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
31 KiB
JSON
61 lines
31 KiB
JSON
{"id": 5716, "in_agent_scope": false, "types": [{"area": "efficiency/compress", "role": "analysis", "why": "Token-level analysis of repetition loops in pruned LLMs"}, {"area": "train/distill", "role": "method", "why": "FOCUS and RePAIR objectives for post-pruning fine-tuning"}], "pattern": "vary-data", "fixes": ["pruned model"], "varies": ["post-pruning fine-tuning objective"], "benchmarks": [], "results": [], "workflow": ["Analyze loop entry and persistence", "Reweight distillation (FOCUS)", "Margin loss on continuation pairs (RePAIR)"], "summary": "Analyzes text degeneration in pruned LLMs and proposes two token-level guidance objectives to reduce repetition after pruning."}
|
|
{"id": 16638, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Expert-oriented RAG with user-aware question entailment for community QA"}], "pattern": "n/a", "fixes": ["LLM generator"], "varies": ["retrieval and user modeling"], "benchmarks": [], "results": [], "workflow": ["Recognize entailed archived questions", "Retrieve expert answers", "Synthesize answer with LLM"], "summary": "Expert-oriented RAG framework for community question answering using user-aware entailment to cut hallucination."}
|
|
{"id": 5687, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "data", "why": "Composes short problems into long dependency chains for RL curriculum"}], "pattern": "vary-data", "fixes": ["RL recipe"], "varies": ["synthetic long-horizon training data", "curriculum"], "benchmarks": [{"name": "ReasoningGym", "purpose": "OOD transfer of long-horizon training", "is_new": false}], "results": [], "workflow": ["Compose problems into chains", "Outcome-reward RL", "Curriculum increases length"], "summary": "Bootstraps long-horizon reasoning by RL on synthetically composed multi-step chains from short-horizon problems."}
|
|
{"id": 12762, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Strategy-aware agent framework for autism intervention dialogue and clinical support"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["domain agent pipeline"], "benchmarks": [], "results": [], "workflow": ["DoctorAgent O-T-A-C loop", "Synthesize intervention dialogues", "Clinical decision support"], "summary": "ASDAgent synthesizes ABA intervention dialogue and supports clinical decisions using a strategy-aware reasoning loop."}
|
|
{"id": 2473, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Information-theoretic test for emergent higher-order structure in multi-agent LLMs"}], "pattern": "vary-conditions", "fixes": ["multi-agent LLM system"], "varies": ["randomized interventions", "measurement framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Information decomposition framework tests for emergent coordination in multi-agent LLMs using a guessing game with interventions."}
|
|
{"id": 772, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Evaluates foundation models' active exploration and spatial belief construction"}, {"area": "embodied-nav", "role": "eval-method", "why": "Proposes Theory of Space definition and belief-probing measures"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["foundation models"], "benchmarks": [], "results": [], "workflow": [], "summary": "Benchmark of whether foundation models build spatial beliefs through active exploration under partial observability."}
|
|
{"id": 13851, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Interactive tool-augmented emotional support benchmark with MCP-style tools"}, {"area": "social-sim", "role": "benchmark", "why": "Emotional support scenarios with process-level metrics"}], "pattern": "fix-env-vary-agent", "fixes": ["tool environment"], "varies": ["nine LLMs", "tool augmentation"], "benchmarks": [{"name": "TEA-Bench", "purpose": "evaluate tool-augmented ESC agents", "is_new": true}], "results": [], "workflow": [], "summary": "TEA-Bench evaluates tool-augmented agents in emotional support conversation, finding capacity-dependent gains."}
|
|
{"id": 1078, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Agentic retrieval and graph-of-thought for Lean autoformalization"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent workflow"], "benchmarks": [{"name": "ProofNet", "purpose": "autoformalization evaluation", "is_new": false}], "results": [], "workflow": ["Decompose into dependency graph", "Retrieve Mathlib definitions", "Construct and score formalization"], "summary": "Aria is an agent for Lean statement autoformalization via dependency graph decomposition and grounded checking."}
|
|
{"id": 14989, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Contrastive decoding using Spotlight and Shadow layers to reduce MLLM hallucination"}], "pattern": "vary-scaffold", "fixes": ["MLLM weights"], "varies": ["decoding procedure"], "benchmarks": [], "results": [], "workflow": ["Pick Spotlight and Shadow layers", "Contrast logits per token"], "summary": "DaID is a contrastive decoding method that mitigates MLLM hallucination via attention-guided layer selection."}
|
|
{"id": 10498, "in_agent_scope": true, "types": [{"area": "science-data", "role": "environment", "why": "Closed-loop materials discovery environments for agents"}, {"area": "science-data", "role": "benchmark", "why": "Benchmarks discovery agents and algorithms under oracle budget"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["discovery environment"], "benchmarks": [{"name": "MADE", "purpose": "benchmark closed-loop materials discovery", "is_new": true}], "results": [], "workflow": [], "summary": "MADE simulates closed-loop materials discovery campaigns to benchmark agents under constrained oracle budgets."}
|
|
{"id": 16968, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "benchmark", "why": "PictureStories dataset of learner answers annotated for task adherence"}], "pattern": "n/a", "fixes": ["models"], "varies": ["task adherence prediction"], "benchmarks": [{"name": "PictureStories", "purpose": "task adherence prediction", "is_new": true}], "results": [], "workflow": [], "summary": "Builds a dataset of learner answers to picture stories and explores automatic task adherence prediction."}
|
|
{"id": 16689, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "LLM-and-decision-tree synthetic anomalies for evaluating tabular anomaly detection"}], "pattern": "vary-evaluation", "fixes": ["AD models"], "varies": ["evaluation set construction"], "benchmarks": [], "results": [], "workflow": ["LLM proposes anomaly conditions", "Decision tree grounds them", "Build pseudo-evaluation set"], "summary": "AutoAnoEval builds semantically grounded pseudo-evaluation sets to select tabular anomaly detection models."}
|
|
{"id": 13456, "in_agent_scope": false, "types": [{"area": "safety/fairness", "role": "defense", "why": "Multi-agent fairness-aware QE to mitigate gender bias in translation"}], "pattern": "n/a", "fixes": ["QE models"], "varies": ["bias-aware aggregation"], "benchmarks": [], "results": [], "workflow": ["Detect gender cues", "Generate gender-flipped variants", "Aggregate QE and LLM scores"], "summary": "FairQE mitigates gender bias in translation quality estimation with a multi-agent plug-and-play framework."}
|
|
{"id": 450, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Chain-of-Decomposition with retrieval, reformulation and LoRA-trained reasoning for KBQA"}], "pattern": "n/a", "fixes": ["KBQA task"], "varies": ["decomposed pipeline"], "benchmarks": [], "results": [], "workflow": ["Retrieve subgraphs", "Reformulate into reasoning paths", "LLM reasoning with LoRA"], "summary": "CoD decomposes KBQA into retrieval, reformulation and lightweight LLM reasoning."}
|
|
{"id": 14495, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Multi-agent adaptive multi-granularity memory management for LLM agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory system"], "benchmarks": [], "results": [], "workflow": ["Hierarchical memory granularities", "Agents retrieve and update", "Resolve inconsistencies"], "summary": "AMA uses coordinated agents to manage memory at multiple granularities for long-term agent interaction."}
|
|
{"id": 15798, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "analysis", "why": "Constructs ConLangs to probe LLMs' metalinguistic grammatical knowledge"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["constructed languages"], "benchmarks": [], "results": [], "workflow": ["Generate language components via prompts", "Refine with commentary"], "summary": "Agentic system for constructed languages used to probe what LLMs know about language."}
|
|
{"id": 14275, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Automatically generated NL2SQL benchmark with multi-faceted ambiguity and unanswerability"}], "pattern": "fix-agent-vary-env", "fixes": ["NL2SQL systems"], "varies": ["ambiguity types", "user behaviors"], "benchmarks": [{"name": "Clarity", "purpose": "ambiguous NL2SQL evaluation", "is_new": true}, {"name": "Spider", "purpose": "base for evaluation", "is_new": false}, {"name": "BIRD", "purpose": "base for evaluation", "is_new": false}], "results": [], "workflow": ["Transform SQL to ambiguous queries", "Add conversational continuations"], "summary": "Clarity generates an NL2SQL benchmark with ambiguity and shows leading systems degrade."}
|
|
{"id": 18532, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "eval-method", "why": "Next-utterance classification as automated evaluation of paralinguistic-aware dialogue systems"}], "pattern": "vary-evaluation", "fixes": ["spoken dialogue systems"], "varies": ["evaluation protocol"], "benchmarks": [], "results": [], "workflow": ["LLM predicts next dialogue act", "Compare scenarios"], "summary": "Corpus-based evaluation using next-utterance classification for paralinguistic-aware spoken dialogue systems."}
|
|
{"id": 16302, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Exploration then deployment dual-agent text-to-SQL with MCTS-style schema exploration"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["exploration and agent pipeline"], "benchmarks": [], "results": [], "workflow": ["MCTS-inspired schema exploration", "Build knowledge base", "Dual-agent query generation"], "summary": "SQLAgent explores databases to build a knowledge base before generating SQL."}
|
|
{"id": 14112, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Federated prototype alignment with geometric refinement for heterogeneous architectures"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["federated method"], "benchmarks": [], "results": [], "workflow": ["Project to shared space", "Refine prototypes at server", "Anchor classifiers"], "summary": "FedPAGR aligns class prototypes across heterogeneous federated clients with geometric regularization."}
|
|
{"id": 14381, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Preference-aware memory update mixing sliding-window and EMA averages"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory update mechanism"], "benchmarks": [], "results": [], "workflow": ["Sliding window average", "Exponential moving average fusion"], "summary": "PAMU dynamically refines preference memory for long-term LLM agents."}
|
|
{"id": 2550, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "LLM-guided evolutionary search with memory for materials design"}], "pattern": "vary-scaffold", "fixes": ["LLM", "oracle"], "varies": ["search framework"], "benchmarks": [], "results": [], "workflow": ["LLM proposes candidates", "Surrogate oracle estimates properties", "Update success and failure memories"], "summary": "LLEMA couples LLMs with evolutionary rules and memory to discover materials across 14 tasks."}
|
|
{"id": 10267, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Safe in-context RL under constrained MDP with exact-penalty dual"}], "pattern": "n/a", "fixes": ["ICRL"], "varies": ["safety constraint method"], "benchmarks": [], "results": [], "workflow": [], "summary": "SCARED enables safe adaptation of in-context reinforcement learning within a safety cost budget."}
|
|
{"id": 10105, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Actor-critic MARL fine-tuning for decentralized LLM collaboration"}], "pattern": "fix-env-vary-agent", "fixes": ["collaboration tasks"], "varies": ["training algorithm"], "benchmarks": [], "results": [], "workflow": ["Centralized or decentralized critic", "Fine-tune collaborating LLMs"], "summary": "MAAC methods CoLLM-CC and CoLLM-DC optimize decentralized LLM collaboration with actor-critic training."}
|
|
{"id": 1938, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "benchmark", "why": "Rubik's Cube benchmark for spatial reasoning, state tracking and active exploration"}, {"area": "planning-decision", "role": "benchmark", "why": "Long-horizon tiered diagnosis; 0% pass on long-horizon tasks"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["LLMs"], "benchmarks": [{"name": "CubeBench", "purpose": "diagnose long-horizon physical intelligence", "is_new": true}], "results": [], "workflow": [], "summary": "CubeBench is a Rubik's Cube benchmark revealing LLM agents fail at long-horizon tasks."}
|
|
{"id": 6441, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "LLM-extracted diagnostic evidence alignment for medical vision-language pretraining"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["pretraining objective"], "benchmarks": [], "results": [], "workflow": ["LLM extracts evidence from reports", "Build evidence space", "Evidence-level alignment"], "summary": "LGDEA aligns images and reports at diagnostic evidence level for medical VLP with limited pairing."}
|
|
{"id": 15782, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Data mixture optimizer for multitask SFT of mobile agents"}, {"area": "web-gui", "role": "benchmark", "why": "PhoneAgentBench for MLLMs on mobile tasks"}], "pattern": "vary-data", "fixes": ["SFT recipe"], "varies": ["data mixture ratios"], "benchmarks": [{"name": "PhoneAgentBench", "purpose": "evaluate MLLMs on mobile phone tasks", "is_new": true}], "results": [], "workflow": ["Predict performance from ratio", "Optimize mixture", "Fine-tune MLLM"], "summary": "DaMo predicts optimal data mixtures for mobile phone agent SFT and introduces PhoneAgentBench."}
|
|
{"id": 18585, "in_agent_scope": false, "types": [{"area": "interp/mech", "role": "analysis", "why": "Probes attention head specialization over Pythia checkpoints using lexical ambiguity"}], "pattern": "vary-conditions", "fixes": ["Pythia models"], "varies": ["checkpoints", "stress tests"], "benchmarks": [{"name": "Pythia", "purpose": "checkpoint suite for developmental probe", "is_new": false}], "results": [], "workflow": [], "summary": "Developmental probe of attention head specialization for word sense disambiguation."}
|
|
{"id": 11910, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Timely-Eval for time-budgeted tool use under varied latency"}, {"area": "tool-use", "role": "train-agent", "why": "Timely-RL trains time-aware reasoning"}], "pattern": "vary-both", "fixes": ["models"], "varies": ["tool latency", "training"], "benchmarks": [{"name": "Timely-Eval", "purpose": "time-aware test-time scaling", "is_new": true}], "results": [], "workflow": ["Cold-start SFT", "RL for time awareness"], "summary": "Redefines test-time as wall-clock time, with Timely-Eval benchmark and Timely-RL training."}
|
|
{"id": 7215, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "Benchmark for expert-level deep research report evaluation with rubrics"}, {"area": "search-research", "role": "eval-method", "why": "Taxonomy, rubric and claim verification for report judging"}], "pattern": "fix-agent-vary-env", "fixes": ["deep research systems"], "varies": ["evaluation criteria"], "benchmarks": [{"name": "DEER", "purpose": "expert report generation evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "DEER benchmarks deep research agents with an expert taxonomy and 101 rubric items."}
|
|
{"id": 17704, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "benchmark", "why": "Japanese parent-child dialogic reading corpus with annotations"}], "pattern": "n/a", "fixes": ["-"], "varies": ["corpus"], "benchmarks": [], "results": [], "workflow": [], "summary": "Constructs and analyzes a multimodal corpus of Japanese parent-child picture-book reading dialogues."}
|
|
{"id": 12162, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "defense", "why": "Safety Guidance Trigger steers fine-tuning toward safety manifold against malicious fine-tuning"}], "pattern": "vary-conditions", "fixes": ["threat model"], "varies": ["protection mechanism"], "benchmarks": [], "results": [], "workflow": ["Optimize safety trigger", "Train with trigger"], "summary": "SGT protects open-weight LLMs from malicious fine-tuning via a learned safety trigger."}
|
|
{"id": 1340, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Rule-guided active inference with wake-sleep learning of human habits"}], "pattern": "n/a", "fixes": ["trajectories"], "varies": ["habit learning method"], "benchmarks": [], "results": [], "workflow": ["Active inference on trajectories", "Harvest candidate rules", "Sleep consolidation"], "summary": "Learns symbolic habit rules alongside human behavior models via wake-sleep active inference."}
|
|
{"id": 15885, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Runtime reflection-driven safety module with verifiers and reflective memory"}], "pattern": "vary-scaffold", "fixes": ["agent workflow"], "varies": ["safety module"], "benchmarks": [], "results": [], "workflow": ["Monitor decision traces", "Retrieve repair exemplars", "Verify before action"], "summary": "Safety Sidecar provides plug-and-play runtime safety control for tool-using agents."}
|
|
{"id": 5806, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "environment", "why": "Synthesizes interactive environments and verifiable tasks from scratch"}, {"area": "tool-use", "role": "train-agent", "why": "RL in synthetic environments improves tool-use generalization"}], "pattern": "fix-agent-vary-env", "fixes": ["training recipe"], "varies": ["training environments"], "benchmarks": [{"name": "tau2-Bench", "purpose": "unseen tool-use generalization", "is_new": false}, {"name": "VitaBench", "purpose": "unseen tool-use generalization", "is_new": false}], "results": [], "workflow": ["Procedurally test environments", "Expand tool dependency graph", "Verify actions"], "summary": "ScaleEnv synthesizes diverse interactive environments for agent training, improving tool-use generalization."}
|
|
{"id": 8438, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "method", "why": "Bayesian inference of latent partial orders from agent traces"}], "pattern": "n/a", "fixes": ["traces"], "varies": ["inference method"], "benchmarks": [{"name": "Cloud-IaC-6", "purpose": "cloud provisioning trace suite", "is_new": true}, {"name": "WFCommons", "purpose": "scientific workflows", "is_new": false}], "results": [], "workflow": ["Infer partial order via MCMC", "Compile executor pruning context"], "summary": "BPOP infers dependency graphs from linear agent traces to enable efficient execution."}
|
|
{"id": 5232, "in_agent_scope": false, "types": [{"area": "reason/math", "role": "eval-method", "why": "DAG-based logical closeness metric for CoT"}, {"area": "reason/math", "role": "benchmark", "why": "DAG-MATH CoT format benchmark"}], "pattern": "vary-evaluation", "fixes": ["LLMs"], "varies": ["metric"], "benchmarks": [{"name": "DAG-MATH", "purpose": "evaluate rule-consistent reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "Models CoT as DAG process and introduces logical closeness metric and benchmark."}
|
|
{"id": 11780, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Multi-Consistency Reward for consistent reasoning in Med-VQA"}, {"area": "align/rlvr", "role": "data", "why": "Med-Zero-17K dataset"}], "pattern": "vary-data", "fixes": ["RLVR"], "varies": ["reward", "dataset"], "benchmarks": [{"name": "Med-Zero-17K", "purpose": "training data for Med-VQA", "is_new": true}], "results": [], "workflow": [], "summary": "Fixes inconsistent RLVR reasoning in medical VQA with a new dataset and consistency reward."}
|
|
{"id": 15399, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free layer-guided model merging to restore MLLM language reasoning"}], "pattern": "n/a", "fixes": ["MLLMs"], "varies": ["merging method"], "benchmarks": [], "results": [], "workflow": ["Mask vision tokens layerwise", "Find plateau stages", "Merge base LM parameters"], "summary": "PlaM merges base LM parameters into MLLMs guided by layerwise plateau analysis."}
|
|
{"id": 4402, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Object-centric prompts refine CLIP patches for zero-shot segmentation"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "OC-ZSS refines CLIP patch representations with object prompts for zero-shot segmentation."}
|
|
{"id": 12009, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "DUDE trains web agents to resist deceptive UI"}, {"area": "web-gui", "role": "benchmark", "why": "RUC benchmark of deceptive scenarios"}], "pattern": "vary-both", "fixes": ["threat"], "varies": ["defense training"], "benchmarks": [{"name": "RUC", "purpose": "deceptive web UI scenarios", "is_new": true}], "results": [], "workflow": ["Hybrid-reward learning", "Experience summarization"], "summary": "DUDE trains web agents against deceptive interfaces and introduces RUC benchmark."}
|
|
{"id": 12221, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Server-side MCP plugin with look-ahead world model to filter risky tools"}], "pattern": "vary-scaffold", "fixes": ["agent"], "varies": ["defense plugin"], "benchmarks": [], "results": [], "workflow": ["Ground environment dynamics", "Filter tools proactively", "Intervene as fail-safe"], "summary": "SafeMCP constrains tool acquisition using look-ahead reasoning to prevent power-seeking."}
|
|
{"id": 5793, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Latent RAG of experiences trained end-to-end with RL"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["experience mechanism"], "benchmarks": [], "results": [], "workflow": ["Encode experiences with hidden states", "Latent retrieval per step", "Cross-attention integration"], "summary": "ExpWeaver lets agents learn from experience via latent retrieval trained with RL."}
|
|
{"id": 6150, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "500-question multimodal search decision-making benchmark"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["VLMs"], "benchmarks": [{"name": "NAVIGATE", "purpose": "evaluate visual-guided search decisions", "is_new": true}], "results": [], "workflow": [], "summary": "NAVIGATE evaluates when and how VLMs search the open web from visual evidence."}
|
|
{"id": 7264, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Evaluates implicit requirement inference by agents"}, {"area": "tool-use", "role": "environment", "why": "Agent-as-a-World YAML simulated worlds"}], "pattern": "fix-env-vary-agent", "fixes": ["worlds"], "varies": ["agents"], "benchmarks": [{"name": "Implicit Intelligence", "purpose": "implicit requirements evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark and LLM-simulated world harness testing whether agents infer unstated constraints."}
|
|
{"id": 7052, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Counterexample-guided workflow graph refinement via failure signatures"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow optimization"], "benchmarks": [], "results": [], "workflow": ["Maintain counterexample pool", "Estimate failure modes", "Propose and verify edits"], "summary": "CE-Graph refines agent workflows using failure distributions rather than scalar scores."}
|
|
{"id": 15279, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "train-agent", "why": "Answer-conditioned variational reasoning to improve RL sampling of reasoning paths"}], "pattern": "vary-data", "fixes": ["RL"], "varies": ["reasoning path derivation"], "benchmarks": [], "results": [], "workflow": ["Condition on reference answer", "Derive reasoning path", "RL refine"], "summary": "RAVR uses reference answers to derive high-utility reasoning paths for RL."}
|
|
{"id": 3158, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Agentic weather science framework with code environment"}, {"area": "science-data", "role": "environment", "why": "ZephyrusWorld environment"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [{"name": "WeatherBench 2", "purpose": "weather data interface", "is_new": false}], "results": [], "workflow": ["Code-based interaction", "Tools for forecasting", "Iterative analysis"], "summary": "Zephyrus is a multi-turn LLM weather agent within ZephyrusWorld."}
|
|
{"id": 12340, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "environment", "why": "HA-Desire home assistance simulation with LLM proxy user"}, {"area": "social-sim", "role": "train-agent", "why": "Communication-efficient desire alignment method"}], "pattern": "vary-both", "fixes": ["-"], "varies": ["environment", "agent method"], "benchmarks": [{"name": "HA-Desire", "purpose": "user adaptation evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "HA-Desire simulates home assistance with LLM user; proposes method for efficient desire alignment."}
|
|
{"id": 15172, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Dynamic procedural memory with distillation, reuse and pruning"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory framework"], "benchmarks": [], "results": [], "workflow": ["Distill experiences", "Context-adaptive reuse", "Utility-based refinement"], "summary": "ReMe manages procedural memory lifecycle for experience-driven agent evolution."}
|
|
{"id": 16160, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Training-free self-correcting T2V refinement with localized regeneration"}], "pattern": "vary-scaffold", "fixes": ["T2V model"], "varies": ["refinement procedure"], "benchmarks": [], "results": [], "workflow": ["Detect misalignment", "Plan refinement", "Regenerate regions"], "summary": "VideoRepair detects and corrects text-video misalignments locally without training."}
|
|
{"id": 6295, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Candidate-aware causal reasoning with GRPO for temporal answer grounding"}], "pattern": "n/a", "fixes": ["task"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": ["Generate candidate segments", "Temporal logic reasoning", "GRPO with rejection reward"], "summary": "CACR grounds answers in instructional videos with candidate selection and RL reasoning."}
|
|
{"id": 16289, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Hierarchical multi-agent scheduler for multi-tenant LLM serving"}], "pattern": "n/a", "fixes": ["serving workloads"], "varies": ["scheduling policy"], "benchmarks": [], "results": [], "workflow": ["Predict request attributes", "Summarize QoS states", "Update scheduling policy"], "summary": "H-MAS adaptively schedules LLM serving workloads with layered agents."}
|
|
{"id": 15086, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Reformulates information seeking as table completion planning"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["planning structure"], "benchmarks": [], "results": [], "workflow": ["Map query to table schema", "Fill cells via search", "Empty cells as plan"], "summary": "Table-as-Search tracks search state in an external table for deep and wide search."}
|
|
{"id": 12824, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Joint RL of policy and tool graph memory"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["policy and memory"], "benchmarks": [], "results": [], "workflow": ["Build structured experience memory", "Train policy with RL"], "summary": "SEARL jointly optimizes policy and tool-graph memory for self-evolving agents."}
|
|
{"id": 11732, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "method", "why": "KG-guided router selecting agents and LLMs for QA"}], "pattern": "n/a", "fixes": ["agents"], "varies": ["routing"], "benchmarks": [], "results": [], "workflow": ["Build heterogeneous KG", "Route with performance supervision"], "summary": "AgentRouter routes QA instances among agents and backbones using knowledge-graph guidance."}
|
|
{"id": 397, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Meta-RL cross-episode training for exploration in LLM agents"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["training framework"], "benchmarks": [], "results": [], "workflow": ["Cross-episode training", "In-context reflection adaptation"], "summary": "LaMer uses meta-RL to induce exploration and test-time adaptation in language agents."}
|
|
{"id": 18465, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Event-guided slot interaction for dialogue state tracking"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["DST model"], "benchmarks": [], "results": [], "workflow": ["Induce latent events", "Dual-stream fusion"], "summary": "Event-DST models latent events to coordinate slot interactions in dialogue state tracking."}
|
|
{"id": 3079, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Self-play RL where agent proposes and solves search tasks"}, {"area": "search-research", "role": "data", "why": "Synthesizes search tasks with ground-truth answers"}], "pattern": "vary-data", "fixes": ["RL"], "varies": ["task source"], "benchmarks": [], "results": [], "workflow": ["Proposer generates queries", "Solver answers via search", "RL on both"], "summary": "Search Self-Play trains deep search agents without supervision by proposing and solving tasks."}
|
|
{"id": 2817, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Physical-world privacy awareness benchmark for embodied LLM agents"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["LLMs"], "benchmarks": [{"name": "EAPrivacy", "purpose": "embodied privacy awareness", "is_new": true}], "results": [{"benchmark": "EAPrivacy", "system": "Gemini 2.5 Pro", "base_model": null, "setting": "baseline", "metric": "accuracy (changing environments)", "score": 59, "unit": "%"}], "workflow": [], "summary": "EAPrivacy measures privacy awareness of embodied LLM agents; top model reaches only 59% in changing environments."}
|