d6fc7c1dfa
- Work type = area x role (fixed vs varied), vocabulary in scripts/worktypes.json - LLM-read labels for 2,678 agent-related papers (data/labels), rule labels for the rest - Per-type synthesis notes (why-same, sub-variants, benchmarks, workflow) - Pages: /work-types matrix and detail (timeline, benchmarks, workflow, score charts), /benchmarks - build_worktypes.py adds work-type tables to data/papers.db Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bhs9LU8eMixk6g1kb1d3j7
61 lines
33 KiB
JSON
61 lines
33 KiB
JSON
{"id": 7102, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Hierarchical agents: SFT sub-goal planner plus offline-to-online RL executor for long-horizon control"}, {"area": "multi-agent", "role": "train-agent", "why": "Decomposes agent behavior into coordinated high-level and low-level agents"}], "pattern": "fix-env-vary-agent", "fixes": ["interactive environments"], "varies": ["agent hierarchy", "training recipe"], "benchmarks": [], "results": [], "workflow": ["SFT high-level sub-goal agent", "RL low-level action agent", "coordinate hierarchically"], "summary": "Hierarchical two-agent LLM framework trained with SFT and RL to reduce objective drift in long-horizon interactive tasks."}
|
|
{"id": 396, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Test-time query embedding refinement using a complementary retriever's scores for hybrid multimodal retrieval"}], "pattern": "fix-env-vary-agent", "fixes": ["retrievers", "benchmarks"], "varies": ["fusion method"], "benchmarks": [{"name": "visual document retrieval benchmarks", "purpose": "evaluate hybrid retrieval", "is_new": false}], "results": [], "workflow": ["take primary query embedding", "score with complementary retriever", "optimize query at test time"], "summary": "Guided Query Refinement optimizes a vision retriever's query embedding at test time with a text retriever's guidance."}
|
|
{"id": 13331, "in_agent_scope": false, "types": [{"area": "knowledge/editing", "role": "defense", "why": "Safety-aware locate-then-edit method aligning updates with estimated safety direction"}], "pattern": "vary-conditions", "fixes": ["editing framework", "threat of unsafe responses"], "varies": ["editing update constraint"], "benchmarks": [{"name": "ZsRE", "purpose": "editing effectiveness", "is_new": false}, {"name": "COUNTERFACT", "purpose": "editing effectiveness", "is_new": false}, {"name": "Mal-KSet", "purpose": "malicious queries safety", "is_new": false}], "results": [], "workflow": ["estimate safety transform", "find safety direction", "constrain parameter updates"], "summary": "Model editing method that keeps sequential edits from degrading safety, reducing unsafe responses while preserving edit success."}
|
|
{"id": 14290, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "application", "why": "Two-phase curriculum fine-tuning for name entity resolution in KYC identity data"}], "pattern": "vary-data", "fixes": ["base LLM"], "varies": ["fine-tuning curriculum"], "benchmarks": [{"name": "Indian identity name pairs", "purpose": "50,000 held-out pairs", "is_new": false}], "results": [{"benchmark": "Indian identity name pairs", "system": "SGER", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 99.02, "unit": "%"}], "workflow": ["train on name structure parsing", "train binary matching"], "summary": "Fine-tunes an LLM via structure-then-matching curriculum for robust Indian name matching."}
|
|
{"id": 3441, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "analysis", "why": "Theoretical noise decomposition explaining when chunked multi-agent processing beats single-shot long-context"}, {"area": "arch/longctx", "role": "theory", "why": "Formal framework of task, model and aggregator noise"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["chunking vs single shot", "input length"], "benchmarks": [{"name": "retrieval, QA, summarization tasks", "purpose": "verify theory", "is_new": false}], "results": [], "workflow": [], "summary": "Theory of three noise types explains when divide-and-conquer chunking helps long-context LLMs."}
|
|
{"id": 16482, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "method", "why": "Question proposing/answering report metric used as RL reward for medical report generation"}, {"area": "eval/judge", "role": "eval-method", "why": "MedQPA question-based factuality metric correlating with human preference"}], "pattern": "vary-evaluation", "fixes": ["report generators"], "varies": ["metric", "reward signal"], "benchmarks": [], "results": [], "workflow": ["propose questions", "answer from report", "use score as RL reward"], "summary": "MedQPA metric and RL-trained generator for medical image report generation."}
|
|
{"id": 9687, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Query-only prompt rewriting attack inflating tool-augmented agent reasoning cost"}], "pattern": "vary-conditions", "fixes": ["agent", "tools"], "varies": ["rewritten input prompt"], "benchmarks": [], "results": [], "workflow": ["multi-agent iterative rewriting", "policy control", "induce verbose trajectories"], "summary": "Sponge Tool Attack rewrites prompts to cause verbose, costly tool-calling reasoning stealthily."}
|
|
{"id": 12437, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "survey", "why": "Position paper formulating uncertainty quantification for LLM agents and open challenges"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [{"name": "tau^2-bench", "purpose": "numerical analysis of agent UQ challenges", "is_new": false}], "results": [], "workflow": [], "summary": "Position paper formalizing agent uncertainty quantification, with challenges and future directions."}
|
|
{"id": 14066, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Agentic workflow for cross-jurisdiction financial report extraction and verification"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["domain workflow"], "benchmarks": [{"name": "USA, Japan, China annual filings", "purpose": "evaluate consistency", "is_new": false}], "results": [], "workflow": ["acquire filings", "extract", "map to canonical ontology", "log anomalies"], "summary": "Agentic workflow using LLMs as constrained verifiers for localized multi-market financial reporting."}
|
|
{"id": 4529, "in_agent_scope": false, "types": [{"area": "arch/attention", "role": "method", "why": "MoE routing guided by temporal multimodal interaction dynamics"}, {"area": "arch/multimodal", "role": "method", "why": "Interaction-aware router for multimodal MoE models"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["router design"], "benchmarks": [], "results": [], "workflow": ["quantify temporal interactions", "route tokens by interaction"], "summary": "Interaction-aware router that guides mixture-of-experts specialization in multimodal models."}
|
|
{"id": 15215, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "environment", "why": "Scalable automatic construction of fully simulated heterogeneous function-calling environments"}, {"area": "tool-use", "role": "train-agent", "why": "Two-phase agent fine-tuning on environment experiences"}], "pattern": "vary-both", "fixes": ["function-calling task type"], "varies": ["environment diversity", "fine-tuning stage"], "benchmarks": [], "results": [], "workflow": ["build simulated environments", "collect interactions", "general then domain fine-tune"], "summary": "Scales simulated function-calling environments and trains agents in two phases for general agentic ability."}
|
|
{"id": 10461, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "DISBench for agentic image retrieval over visual histories"}, {"area": "embodied-nav", "role": "benchmark", "why": "Multimodal context-dependent exploration tasks"}], "pattern": "fix-env-vary-agent", "fixes": ["agents under test"], "varies": ["tasks"], "benchmarks": [{"name": "DISBench", "purpose": "agentic image retrieval in visual histories", "is_new": true}], "results": [], "workflow": ["mine spatiotemporal associations with VLMs", "human verify"], "summary": "Introduces DISBench and an agent baseline with tools and dual memory for context-aware image retrieval."}
|
|
{"id": 706, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Hypothesis-driven clinical agent trained with SFT and RL to request tests and diagnose"}], "pattern": "fix-env-vary-agent", "fixes": ["clinical environment"], "varies": ["training objectives"], "benchmarks": [], "results": [], "workflow": ["request tests", "update hypothesis", "SFT then RL", "diagnose"], "summary": "LA-CDM, an uncertainty-aware language agent trained with hybrid SFT/RL for interactive diagnosis."}
|
|
{"id": 15849, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Configurable environment with 31 user preference settings and UX metric"}, {"area": "domain-agents", "role": "benchmark", "why": "Evaluates agent interaction quality alongside task accuracy"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["user preference settings"], "benchmarks": [{"name": "PrefIx", "purpose": "evaluate interaction and UX", "is_new": true}], "results": [], "workflow": ["treat interaction as tool calls", "LLM judge on UX dimensions"], "summary": "Environment benchmarking agents on adapting to implicit user preferences with a UX metric."}
|
|
{"id": 2269, "in_agent_scope": false, "types": [{"area": "reason/code", "role": "data", "why": "AesCode-358K instruction dataset for code aesthetics"}, {"area": "reason/code", "role": "train-agent", "why": "GRPO with agentic reward feedback for aesthetics"}, {"area": "reason/code", "role": "benchmark", "why": "OpenDesign benchmark for code aesthetics"}], "pattern": "vary-data", "fixes": ["base LLM"], "varies": ["data", "reward"], "benchmarks": [{"name": "OpenDesign", "purpose": "assess code aesthetics", "is_new": true}, {"name": "PandasPlotBench", "purpose": "existing benchmark", "is_new": false}], "results": [], "workflow": ["build dataset", "SFT", "RL with multi-agent reward"], "summary": "Dataset, agentic reward RL and benchmark to improve aesthetics of LLM-generated code."}
|
|
{"id": 6967, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Backdoor attack hiding payload in skill implementations of skill-based agents"}], "pattern": "vary-conditions", "fixes": ["agent system"], "varies": ["backdoored skills"], "benchmarks": [], "results": [], "workflow": ["embed split payload in skills", "activate by trigger"], "summary": "SkillTrojan backdoors skill-based agents by distributing encrypted payloads across skill invocations; releases 3,000+ skill dataset."}
|
|
{"id": 4103, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "method", "why": "N-gram hierarchical linguistic distribution detector of LLM-generated text"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["detector"], "benchmarks": [], "results": [], "workflow": ["n-gram features at three levels", "Bayesian comparison"], "summary": "Zero-shot AI text detector comparing hierarchical n-gram distributions of human and machine text."}
|
|
{"id": 16159, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "NewsAgent benchmark for agentic newswriting with search and revision"}, {"area": "domain-agents", "role": "benchmark", "why": "Journalistic workflow tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "NEWSAGENT", "purpose": "journalistic writing agents", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 6k examples evaluating agents searching, selecting and revising for newswriting."}
|
|
{"id": 7045, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Multi-agent preference training internalizing contextual privacy in agent actions"}], "pattern": "vary-conditions", "fixes": ["threat model", "benchmarks"], "varies": ["training for privacy"], "benchmarks": [], "results": [], "workflow": ["embed privacy preferences", "multi-agent preference training"], "summary": "PrivAct internalizes contextual privacy into agents via preference training, cutting leakage with comparable helpfulness."}
|
|
{"id": 7022, "in_agent_scope": true, "types": [{"area": "search-research", "role": "data", "why": "Tree-guided data synthesis for information-seeking optimization modeling"}, {"area": "science-data", "role": "train-agent", "why": "R-GRPO training of retrieval-using optimization modeling agent"}], "pattern": "vary-data", "fixes": ["environment"], "varies": ["synthesized data", "RL algorithm"], "benchmarks": [], "results": [], "workflow": ["synthesize tree-based problems", "train with R-GRPO", "retrieve docs and model"], "summary": "Opt-Miner synthesizes data and trains an agent to retrieve web knowledge for optimization modeling."}
|
|
{"id": 3247, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Dual-agent perception and reasoning loop using 3D expert tools for minimal sufficient information"}], "pattern": "vary-scaffold", "fixes": ["VLM", "3D tools"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["perception agent queries 3D scene", "reasoning agent prunes", "iterate until minimal"], "summary": "MSSR dual-agent framework for spatial reasoning that curates a minimal sufficient set of 3D information."}
|
|
{"id": 13651, "in_agent_scope": true, "types": [{"area": "search-research", "role": "application", "why": "Multi-agent literature discovery and analysis system"}, {"area": "multi-agent", "role": "application", "why": "Coder-LLM orchestrated pipelines"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["system pipelines"], "benchmarks": [], "results": [], "workflow": ["discovery retrieval and ranking", "paper knowledge graphs"], "summary": "Open-source multi-agent system for finding, scoring and analyzing academic papers."}
|
|
{"id": 1597, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "DRBench of enterprise deep research mixing web and private data"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "DRBench", "purpose": "enterprise deep research", "is_new": true}], "results": [], "workflow": ["synthesize tasks", "human verify"], "summary": "Benchmark of 100 enterprise deep-research tasks across 10 domains requiring web and private knowledge."}
|
|
{"id": 669, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "MCP-SafetyBench with real MCP servers and 20 attack types"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["MCP attack tasks"], "benchmarks": [{"name": "MCP-SafetyBench", "purpose": "MCP safety evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Safety benchmark on real MCP servers across five domains and 20 attack types."}
|
|
{"id": 16565, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "analysis", "why": "Controlled arithmetic study of when answers form during CoT and its causal effect"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["CoT conditions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Finds LLMs compute sub-answers on the fly during CoT, so reasoning chains faithfully reflect computation."}
|
|
{"id": 8395, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Dynamic backward contribution propagation to isolate malicious agents in MAS"}, {"area": "multi-agent", "role": "defense", "why": "Signed DAG model of agent communication"}], "pattern": "vary-conditions", "fixes": ["threat model"], "varies": ["defense mechanism"], "benchmarks": [], "results": [], "workflow": ["model communication as signed DAG", "backpropagate contributions", "isolate malicious agents"], "summary": "Dynamic defense for multi-agent LLM systems identifying corrupted agents via contribution backpropagation."}
|
|
{"id": 14005, "in_agent_scope": true, "types": [{"area": "science-data", "role": "data", "why": "Distills GitHub repositories into vetted code example library for scientific coding agents"}, {"area": "code-swe", "role": "application", "why": "Code library generation system"}], "pattern": "vary-scaffold", "fixes": ["ASD agent"], "varies": ["code library"], "benchmarks": [], "results": [], "workflow": ["distill repositories", "vet examples", "augment agent"], "summary": "CodeDistiller builds working code libraries from 250 materials science repositories to improve scientific discovery agents."}
|
|
{"id": 13683, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "application", "why": "Research idea advising system with compressed literature and structured reasoning"}, {"area": "eval/bench", "role": "analysis", "why": "Studies factors for advising systems on ICLR acceptance"}], "pattern": "vary-conditions", "fixes": ["test set"], "varies": ["model size", "context", "confidence"], "benchmarks": [{"name": "ICLR 2025 test set", "purpose": "acceptance rate evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "Small advising model with literature database and structured reasoning outperforms Deepseek-R1 on ICLR acceptance prediction."}
|
|
{"id": 14491, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Multi-agent functional mining and refactoring for cross-architecture code migration"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [{"name": "OpenCV migration", "purpose": "real-world library migration", "is_new": false}], "results": [], "workflow": ["mine function semantics", "refactor", "migrate"], "summary": "FSCM multi-agent framework avoids surface-pattern mapping in cross-architecture migration of OpenCV."}
|
|
{"id": 14512, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Adaptive meta-hint guidance with Affinity metric to keep RL rollouts on-policy"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["guidance method"], "benchmarks": [], "results": [], "workflow": ["give abstract meta-hints", "preserve Affinity"], "summary": "HINT guides RL reasoning training with meta-hints and a metric of guidance-policy compatibility."}
|
|
{"id": 17670, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "analysis", "why": "Compares human and LLM word associations across temperatures and lexical factors"}], "pattern": "vary-conditions", "fixes": ["SWOW-EN data"], "varies": ["models", "temperature"], "benchmarks": [{"name": "SWOW-EN", "purpose": "human word associations", "is_new": false}], "results": [], "workflow": [], "summary": "LLMs mirror human frequency and concreteness trends but differ in variability and typicality of associations."}
|
|
{"id": 11964, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multi-agent financial report framework with code agent and variable memory"}, {"area": "search-research", "role": "application", "why": "Deep research for financial reporting"}], "pattern": "n/a", "fixes": ["LLMs"], "varies": ["framework"], "benchmarks": [], "results": [], "workflow": ["code agent with variable memory", "two-stage writing", "vision refinement"], "summary": "FinSight multi-agent framework producing multimodal professional financial reports."}
|
|
{"id": 2267, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "WideSearch benchmark of broad large-scale information collection tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["search agents"], "varies": ["tasks"], "benchmarks": [{"name": "WideSearch", "purpose": "wide info-seeking agents", "is_new": true}], "results": [], "workflow": [], "summary": "200 curated bilingual tasks testing search agents on reliable, complete large-scale information collection."}
|
|
{"id": 2004, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Multi-frame node-guided 4D Gaussian splatting for dynamic scene reconstruction"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["temporal transformer on control nodes"], "summary": "Mango-GS improves temporal consistency in dynamic 4D scene reconstruction."}
|
|
{"id": 12377, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "YIELD dataset and metrics for information elicitation agents"}, {"area": "social-sim", "role": "data", "why": "Training on dialogues improves elicitation behavior"}], "pattern": "fix-agent-vary-env", "fixes": ["foundation LLMs"], "varies": ["dataset"], "benchmarks": [{"name": "YIELD", "purpose": "information elicitation dialogues", "is_new": true}], "results": [], "workflow": [], "summary": "26M-token human dialogue dataset with POMDP formalization and metrics for information elicitation agents."}
|
|
{"id": 15438, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Training-free logprob-margin guided early stopping of reasoning"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["decoding stopping rule"], "benchmarks": [], "results": [], "workflow": ["monitor margin at sentences", "stop thinking"], "summary": "ThinkBrake stops overthinking in reasoning models, reducing thinking tokens up to 30%."}
|
|
{"id": 4059, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Document and web-specific VLM vision encoder with self-supervised and supervised pretraining"}, {"area": "arch/multimodal", "role": "method", "why": "Model merging of encoders"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["vision encoder"], "benchmarks": [], "results": [], "workflow": ["self-supervised pretraining", "autoregressive pretraining", "merge encoders"], "summary": "DAVE vision encoder tailored for document understanding and web agents."}
|
|
{"id": 18031, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Swedish minimal pairs benchmark for syntactic evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["pairs"], "benchmarks": [{"name": "Swedish Benchmark of Linguistic Minimal Pairs", "purpose": "syntactic evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "2,500 Swedish minimal pairs over 25 phenomena evaluating eleven models."}
|
|
{"id": 4959, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Agentic RL with two-tier reward and experience repository for clinical inquiry"}, {"area": "domain-agents", "role": "environment", "why": "Multi-agent interactive consultation environment"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training"], "benchmarks": [{"name": "HealthBench", "purpose": "medical evaluation", "is_new": false}, {"name": "MAQuE", "purpose": "consultation evaluation", "is_new": false}], "results": [], "workflow": ["multi-agent environment", "two-tier reward", "experience repository"], "summary": "Doctor-R1 trained with experiential agentic RL for strategic patient inquiry and decisions."}
|
|
{"id": 7610, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Scheduling guidance and stochasticity in diffusion posterior sampling for inverse problems"}], "pattern": "fix-env-vary-agent", "fixes": ["diffusion model"], "varies": ["schedules"], "benchmarks": [], "results": [], "workflow": ["time-varying control schedules"], "summary": "TriPS optimizes DC, CFG and stochasticity schedules for imaging inverse problems."}
|
|
{"id": 16154, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "survey", "why": "Survey of LLM-based human-agent collaboration systems"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "First structured survey of LLM human-agent systems: components, interaction types, applications and challenges."}
|
|
{"id": 6824, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "environment", "why": "Large-scale open-web simulator world model trained on 1M+ interactions"}, {"area": "web-gui", "role": "data", "why": "Synthesized trajectories to train web agents"}], "pattern": "fix-agent-vary-env", "fixes": ["web agent"], "varies": ["simulator"], "benchmarks": [{"name": "WebArena", "purpose": "extrinsic web agent eval", "is_new": false}, {"name": "WebWorld-Bench", "purpose": "intrinsic simulator eval", "is_new": true}], "results": [], "workflow": ["train on open-web interactions", "simulate trajectories", "train agents"], "summary": "WebWorld open-web simulator whose synthesized trajectories improve Qwen3-14B by 9.2% on WebArena."}
|
|
{"id": 14596, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Training-free NL2SQL with structured decomposition and error-fix memory"}], "pattern": "vary-scaffold", "fixes": ["model weights"], "varies": ["decomposition", "memory"], "benchmarks": [], "results": [], "workflow": ["decompose question", "retrieve error-fix pairs", "self-correct"], "summary": "Memo-SQL uses diverse decomposition and experience memory for training-free text-to-SQL."}
|
|
{"id": 5543, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "Training-free KV cache reuse across agents' decoding to cut prefill"}], "pattern": "vary-scaffold", "fixes": ["accuracy target"], "varies": ["cache reuse"], "benchmarks": [], "results": [], "workflow": ["reuse decoding KV caches", "recompute sparse positions"], "summary": "RelayCaching accelerates multi-agent LLM collaboration by reusing prior agents' KV caches."}
|
|
{"id": 15371, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Training-free signalling-game equilibrium for inference-time multi-agent dialogue adjustment"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["interaction process"], "benchmarks": [], "results": [], "workflow": ["model dialogue as signalling game", "approximate equilibrium"], "summary": "LinguaGame treats multi-agent dialogue as a signalling game to improve communication efficiency."}
|
|
{"id": 14042, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "application", "why": "OS-style agentic architecture with MCP and ROS for embodied LLMs"}], "pattern": "n/a", "fixes": [], "varies": ["system architecture"], "benchmarks": [], "results": [], "workflow": ["agentic kernel", "process agents call tools", "ROS devices"], "summary": "rosaOS open-source agentic operating system controlling robots and devices through MCP and ROS."}
|
|
{"id": 5110, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Mask-guided subject swapping for video editing with condition fusion"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [{"name": "VBench", "purpose": "video quality evaluation", "is_new": false}], "results": [], "workflow": ["fuse conditions", "adaptive mask"], "summary": "DreamSwapV swaps subjects in videos using mask and reference image."}
|
|
{"id": 14245, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Iterative PEFT training for tool-conditioned agents in SME settings"}, {"area": "tool-use", "role": "data", "why": "Self-evolving trajectory collection pipeline"}], "pattern": "vary-data", "fixes": ["compute budget"], "varies": ["data", "training"], "benchmarks": [], "results": [], "workflow": ["bootstrap trajectories", "iterative PEFT", "deploy and evaluate"], "summary": "EfficientTool low-cost closed loop aligning tool-calling agents for SMEs."}
|
|
{"id": 6452, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "DocOS benchmark for document-guided GUI actions in web"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "DocOS", "purpose": "document-guided GUI agents", "is_new": true}], "results": [], "workflow": [], "summary": "Introduces proactive document-guided action paradigm and DocOS benchmark for GUI agents."}
|
|
{"id": 2123, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "method", "why": "Theoretically grounded epistemic uncertainty quantification for contextual QA"}], "pattern": "fix-env-vary-agent", "fixes": ["models"], "varies": ["UQ method"], "benchmarks": [], "results": [], "workflow": ["decompose cross-entropy", "estimate feature gaps"], "summary": "Epistemic uncertainty for contextual QA as semantic feature gaps relative to an ideal model."}
|
|
{"id": 15131, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "White-box visual adversarial perturbation attack on mobile agents' price constraints"}], "pattern": "vary-conditions", "fixes": ["agent"], "varies": ["adversarial images"], "benchmarks": [], "results": [], "workflow": ["semantic-decoupling loss on CLIP encoders", "perturb screenshot"], "summary": "PriceBlind adversarial perturbations make multimodal mobile agents ignore price evidence."}
|
|
{"id": 14074, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Human-agent multi-agent system generating interactive educational documents"}], "pattern": "n/a", "fixes": [], "varies": ["system"], "benchmarks": [], "results": [], "workflow": ["Planner", "Executor", "Evaluator"], "summary": "ViviDoc demo generating interactive documents via DocSpec; user study shows quality comparable to human-authored."}
|
|
{"id": 15686, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "application", "why": "Standard-native RAG over 5G specs with metadata indices"}, {"area": "knowledge/rag", "role": "benchmark", "why": "Two curated 5G QA datasets"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval system"], "benchmarks": [{"name": "5G expert QA dataset", "purpose": "573 expert questions", "is_new": true}, {"name": "5G evolution QA dataset", "purpose": "350 questions", "is_new": true}], "results": [], "workflow": ["index specs, changes, requests", "resolve cross-references"], "summary": "DeepSpecs RAG handles cross-references and specification evolution for 5G QA."}
|
|
{"id": 14611, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Multimodal RAG assigning VLM refiner and inspector roles for KB-VQA"}], "pattern": "vary-scaffold", "fixes": ["models"], "varies": ["RAG roles"], "benchmarks": [{"name": "EVQA", "purpose": "KB-VQA", "is_new": false}, {"name": "InfoSeek", "purpose": "KB-VQA", "is_new": false}], "results": [], "workflow": ["refine query", "retrieve", "inspect and route"], "summary": "WikiSeeker uses VLM as query refiner and context inspector in multimodal RAG."}
|
|
{"id": 16425, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Two-agent framework modeling explicit and implicit dimensions for proactive personalization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["generate dimensions", "generate response"], "summary": "PROPER agents model user knowledge gaps via generated dimensions for proactive assistance."}
|
|
{"id": 3400, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "AlphaBench for LLM formulaic alpha factor mining"}, {"area": "apps/domain", "role": "analysis", "why": "Effect of model, prompting and reasoning settings"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["tasks", "settings"], "benchmarks": [{"name": "AlphaBench", "purpose": "factor generation, evaluation, searching", "is_new": true}], "results": [], "workflow": [], "summary": "First benchmark for LLMs in formulaic alpha factor mining across three tasks."}
|
|
{"id": 15623, "in_agent_scope": false, "types": [{"area": "knowledge/halluc", "role": "eval-method", "why": "Metrics and Hinter-Guesser protocol for multi-turn confidence estimation"}, {"area": "knowledge/halluc", "role": "method", "why": "P(Sufficient) logit probe"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["metrics", "protocol"], "benchmarks": [], "results": [], "workflow": [], "summary": "Systematic study of multi-turn confidence with new InfoECE metric and P(Sufficient) probe."}
|
|
{"id": 13504, "in_agent_scope": false, "types": [{"area": "efficiency/compress", "role": "method", "why": "Efficient Kronecker-factored Fisher approximation for quantization"}], "pattern": "fix-env-vary-agent", "fixes": ["quantization schemes"], "varies": ["Hessian approximation"], "benchmarks": [], "results": [], "workflow": ["Kronecker factorization", "integrate into quantization"], "summary": "Fast Fisher-guided quantization: near-baseline at 4-bit, 5-6% degradation at 2-bit, up to 10x speedup."}
|
|
{"id": 13735, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Evaluates diffusion LLMs as agent backbones on embodied and tool tasks"}, {"area": "tool-use", "role": "analysis", "why": "dLLM failures in strict JSON tool calling"}], "pattern": "vary-conditions", "fixes": ["benchmarks"], "varies": ["backbone model"], "benchmarks": [{"name": "AgentBoard", "purpose": "embodied agents", "is_new": false}, {"name": "BFCL", "purpose": "tool calling", "is_new": false}], "results": [], "workflow": [], "summary": "dLLMs fail as reliable agentic backbones; introduces DiffuAgent evaluation framework."}
|
|
{"id": 15612, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "AJ-Bench for Agent-as-a-Judge across search, data systems and GUI"}, {"area": "web-gui", "role": "eval-method", "why": "Environment-aware agent judges"}], "pattern": "vary-evaluation", "fixes": ["judged trajectories"], "varies": ["judge type"], "benchmarks": [{"name": "AJ-Bench", "purpose": "agent-as-a-judge", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 155 tasks and 516 trajectories evaluating Agent-as-a-Judge verification."}
|