instructions", "Train GroundNext models"], "summary": "GroundCUA is a large desktop grounding dataset from human demonstrations used to train GroundNext grounding models."}
{"id": 13437, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Multi-agent graph RAG with hierarchical legal graph and verification"}], "pattern": "n/a", "fixes": ["base LLM"], "varies": ["RAG pipeline"], "benchmarks": [], "results": [], "workflow": ["Build hierarchical legal graph", "Retrieve evidence", "Multi-agent verify reasoning"], "summary": "LegalGraphRAG uses a hierarchical legal graph and multi-agent verification for reliable legal reasoning."}
{"id": 15751, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Graph-based memory with filtering, provenance and query-adaptive retrieval for conversational agents"}], "pattern": "vary-scaffold", "fixes": ["base LLM"], "varies": ["memory system"], "benchmarks": [{"name": "LOCOMO", "purpose": "evaluate memory retrieval", "is_new": false}, {"name": "LongMemEval", "purpose": "evaluate memory retrieval", "is_new": false}], "results": [], "workflow": ["Filter and compress memories", "Build provenance graph", "Query-adaptive subgraph retrieval"], "summary": "MemORAI is a graph memory framework for conversational agents with selective filtering and query-adaptive PageRank retrieval."}
{"id": 5206, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "analysis", "why": "Shows LLM agents have latent source preferences when selecting information"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["source attribution"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies latent source preferences of LLM agents when filtering and recommending retrieved information."}
{"id": 2259, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "SE(3)-equivariant diffusion transformer for long protein dynamics"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["generative model"], "benchmarks": [{"name": "ATLAS", "purpose": "evaluate protein dynamics generation", "is_new": false}], "results": [], "workflow": [], "summary": "STAR-MD is a causal spatiotemporal diffusion transformer generating microsecond protein trajectories."}
{"id": 2783, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "eval-method", "why": "Expert rubric benchmark ProfBench with cheap robust LLM judges"}, {"area": "eval/judge", "role": "eval-method", "why": "Judges mitigating self-enhancement bias"}], "pattern": "vary-evaluation", "fixes": ["models under test"], "varies": ["rubrics and judges"], "benchmarks": [{"name": "ProfBench", "purpose": "professional rubric benchmark", "is_new": true}], "results": [], "workflow": [], "summary": "ProfBench provides 7000+ expert response-criterion pairs and affordable LLM judges for professional-knowledge evaluation."}
{"id": 12426, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Agentic black-box graph extraction attack on GraphRAG under query budgets"}], "pattern": "vary-conditions", "fixes": ["GraphRAG systems"], "varies": ["attack strategy"], "benchmarks": [], "results": [], "workflow": ["Novelty-guided exploration", "Graph memory", "Discovery plus LLM filtering"], "summary": "AGEA is an agentic, query-efficient attack that steals hidden entity-relation graphs from GraphRAG systems."}
{"id": 12210, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "SVG font generation with structured linguistic supervision and efficient tokenization"}], "pattern": "n/a", "fixes": ["benchmark"], "varies": ["training supervision and tokenization"], "benchmarks": [], "results": [], "workflow": [], "summary": "Vector Calligrapher generates SVG fonts with structured linguistic descriptions and efficient representation in MLLMs."}
{"id": 13574, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Emotion-wheel-guided audio-referred text representation for multimodal ERC"}], "pattern": "n/a", "fixes": ["datasets"], "varies": ["fusion method"], "benchmarks": [], "results": [], "workflow": [], "summary": "EMART improves multimodal emotion recognition in conversation with audio-referred text fusion and emotion-wheel-aware loss."}
{"id": 14811, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-LLM agent framework for motivational interviewing dialogue generation"}, {"area": "domain-agents", "role": "data", "why": "6K simulated MI dialogue dataset"}], "pattern": "vary-scaffold", "fixes": ["base LLM"], "varies": ["multi-agent dialogue framework"], "benchmarks": [], "results": [], "workflow": ["Expand profiles into stories", "Interaction agent selects MI codes", "Generate dialogue"], "summary": "StoryMI uses therapist, client and interaction agents to generate controllable MI dialogues and a 6K dialogue dataset."}
{"id": 7754, "in_agent_scope": true, "types": [{"area": "science-data", "role": "benchmark", "why": "DDR-Bench for autonomous open-ended data investigation"}, {"area": "science-data", "role": "analysis", "why": "Analyzes scaffolding and scaling effects on investigatory intelligence"}], "pattern": "fix-agent-vary-env", "fixes": ["frontier LLM agents"], "varies": ["tasks"], "benchmarks": [{"name": "DDR-Bench", "purpose": "checklist-based deep data research", "is_new": true}], "results": [], "workflow": [], "summary": "Introduces Deep Data Research task and DDR-Bench evaluating autonomous insight extraction from databases."}
{"id": 15836, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "Concept diffusion with KG constraints to synthesize diverse, factual QA data"}], "pattern": "vary-data", "fixes": ["post-training recipe"], "varies": ["QA generation method"], "benchmarks": [], "results": [], "workflow": [], "summary": "DiFRa balances semantic diversity and factual consistency in QA generation using concept diffusion and knowledge graph constraints."}
{"id": 11949, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "RL training of persuasive utterances framed as Stackelberg game in social deduction games"}], "pattern": "fix-env-vary-agent", "fixes": ["SDG environments"], "varies": ["training objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "Formalizes SDG dialogue as Stackelberg competition and trains agents with RL to produce persuasive utterances."}
{"id": 14859, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "eval-method", "why": "Neuro-symbolic step-wise verification of CoT faithfulness"}], "pattern": "vary-evaluation", "fixes": ["models"], "varies": ["faithfulness check"], "benchmarks": [], "results": [], "workflow": ["Convert facts to logic", "Counterfactual attribution", "SMT and entailment check"], "summary": "NSF-CoT verifies chain-of-thought step faithfulness using logical conversion, counterfactual attribution and SMT plus LLM checking."}
{"id": 6992, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Visual-guided prompt evolution for training-free open-vocabulary segmentation"}], "pattern": "n/a", "fixes": ["dino.txt"], "varies": ["text query refinement"], "benchmarks": [], "results": [], "workflow": [], "summary": "VIP refines text queries in dino.txt with alias expansion and visual-guided distillation for training-free segmentation."}
{"id": 17272, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Induces reusable task instructions from few audio examples for AQA"}], "pattern": "n/a", "fixes": ["LALMs"], "varies": ["prompting method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Audio-Induct induces task-level instructions from few examples, outperforming few-shot prompting at lower inference cost."}
{"id": 4653, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "OSWorld-MCP benchmark for MCP tool invocation in computer-use agents"}, {"area": "tool-use", "role": "benchmark", "why": "158 validated tools across 7 applications"}], "pattern": "fix-agent-vary-env", "fixes": ["agents under test"], "varies": ["tools in OSWorld"], "benchmarks": [{"name": "OSWorld-MCP", "purpose": "tool invocation in computer use", "is_new": true}], "results": [], "workflow": [], "summary": "OSWorld-MCP benchmarks tool invocation alongside GUI operation for computer-use agents."}
{"id": 1731, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Conditional graph generation for adaptive communication topology"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["communication topology"], "benchmarks": [], "results": [], "workflow": [], "summary": "CARD generates environment-conditioned communication graphs for LLM multi-agent systems via a conditional variational graph encoder."}
{"id": 4582, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Shared identical KV cache across multiple task-specialized models"}], "pattern": "n/a", "fixes": ["serving workloads"], "varies": ["cache sharing architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "ICaRus lets multiple fine-tuned models share identical KV caches to cut memory and recomputation in multi-model inference."}
{"id": 12034, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "DeepPlanning long-horizon planning benchmark with verifiable constraints"}, {"area": "tool-use", "role": "benchmark", "why": "Travel planning and shopping tasks needing tool use"}], "pattern": "fix-agent-vary-env", "fixes": ["frontier LLMs"], "varies": ["tasks"], "benchmarks": [{"name": "DeepPlanning", "purpose": "long-horizon planning with constraints", "is_new": true}], "results": [], "workflow": [], "summary": "DeepPlanning benchmarks constrained long-horizon planning in multi-day travel and shopping tasks."}
{"id": 14750, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "environment", "why": "EnvScaler programmatic synthesis of tool-interaction environments"}, {"area": "tool-use", "role": "train-agent", "why": "SFT and RL on synthesized scenarios"}], "pattern": "fix-agent-vary-env", "fixes": ["agent training recipe"], "varies": ["synthesized environments"], "benchmarks": [], "results": [], "workflow": ["Build environment skeletons", "Generate scenarios and validators", "SFT and RL"], "summary": "EnvScaler synthesizes 191 tool-interaction environments and about 7K scenarios for SFT and RL of agents."}
{"id": 15381, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "application", "why": "Graph-based retrieval-generation agent with planning for long meeting audio QA"}, {"area": "domain-agents", "role": "benchmark", "why": "LongAudioQA dataset"}], "pattern": "vary-scaffold", "fixes": ["speech LLMs"], "varies": ["graph agent"], "benchmarks": [{"name": "LongAudioQA", "purpose": "long meeting audio QA", "is_new": true}], "results": [], "workflow": [], "summary": "GRGA builds a multi-dimensional audio graph and uses agent planning for retrieval and QA on long meetings."}
{"id": 18572, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "analysis", "why": "Module-level analysis of GraphRAG: triple extraction, clustering, report generation"}], "pattern": "vary-conditions", "fixes": ["GraphRAG framework"], "varies": ["module design choices"], "benchmarks": [], "results": [], "workflow": [], "summary": "Systematically analyzes GraphRAG module choices, finding triple extraction quality and community granularity matter."}
{"id": 6687, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Wavelet-domain transformers for operator learning of dynamical systems"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "MSWTs learn dynamics in a tokenized wavelet domain to mitigate spectral bias in neural operators."}
{"id": 12611, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "benchmark", "why": "TraceElephant failure attribution benchmark with full execution traces"}, {"area": "multi-agent", "role": "analysis", "why": "Studies attribution under full observability"}], "pattern": "vary-conditions", "fixes": ["attribution techniques"], "varies": ["trace observability"], "benchmarks": [{"name": "TraceElephant", "purpose": "failure attribution in MAS", "is_new": true}], "results": [], "workflow": [], "summary": "TraceElephant benchmarks failure attribution in LLM multi-agent systems with full traces and reproducible environments."}
{"id": 18200, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "application", "why": "Multi-agent critique-and-judge translation of interlinear glosses"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-agent critique procedure"], "benchmarks": [], "results": [], "workflow": ["Retrieve linguistic resources", "Generate candidates", "Critique and judge"], "summary": "GlossMATE translates IGT gloss lines to English via multi-agent critique-and-judge with linguistic resources and human-in-the-loop."}
{"id": 11605, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "3D world model disentangling ego-motion for dynamic 3D reconstruction forecasting"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["world model design"], "benchmarks": [], "results": [], "workflow": [], "summary": "FR3D predicts a persistent 3D latent for future dynamic reconstruction with ego-motion disentangled from scene dynamics."}
{"id": 10216, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Vision2Web hierarchical visual website development benchmark"}, {"area": "code-swe", "role": "eval-method", "why": "Agent verification with GUI agent verifier and VLM judge"}], "pattern": "fix-agent-vary-env", "fixes": ["coding agent frameworks"], "varies": ["tasks"], "benchmarks": [{"name": "Vision2Web", "purpose": "visual website development", "is_new": true}], "results": [], "workflow": [], "summary": "Vision2Web has 193 website tasks with workflow-based agent verification for coding agents."}
{"id": 14146, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "scaffold", "why": "Case-adaptive multi-agent specialist panel with voting and arbitration for clinical prediction"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["panel composition and routing"], "benchmarks": [], "results": [], "workflow": ["Attending agent assembles panel", "Specialists vote", "Hybrid router arbitrates"], "summary": "CAMP dynamically assembles specialist agent panels with three-valued voting and arbitration for clinical prediction."}
{"id": 12443, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "method", "why": "Sigmoid unembedding head for quality estimation under ambiguity"}], "pattern": "n/a", "fixes": ["pretrained LM"], "varies": ["quality estimation head"], "benchmarks": [], "results": [], "workflow": [], "summary": "Sigmoid Head adds a sigmoid unembedding head trained with heuristic negative sampling for better LM quality estimation."}
{"id": 15297, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Bi-directional rank reallocation guided by Hebbian importance"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["rank allocation"], "benchmarks": [], "results": [], "workflow": [], "summary": "HeBiRA adaptively prunes and expands LoRA ranks using Hebbian-inspired importance estimation."}
{"id": 13839, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent rebuttal generation as evidence-centric planning with external search"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent workflow"], "benchmarks": [], "results": [], "workflow": ["Decompose concerns", "Build hybrid context", "Plan response", "Draft rebuttal"], "summary": "RebuttalAgent decomposes reviews into concerns, builds contexts, searches literature and plans responses before drafting."}
{"id": 12793, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "benchmark", "why": "InsideOut benchmark of 4,000 prompts measuring insider-outsider cultural bias"}, {"area": "safety/fairness", "role": "defense", "why": "Mitigation of the bias"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["culture prompts"], "benchmarks": [{"name": "InsideOut", "purpose": "cultural insider-outsider bias", "is_new": true}], "results": [], "workflow": [], "summary": "Identifies insider-outsider bias in interview script generation with a 4,000-prompt benchmark and mitigations."}
{"id": 11487, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Reaction-center-guided atom ordering with discrete flow matching for retrosynthesis"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["generative model"], "benchmarks": [], "results": [], "workflow": [], "summary": "RetroDiT places reaction-center atoms first with discrete flow matching for template-free retrosynthesis."}
{"id": 13314, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "S2S-Arena speech-native paralinguistic instruction-following benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["S2S models"], "varies": ["tasks"], "benchmarks": [{"name": "S2S-Arena", "purpose": "paralinguistic S2S evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "S2S-Arena evaluates speech-to-speech models on paralinguistic instruction following with 1,243 samples and arena-style comparison."}
{"id": 10980, "in_agent_scope": false, "types": [{"area": "train/continual", "role": "method", "why": "Probe-guided expert expansion for replay-free continual learning in LVLMs"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["expansion method"], "benchmarks": [], "results": [], "workflow": [], "summary": "CoPE uses probe experts to decide where to add MoE experts, enabling replay-free continual multimodal learning."}
{"id": 1802, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Polymorphic skill abstraction for reusable web skills"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Skill induction from interaction"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["skill abstraction"], "benchmarks": [{"name": "Mind2Web", "purpose": "web agent success", "is_new": false}], "results": [], "workflow": [], "summary": "PolySkill decouples skill goals from implementations to learn generalizable web skills."}
{"id": 4655, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Sparse multi-agent debate with symbolic and natural-language reasoning"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["debate structure"], "benchmarks": [], "results": [], "workflow": [], "summary": "Multi-agent debate over symbolic and natural-language reasoning with adaptive sparse communication improves logical reasoning."}
{"id": 13149, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Retrieval-augmented multimodal multi-agent debate for stance detection"}], "pattern": "vary-scaffold", "fixes": ["base models"], "varies": ["agent pipeline"], "benchmarks": [], "results": [], "workflow": ["Retrieve context", "Multimodal analysis agents", "Debate", "Self-reflect adjudication"], "summary": "MM-StanceDet combines retrieval, multimodal analysis agents, debate and self-reflection for stance detection on five datasets."}
{"id": 2792, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "TDBench temporal database benchmark for time-sensitive QA"}, {"area": "eval/judge", "role": "eval-method", "why": "Time accuracy metric"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["benchmark and metric"], "benchmarks": [{"name": "TDBench", "purpose": "time-sensitive QA", "is_new": true}], "results": [], "workflow": [], "summary": "TDBench builds time-sensitive QA from temporal databases and adds a time accuracy metric."}
{"id": 2530, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Hierarchical group policy optimization fixing context inconsistency in stepwise advantage"}], "pattern": "fix-env-vary-agent", "fixes": ["agentic tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "HGPO groups steps by consistent history for unbiased stepwise advantage in long-horizon agentic RL."}
{"id": 13368, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "AndroidIntent benchmark for implicit intent alignment"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Hierarchical Intent Memory"}], "pattern": "vary-both", "fixes": ["GUI agents"], "varies": ["tasks", "memory"], "benchmarks": [{"name": "AndroidIntent", "purpose": "implicit intent GUI", "is_new": true}], "results": [], "workflow": [], "summary": "PersonalAlign defines personalized GUI agent task, AndroidIntent benchmark and hierarchical intent memory."}
{"id": 15808, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Agent-oriented sitemap WASP providing structure knowledge to web agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["site knowledge"], "benchmarks": [], "results": [], "workflow": [], "summary": "WASP and LightASM build dual-granularity sitemaps that improve web agent robustness and efficiency."}
{"id": 12581, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "LLM simulation of supply chain agents with varied cognitive heterogeneity"}], "pattern": "vary-conditions", "fixes": ["supply chain setup"], "varies": ["agent reasoning level"], "benchmarks": [], "results": [], "workflow": [], "summary": "Simulates multi-stage supply chains with DeepSeek and GPT agents to study cognitive heterogeneity effects."}
{"id": 10057, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Isolated planning defense against tool description poisoning"}], "pattern": "fix-agent-vary-env", "fixes": ["threat model"], "varies": ["defense mechanism"], "benchmarks": [], "results": [], "workflow": [], "summary": "Tool-Guard isolates misaligned tool invocations to defend LLM agents against cross-tool description poisoning."}
{"id": 14894, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Training ToM dialogue agent with lookahead"}, {"area": "social-sim", "role": "scaffold", "why": "Prompting mental-state generation"}], "pattern": "fix-env-vary-agent", "fixes": ["Sotopia"], "varies": ["ToM use"], "benchmarks": [{"name": "Sotopia", "purpose": "social interaction evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "ToMA trains dialogue agents to produce mental states useful for goals, evaluated on Sotopia."}
{"id": 6440, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Rubric rewards and offline guidance in GRPO for multi-domain reasoning"}], "pattern": "n/a", "fixes": ["14 benchmarks"], "varies": ["RL reward and guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "RGR-GRPO uses rubrics for dense rewards and offline guidance in multi-domain RL reasoning."}
{"id": 4392, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "RL agent with LLM-derived memory graph guidance"}], "pattern": "fix-env-vary-agent", "fixes": ["RL environments"], "varies": ["memory guidance"], "benchmarks": [], "results": [], "workflow": [], "summary": "MIRA uses an evolving memory graph from experience and LLM priors to cut RL sample complexity."}
{"id": 2653, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "analysis", "why": "Shows SAE metrics fail to distinguish trained from random transformers"}], "pattern": "vary-conditions", "fixes": ["SAE metrics"], "varies": ["trained vs random models"], "benchmarks": [], "results": [], "workflow": [], "summary": "Finds auto-interpretability and reconstruction metrics similar for SAEs on random and trained Pythia models."}
{"id": 14675, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "HiEval hierarchical QA scoring and error analysis models"}], "pattern": "vary-evaluation", "fixes": ["QA systems"], "varies": ["evaluator"], "benchmarks": [], "results": [], "workflow": [], "summary": "HiEval is a curriculum-trained QA evaluator giving quick correctness scoring and error type analysis."}
{"id": 6675, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Reusable latent building blocks for multi-agent systems"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["MAS architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "Agent Primitives decomposes MAS into reusable latent primitives communicating via latent states."}
{"id": 10260, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Score-based diffusion for task-adaptive communication graph generation"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["graph topology"], "benchmarks": [], "results": [], "workflow": [], "summary": "MAGE generates multi-agent communication graphs with score-based diffusion framed as energy minimization."}
{"id": 16849, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Abductive visual-linguistic reasoning for knowledge-based VQA"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["reasoning method"], "benchmarks": [], "results": [], "workflow": [], "summary": "ViLA hypothesizes answers, generates visual and textual premises and uses fuzzy scoring for KB-VQA."}
{"id": 12709, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Head-tail rebalancing in LVLM self-improvement"}], "pattern": "vary-data", "fixes": ["LVLMs"], "varies": ["resampling"], "benchmarks": [], "results": [], "workflow": [], "summary": "Counters Matthew effect in LVLM self-improvement via distribution reshaping and trajectory resampling."}
{"id": 18509, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Q-learning critic selecting persuasion policies for fire evacuation"}], "pattern": "fix-env-vary-agent", "fixes": ["simulated resident"], "varies": ["policy selection"], "benchmarks": [], "results": [], "workflow": [], "summary": "DiPS trains a critic to select persuasion strategies per turn, beating zero-shot and RAG baselines."}
{"id": 17290, "in_agent_scope": false, "types": [{"area": "multiling/multi", "role": "method", "why": "Language-family connector sharing for LLM-based ASR"}], "pattern": "n/a", "fixes": ["frozen encoder and LLM"], "varies": ["connector sharing"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shares one connector per language family in LLM-based ASR, reducing parameters and improving generalization."}