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
36 KiB
JSON
61 lines
36 KiB
JSON
{"id": 12134, "in_agent_scope": false, "types": [{"area": "train/data", "role": "data", "why": "Multi-agent lineage reconstruction of post-training datasets; builds lineage-aware diverse dataset"}, {"area": "train/data", "role": "analysis", "why": "Large-scale lineage analysis reveals redundancy and contamination propagation"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["dataset sampling via lineage"], "benchmarks": [], "results": [], "workflow": ["Reconstruct dataset lineage graph", "Analyze structural patterns", "Sample from upstream leaf sources"], "summary": "Introduces data lineage for post-training datasets via a multi-agent framework, revealing redundancy and contamination and yielding a lineage-aware diverse dataset."}
|
|
{"id": 12555, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Scalable pipeline generating long-horizon web tasks with executable trajectories"}, {"area": "web-gui", "role": "benchmark", "why": "Instantiates tasks on 50+ websites for training and evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["web agents under test"], "varies": ["generated tasks and sites"], "benchmarks": [], "results": [], "workflow": ["Crawl sites", "Retrieval-based seeding", "In-context task generation", "Replay and validate"], "summary": "GTA generates realistic long-horizon web tasks with dense executable trajectories across 50+ websites for training and evaluating web agents."}
|
|
{"id": 1661, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Trains agent to fold context at multiple scales for long-horizon web tasks"}, {"area": "search-research", "role": "scaffold", "why": "Context folding for information-seeking agents"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["context management paradigm"], "benchmarks": [{"name": "BrowseComp", "purpose": "long-horizon web search", "is_new": false}, {"name": "BrowseComp-ZH", "purpose": "Chinese web search", "is_new": false}], "results": [{"benchmark": "BrowseComp", "system": "AgentFold-30B-A3B", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 36.2, "unit": "%"}, {"benchmark": "BrowseComp-ZH", "system": "AgentFold-30B-A3B", "base_model": null, "setting": "proposed", "metric": "accuracy", "score": 47.3, "unit": "%"}], "workflow": ["Observe step", "Choose fold operation", "Condense or consolidate history"], "summary": "AgentFold agent learns proactive multi-scale context folding, reaching 36.2% on BrowseComp with a 30B-A3B model."}
|
|
{"id": 13086, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Compresses multi-agent workflows by pruning or downgrading redundant agents"}], "pattern": "vary-scaffold", "fixes": ["models", "tasks"], "varies": ["agent topology and cost"], "benchmarks": [], "results": [], "workflow": ["Score agent importance", "Remove or replace agents", "Validate with acceptance rule"], "summary": "AgentSlimming prunes or swaps redundant agents in graph workflows, cutting token cost up to 78.9% with little accuracy loss."}
|
|
{"id": 7949, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Controlled study of task horizon length effects on agent RL training"}], "pattern": "vary-conditions", "fixes": ["decision rules", "reasoning structure"], "varies": ["horizon length"], "benchmarks": [], "results": [], "workflow": [], "summary": "Empirical study showing horizon length alone destabilizes RL training of LLM agents, and horizon reduction helps stability and generalization."}
|
|
{"id": 15573, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Evolves semantic payloads for environmental injection attacks on GUI agents"}, {"area": "agent-safety", "role": "analysis", "why": "Finds semantic content dominates over visual appearance in attack success"}], "pattern": "vary-conditions", "fixes": ["target GUI agents"], "varies": ["injection payload semantics"], "benchmarks": [], "results": [], "workflow": ["Analyze visual vs semantic factors", "Evolve semantic payloads", "Deploy discovered attacks"], "summary": "EVA evolves semantic injection payloads against GUI agents, achieving 59-85% average ASR within about 1-2 iterations."}
|
|
{"id": 14915, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "LLM agent autonomously conducts inference attacks against ML services"}], "pattern": "n/a", "fixes": ["target services"], "varies": ["attacker agent design"], "benchmarks": [], "results": [{"benchmark": "20 target services", "system": "InferPilot (GPT-4o)", "base_model": "GPT-4o", "setting": "proposed", "metric": "task completion rate", "score": 100.0, "unit": "%"}], "workflow": ["Plan attack", "Use task-specific actions", "Adapt to service constraints"], "summary": "InferPilot is an autonomous LLM agent that carries out inference attacks on ML services with 100% task completion on 20 targets."}
|
|
{"id": 3782, "in_agent_scope": false, "types": [{"area": "arch/diffusion", "role": "method", "why": "Inpainting-guided exploration for RL of diffusion LLMs"}, {"area": "align/rlvr", "role": "train-agent", "why": "GRPO variant inserting partial ground-truth traces to restore gradients"}], "pattern": "n/a", "fixes": ["GRPO framework"], "varies": ["exploration via inpainting"], "benchmarks": [], "results": [], "workflow": ["Insert partial reasoning traces", "Sample with inpainting", "Optimize with group RL"], "summary": "IGPO uses dLLM inpainting to inject partial ground-truth reasoning during RL sampling, restoring gradients under sparse rewards."}
|
|
{"id": 16967, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Gaze-prediction-guided decoding to control text reading ease"}], "pattern": "n/a", "fixes": ["language model"], "varies": ["decoding guidance"], "benchmarks": [], "results": [], "workflow": ["Predict gaze patterns", "Steer LM generation", "Eye-tracking evaluation"], "summary": "Steers LM generation with a gaze-prediction model to make texts easier or harder to read, validated in an eye-tracking study."}
|
|
{"id": 386, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Diffusion-based generative framework for Pareto front refinement"}], "pattern": "n/a", "fixes": ["MOO benchmarks"], "varies": ["sampling method"], "benchmarks": [], "results": [], "workflow": ["Learn conditional diffusion", "Refine with MGD-style update", "Add RBF repulsion"], "summary": "SPREAD is a DDPM-based multi-objective optimizer with adaptive gradient and repulsion terms for Pareto coverage."}
|
|
{"id": 13276, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Conjunctive prompt attacks via trigger plus compromised remote agent template"}, {"area": "agent-safety", "role": "analysis", "why": "Shows existing defenses fail across star, chain, DAG topologies"}], "pattern": "vary-conditions", "fixes": ["model weights", "client agent"], "varies": ["trigger placement", "template insertion", "topology"], "benchmarks": [], "results": [], "workflow": ["Place trigger in query", "Insert template in remote agent", "Optimize routing-aware"], "summary": "Studies conjunctive attacks in multi-agent LLM systems where benign parts combine via routing to cause harm, evading existing defenses."}
|
|
{"id": 12761, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "Trajectory-level preference benchmark for reward models in tool-use scenarios"}, {"area": "tool-use", "role": "eval-method", "why": "Evaluates RMs and judges on distinguishing agent trajectories"}], "pattern": "fix-agent-vary-env", "fixes": ["reward models under test"], "varies": ["trajectory preference tasks"], "benchmarks": [{"name": "Plan-RewardBench", "purpose": "trajectory-level reward model evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Plan-RewardBench evaluates reward models and judges on preferred vs distractor tool-using agent trajectories across four task families."}
|
|
{"id": 12373, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "Multimodal industry classification benchmark with text and geospatial sources"}], "pattern": "n/a", "fixes": ["MLLMs"], "varies": ["input modalities and prompting design"], "benchmarks": [{"name": "MONETA", "purpose": "industry classification", "is_new": true}], "results": [{"benchmark": "MONETA", "system": "open-source MLLM baseline", "base_model": null, "setting": "baseline", "metric": "accuracy", "score": 62.1, "unit": "%"}, {"benchmark": "MONETA", "system": "closed-source MLLM baseline", "base_model": null, "setting": "baseline", "metric": "accuracy", "score": 74.1, "unit": "%"}], "workflow": [], "summary": "MONETA is a 1,000-business multimodal industry classification benchmark with training-free MLLM baselines."}
|
|
{"id": 3526, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "eval-method", "why": "Human-preference reward model for image editing"}, {"area": "align/pref", "role": "data", "why": "Uses reward model to select high-quality editing training subset"}], "pattern": "vary-data", "fixes": ["Step1X-Edit recipe"], "varies": ["data selection by reward"], "benchmarks": [{"name": "GenAI-Bench", "purpose": "human correlation", "is_new": false}, {"name": "AURORA-Bench", "purpose": "human correlation", "is_new": false}, {"name": "ImagenHub", "purpose": "human correlation", "is_new": false}, {"name": "EditReward-Bench", "purpose": "new editing reward benchmark", "is_new": true}, {"name": "ShareGPT-4o-Image", "purpose": "data selection source", "is_new": false}], "results": [], "workflow": ["Collect 200K preference pairs", "Train reward model", "Select subset for training"], "summary": "EditReward is a reward model trained on 200K expert preference pairs for image editing, used to filter training data."}
|
|
{"id": 8908, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Trains MLLM with SFT and agentic RL for multi-turn segmentation tool use"}, {"area": "tool-use", "role": "data", "why": "Pipeline synthesizing multi-turn reasoning segmentation trajectories"}], "pattern": "fix-env-vary-agent", "fixes": ["segmentation benchmarks"], "varies": ["agent model training"], "benchmarks": [{"name": "ReasonSeg", "purpose": "reasoning segmentation", "is_new": false}], "results": [{"benchmark": "ReasonSeg", "system": "RSAgent", "base_model": null, "setting": "proposed", "metric": "gIoU", "score": 66.5, "unit": "%"}], "workflow": ["Synthesize trajectories", "Cold-start SFT", "Agentic RL"], "summary": "RSAgent is an agentic MLLM that iteratively invokes visual tools to refine segmentation, trained with SFT then RL."}
|
|
{"id": 12079, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "benchmark", "why": "Benchmark combining commonsense and math reasoning steps per task"}, {"area": "planning-decision", "role": "analysis", "why": "61 LLMs show accuracy drop when steps combine"}], "pattern": "fix-agent-vary-env", "fixes": ["61 LLMs"], "varies": ["compositional tasks"], "benchmarks": [{"name": "AgentCoMa", "purpose": "mixed commonsense-math reasoning", "is_new": true}], "results": [], "workflow": [], "summary": "AgentCoMa shows LLMs solving both steps in isolation lose nearly 30% accuracy when commonsense and math are combined."}
|
|
{"id": 6981, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Multi-scale self-evolving conversational memory inspired by renormalization group"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["memory structure"], "benchmarks": [], "results": [], "workflow": ["Convert episodes to facts", "Coarse-grain hierarchically", "Update user profile"], "summary": "RGMem evolves long-term conversational memory through hierarchical coarse-graining into a dynamic user profile."}
|
|
{"id": 16405, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "DAG-planning agent for speech data quality management"}, {"area": "domain-agents", "role": "benchmark", "why": "Releases SpeechQM-Dataset of 24 verification tasks"}], "pattern": "n/a", "fixes": ["general LLM"], "varies": ["workflow planning"], "benchmarks": [{"name": "SpeechQM-Dataset", "purpose": "speech QC verification", "is_new": true}], "results": [], "workflow": ["Compile requirements to DAG", "Run modular tools", "Replan on failure"], "summary": "SpeechQM-Agent plans DAG workflows for speech data quality checks, reaching 80-90% agreement with experts at under 20% cost."}
|
|
{"id": 1999, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Agent revises dependency and action knowledge from experience in Minecraft"}, {"area": "memory-selfevolve", "role": "scaffold", "why": "Adaptive dependency graph and failure-aware action memory"}], "pattern": "vary-scaffold", "fixes": ["7B LLM", "Minecraft benchmarks"], "varies": ["knowledge correction mechanisms"], "benchmarks": [{"name": "Minecraft benchmarks", "purpose": "knowledge learning and planning", "is_new": false}], "results": [], "workflow": ["Correct dependencies from successes", "Correct actions from failures", "Plan with revised knowledge"], "summary": "XENON corrects flawed LLM priors from experience in Minecraft, letting a 7B model beat larger proprietary-model agents."}
|
|
{"id": 16015, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Multi-agent TableQA with diverse reasoning paths and small-model tools"}], "pattern": "vary-scaffold", "fixes": ["LLMs"], "varies": ["multi-agent pipeline"], "benchmarks": [], "results": [], "workflow": ["Generate candidate answers", "Refine or select with tools", "Minimize LLM calls"], "summary": "MATA generates diverse candidate TableQA answers and selects them with small-LM tools, giving strong accuracy and efficiency."}
|
|
{"id": 13121, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Three-stage curriculum RL training for VLM mobile agents"}], "pattern": "fix-env-vary-agent", "fixes": ["mobile environment"], "varies": ["training recipe"], "benchmarks": [], "results": [], "workflow": ["Format alignment", "On-policy exploration", "Multi-turn task-level RL"], "summary": "Mobile-R1 trains VLM mobile agents with a hierarchical curriculum culminating in multi-turn task-level RL."}
|
|
{"id": 466, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Information-theoretic debiasing of reward model training"}], "pattern": "n/a", "fixes": ["RLHF setup"], "varies": ["RM training objective"], "benchmarks": [], "results": [], "workflow": ["Maximize MI with preference", "Minimize MI with bias attributes"], "summary": "DIR debiases reward models by maximizing preference mutual information while minimizing mutual information with biased attributes."}
|
|
{"id": 17584, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "data", "why": "Large-scale Basque speech and text corpora infrastructure"}, {"area": "multiling/lowres", "role": "application", "why": "Framework for open Basque language models"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Euskorpora is an initiative building data infrastructure and open models for Basque digital sovereignty."}
|
|
{"id": 17551, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Ranking-based passage retrieval with graph attention refinement for ODQA"}], "pattern": "n/a", "fixes": ["RAG generator"], "varies": ["retriever"], "benchmarks": [], "results": [], "workflow": ["Rank candidate passages", "Refine scores with GAT", "Generate answer"], "summary": "RBR improves RAG-based ODQA retrieval by modeling relative passage relevance with a ranking approach and graph attention."}
|
|
{"id": 17113, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Uses multi-agent disagreement to recruit visual tools"}, {"area": "tool-use", "role": "scaffold", "why": "Tool selection guided by debate disagreement"}], "pattern": "vary-scaffold", "fixes": ["VLMs"], "varies": ["tool recruitment scheme"], "benchmarks": [{"name": "A-OKVQA", "purpose": "VQA evaluation", "is_new": false}, {"name": "MMMU", "purpose": "multimodal reasoning", "is_new": false}], "results": [], "workflow": ["Agents debate", "Disagreement picks tools", "Aggregator selects answer"], "summary": "DART recruits visual tools based on debating-agent disagreement, beating multi-agent debate by 3.4% on A-OKVQA and 2.4% on MMMU."}
|
|
{"id": 2085, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "benchmark", "why": "GeoPerceive benchmark isolating geometric perception with DSL"}, {"area": "align/pref", "role": "train-agent", "why": "GeoDPO RL with translator-based DSL rewards"}], "pattern": "n/a", "fixes": ["VLMs"], "varies": ["training reward"], "benchmarks": [{"name": "GeoPerceive", "purpose": "geometric perception", "is_new": true}], "results": [], "workflow": ["Generate diagram-DSL pairs", "Train NL-to-DSL translator", "RL with DSL scores"], "summary": "GeoPerceive benchmarks VLM geometric perception and GeoDPO improves it using translator-derived fine-grained RL rewards."}
|
|
{"id": 14302, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "method", "why": "Backward-question refinement distilling evaluation plans for contact center QA"}], "pattern": "n/a", "fixes": ["small LMs"], "varies": ["evaluation plan"], "benchmarks": [], "results": [], "workflow": ["Generate backward questions", "Distill reasoning into plans", "Refine criteria"], "summary": "BQR uses backward questions to refine evaluation plans for contact center QA, with gains up to 27.8% Macro F1."}
|
|
{"id": 2607, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "RL-trained Conductor orchestrates worker LLMs and topologies"}], "pattern": "fix-env-vary-agent", "fixes": ["worker LLM pools"], "varies": ["coordination policy"], "benchmarks": [{"name": "LiveCodeBench", "purpose": "coding reasoning", "is_new": false}, {"name": "GPQA", "purpose": "science reasoning", "is_new": false}], "results": [], "workflow": ["Design topology", "Write worker instructions", "Aggregate worker outputs"], "summary": "A 7B Conductor trained with RL learns coordination topologies and prompts for worker LLMs, surpassing individual workers."}
|
|
{"id": 17275, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "benchmark", "why": "Evaluation framework of three tasks on 13 Indigenous languages"}, {"area": "multiling/lowres", "role": "analysis", "why": "Shows large variation across model families and languages"}], "pattern": "n/a", "fixes": ["models"], "varies": ["languages and tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "Evaluates GPT, Gemini, DeepSeek, Qwen and LLaMA on Indigenous American languages, finding many near chance."}
|
|
{"id": 16948, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RLVR-trained search agents over scientific paper corpus"}, {"area": "search-research", "role": "benchmark", "why": "PaperSearchQA 60k factoid QA over 16M abstracts"}], "pattern": "fix-env-vary-agent", "fixes": ["search corpus"], "varies": ["agent training"], "benchmarks": [{"name": "PaperSearchQA", "purpose": "scientific paper QA", "is_new": true}], "results": [], "workflow": ["Build corpus", "Train with RLVR", "Analyze behaviors"], "summary": "Releases a 16M-abstract corpus and PaperSearchQA, training RLVR search agents that beat non-RL retrieval baselines."}
|
|
{"id": 17109, "in_agent_scope": false, "types": [{"area": "interp/icl", "role": "analysis", "why": "GPT-2 learning of natural vs impossible languages"}], "pattern": "vary-conditions", "fixes": ["GPT-2"], "varies": ["languages", "perturbations"], "benchmarks": [], "results": [], "workflow": [], "summary": "Re-examines impossible-language learning, finding GPT-2 learns natural and impossible languages about equally easily."}
|
|
{"id": 1335, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "EduVisBench for pedagogical visual explanations"}, {"area": "domain-agents", "role": "scaffold", "why": "EduVisAgent multi-agent framework for visualization"}], "pattern": "vary-both", "fixes": ["foundation models"], "varies": ["benchmark", "agent framework"], "benchmarks": [{"name": "EduVisBench", "purpose": "pedagogical visualization", "is_new": true}], "results": [], "workflow": [], "summary": "Introduces EduVisBench for educational visual reasoning and EduVisAgent, a multi-agent framework for pedagogical visualization."}
|
|
{"id": 13450, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "Distillation-guided RL for compact agentic RAG models"}, {"area": "search-research", "role": "analysis", "why": "ARC metric analyzes reasoning, search, synthesis"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["training algorithm"], "benchmarks": [], "results": [], "workflow": ["Cold-start from teacher", "Teacher-guided RL", "Measure with ARC"], "summary": "DGPO trains 0.5-1B models for agentic RAG via teacher demonstrations and guidance, with ARC metric analyzing capabilities."}
|
|
{"id": 4316, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Evaluation of computer-use agents exploiting web vulnerabilities"}, {"area": "web-gui", "role": "benchmark", "why": "GUI-based cybersecurity tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["CUAs"], "varies": ["vulnerability tasks"], "benchmarks": [{"name": "HackWorld", "purpose": "CUA web exploitation", "is_new": true}], "results": [], "workflow": [], "summary": "HackWorld evaluates whether computer-use agents can discover and exploit web application vulnerabilities through GUIs."}
|
|
{"id": 13346, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "environment", "why": "Large-scale Chinese shopping simulation environment"}, {"area": "domain-agents", "role": "benchmark", "why": "Evaluates LLMs, best under 40% full success"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs"], "varies": ["shopping environment"], "benchmarks": [{"name": "ShopSimulator", "purpose": "shopping assistant evaluation and RL", "is_new": true}], "results": [], "workflow": [], "summary": "ShopSimulator is a Chinese shopping environment for evaluating and RL-training agents; best models get under 40% full success."}
|
|
{"id": 1750, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Training-free VLM-guided negative prompting for creative image generation"}], "pattern": "n/a", "fixes": ["diffusion model"], "varies": ["inference-time prompting"], "benchmarks": [], "results": [], "workflow": ["Generate image", "VLM identifies concept", "Add adaptive negative prompt"], "summary": "Training-free inference-time method using VLM-guided negative prompts to make text-to-image outputs more creative while valid."}
|
|
{"id": 11492, "in_agent_scope": false, "types": [{"area": "efficiency/inference", "role": "method", "why": "Entropy-guided dynamic KV-cache budget allocation across heads"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["budget allocation"], "benchmarks": [], "results": [], "workflow": ["Measure attention entropy", "Allocate head budgets", "Compress KV cache"], "summary": "EntroKV allocates KV-cache budgets per head using attention entropy as proxy for compression sensitivity."}
|
|
{"id": 15548, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "method", "why": "Dual-stream latent persona steering preserving quirks"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["personalization method"], "benchmarks": [], "results": [], "workflow": ["Anchor traits dual-stream", "Narrative refinement", "Generate persona-faithful text"], "summary": "QuirkyMind steers LLMs toward faithful personas including atypical quirks via a dual-stream latent state."}
|
|
{"id": 10805, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Benchmark for personalized web agents using user history on open web"}, {"area": "social-sim", "role": "benchmark", "why": "Ambiguous queries requiring implicit preference inference"}], "pattern": "fix-agent-vary-env", "fixes": ["agent architectures"], "varies": ["user histories and queries"], "benchmarks": [{"name": "Persona2Web", "purpose": "personalized web agents", "is_new": true}], "results": [], "workflow": [], "summary": "Persona2Web benchmarks web agents on resolving ambiguous queries from long user histories on the real web."}
|
|
{"id": 13414, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "GRPO with composite reasoning-aware rewards for function calling"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["reward design"], "benchmarks": [{"name": "BFCL", "purpose": "function calling", "is_new": false}, {"name": "ACEBench", "purpose": "function calling", "is_new": false}], "results": [], "workflow": ["Define composite rewards", "Optimize with GRPO"], "summary": "R2IF aligns reasoning with tool-call decisions via composite RL rewards, improving BFCL by up to 34.62% over baselines."}
|
|
{"id": 12269, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "environment", "why": "Simulated user and mock tool environments for RL"}, {"area": "tool-use", "role": "data", "why": "Synthesizes diverse underspecified tool-use tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["RL recipe"], "varies": ["tasks", "simulated environment"], "benchmarks": [], "results": [], "workflow": ["Teacher creates tasks and tools", "Simulate users and tools", "Rubric-based rewards"], "summary": "SynthAgent synthesizes tool-use tasks and simulated environments with rubric rewards to train small agentic LMs."}
|
|
{"id": 5652, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Grid-world study of constitutions for multi-agent stability"}, {"area": "multi-agent", "role": "scaffold", "why": "Evolves behavioral norms automatically"}], "pattern": "vary-conditions", "fixes": ["LLM agents", "grid-world"], "varies": ["constitutions"], "benchmarks": [], "results": [{"benchmark": "grid-world", "system": "Claude 4.5 Opus designed constitution", "base_model": null, "setting": "baseline", "metric": "Societal Stability Score", "score": 0.332, "unit": "score"}], "workflow": [], "summary": "Constitutional Evolution automatically discovers norms for multi-agent LLM coordination, outperforming hand-designed constitutions."}
|
|
{"id": 8802, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "MM-DeceptionBench for multimodal deception"}, {"area": "agent-safety", "role": "eval-method", "why": "Debate with images multi-agent evaluation"}], "pattern": "vary-evaluation", "fixes": ["VLMs"], "varies": ["evaluation framework"], "benchmarks": [{"name": "MM-DeceptionBench", "purpose": "multimodal deception", "is_new": true}], "results": [], "workflow": [], "summary": "Introduces MM-DeceptionBench and a debate-with-images multi-agent evaluation to detect deception in multimodal LLMs."}
|
|
{"id": 11650, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "analysis", "why": "Studies persistent contamination across 20 models in tool-augmented agents"}], "pattern": "vary-conditions", "fixes": ["agent setup"], "varies": ["models", "contamination type"], "benchmarks": [], "results": [], "workflow": [], "summary": "Formalizes persistent semantic entities in tool-augmented agents, showing contamination persists across sessions and agent boundaries."}
|
|
{"id": 2453, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Flatness-guided test-time adaptation for VLMs"}], "pattern": "n/a", "fixes": ["VLMs"], "varies": ["TTA method"], "benchmarks": [], "results": [], "workflow": [], "summary": "FGA uses training-time flatness to guide test-time adaptation of vision-language models."}
|
|
{"id": 12230, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "train-agent", "why": "Multi-objective RL for role-playing agents with rubrics"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL optimization"], "benchmarks": [], "results": [], "workflow": ["Multi-rubric training", "Thought-augmented rollout", "Off-policy guidance"], "summary": "MOA trains role-playing agents with multi-objective RL over fine-grained rubrics and thought-augmented rollouts."}
|
|
{"id": 641, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "RL sampling policy for active prompt learning with CLIP"}], "pattern": "n/a", "fixes": ["CLIP"], "varies": ["sample selection"], "benchmarks": [], "results": [], "workflow": ["Soft Actor-Critic selection", "Hybrid reward", "Self-training"], "summary": "PSP uses a prompt-guided RL sampling policy to select samples for active prompt learning."}
|
|
{"id": 8526, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "RL with recursive reasoning for strategic multi-agent games"}], "pattern": "fix-env-vary-agent", "fixes": ["games"], "varies": ["RL framework"], "benchmarks": [], "results": [], "workflow": ["Recursive reasoning", "Intermediate rewards", "RL training"], "summary": "Strat-Reasoner uses recursive reasoning and RL credit assignment to improve LLM strategic reasoning in multi-agent games."}
|
|
{"id": 7349, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "benchmark", "why": "ConvexBench for convexity recognition under composition depth"}, {"area": "tool-use", "role": "scaffold", "why": "Agentic divide-and-conquer with external parsing tool"}], "pattern": "vary-both", "fixes": ["frontier LLMs"], "varies": ["depth", "agentic framework"], "benchmarks": [{"name": "ConvexBench", "purpose": "convexity recognition", "is_new": true}], "results": [{"benchmark": "ConvexBench", "system": "frontier LLMs depth 2", "base_model": null, "setting": "baseline", "metric": "F1", "score": 1.0, "unit": "score"}, {"benchmark": "ConvexBench", "system": "frontier LLMs depth 100", "base_model": null, "setting": "baseline", "metric": "F1", "score": 0.2, "unit": "score"}], "workflow": [], "summary": "ConvexBench reveals LLM compositional reasoning gaps in convexity recognition and proposes an agentic divide-and-conquer fix."}
|
|
{"id": 15099, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "DisastQA, 3,000 disaster QA questions under varied evidence"}], "pattern": "n/a", "fixes": ["20 models"], "varies": ["evidence conditions"], "benchmarks": [{"name": "DisastQA", "purpose": "disaster management QA", "is_new": true}], "results": [], "workflow": [], "summary": "DisastQA benchmarks 20 models on disaster QA with noisy evidence and a keypoint-based open-ended evaluation."}
|
|
{"id": 14881, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Supervision-free cold-start RL for small VLM tool use"}], "pattern": "fix-env-vary-agent", "fixes": ["SVLMs"], "varies": ["training framework"], "benchmarks": [], "results": [], "workflow": ["Structured multi-turn rollouts", "Multi-objective reward", "Cold-start RL"], "summary": "SPECTRA bootstraps tool-using agentic abilities of small VLMs via supervision-free cold-start RL."}
|
|
{"id": 7920, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "eval-method", "why": "Reference-free trajectory evaluation with evidence bank"}], "pattern": "vary-evaluation", "fixes": ["tool-augmented agents"], "varies": ["evaluation metric"], "benchmarks": [], "results": [], "workflow": ["Accumulate evidence bank", "Assess trajectory dimensions"], "summary": "TRACE evaluates tool-augmented agent trajectories reference-free on efficiency, hallucination and adaptivity, validated on a meta-evaluation dataset."}
|
|
{"id": 7966, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Agent-world-model synergy learning task-sufficient representations"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["world model learning"], "benchmarks": [], "results": [], "workflow": ["Active probing", "Curriculum", "Structured world model learning"], "summary": "Learns minimal task-sufficient world models via closed-loop agentic exploration and structured modeling."}
|
|
{"id": 4353, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Search-based discovery of privacy attack strategies in agent simulation"}, {"area": "agent-safety", "role": "defense", "why": "Alternates improving defense strategies"}], "pattern": "vary-conditions", "fixes": ["simulated interactions"], "varies": ["attack and defense instructions"], "benchmarks": [], "results": [], "workflow": ["Simulate interactions", "LLM optimizer proposes instructions", "Parallel search"], "summary": "Searches via simulation for privacy attack and defense strategies among LLM agents, finding attacks escalate to sophisticated tactics."}
|
|
{"id": 15914, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Dialogue-corpus adaptation for RAG with dual querying and graph structure"}], "pattern": "n/a", "fixes": ["RAG pipeline"], "varies": ["dialogue structuring"], "benchmarks": [], "results": [], "workflow": ["Dual dialogue querying", "Graph-structured dialogues", "Subgraph selection"], "summary": "SDR adapts goal-oriented dialogue corpora for RAG via issue/solution pseudo-documents and graph-based evidence selection."}
|
|
{"id": 14938, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "survey", "why": "Position on collaborative agents and effort scaling"}, {"area": "social-sim", "role": "analysis", "why": "Case studies show SOTA agents underperform in multi-turn collaboration"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Argues for collaborative agents and introduces collaborative effort scaling to measure utility growth with user involvement."}
|
|
{"id": 15595, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "analysis", "why": "Trajectory analysis of 15 LLMs in evolutionary search"}], "pattern": "vary-conditions", "fixes": ["evolutionary search setup"], "varies": ["LLM optimizers"], "benchmarks": [], "results": [], "workflow": [], "summary": "Large-scale study of what makes LLMs good optimizers in evolutionary search, finding strong ones act as local refiners."}
|
|
{"id": 11904, "in_agent_scope": false, "types": [{"area": "multiling/mt", "role": "benchmark", "why": "MENT meta-evaluation dataset for non-literal translation"}, {"area": "multiling/mt", "role": "eval-method", "why": "RATE agentic translation evaluation"}], "pattern": "vary-evaluation", "fixes": ["MT systems"], "varies": ["metrics"], "benchmarks": [{"name": "MENT", "purpose": "MT metric meta-evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "Builds MENT to test MT metrics on non-literal translation and proposes RATE, an agentic evaluation framework."}
|
|
{"id": 15011, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Online Chinese mobile GUI benchmark with 1080 tasks and noise"}], "pattern": "fix-agent-vary-env", "fixes": ["13 agents"], "varies": ["tasks", "noise"], "benchmarks": [{"name": "MobileBench-OL", "purpose": "mobile GUI agents online", "is_new": true}], "results": [], "workflow": [], "summary": "MobileBench-OL provides 1080 tasks over 80 Chinese apps measuring execution, reasoning and noise robustness of mobile agents."}
|
|
{"id": 7879, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Bi-level co-evolution of context-engineering skills and contexts"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["context engineering skills"], "benchmarks": [], "results": [], "workflow": ["Meta agent evolves skills", "Base agent executes", "Optimize context files"], "summary": "MCE co-evolves context-engineering skills and artifacts via a meta-level agent rather than manual harnesses."}
|