{"id": 13824, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "application", "why": "Memory OS consolidating dialogue into MemCells and MemScenes with agentic retrieval"}], "pattern": "vary-scaffold", "fixes": ["base LLM"], "varies": ["memory architecture"], "benchmarks": [], "results": [], "workflow": ["form MemCells", "consolidate MemScenes", "agentic retrieval"], "summary": "Self-organizing agent memory system with episodic formation, semantic consolidation and reconstructive retrieval for long-horizon dialogue."} {"id": 12883, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL-trained TKGQA agent with expanded actions and reverse curriculum"}], "pattern": "fix-env-vary-agent", "fixes": ["TKGQA benchmarks"], "varies": ["RL training", "curriculum"], "benchmarks": [{"name": "MultiTQ", "purpose": "SOTA evaluation", "is_new": false}, {"name": "TimelineKGQA", "purpose": "SOTA evaluation", "is_new": false}], "results": [], "workflow": ["add internal actions", "reverse curriculum", "RL training"], "summary": "RL-trained autonomous agent for temporal KG question answering using internal actions and reverse curriculum."} {"id": 16178, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Lightweight hierarchical graph memory with temporal-aware search and reranking"}], "pattern": "vary-scaffold", "fixes": ["base LLM"], "varies": ["memory structure", "retrieval"], "benchmarks": [], "results": [], "workflow": ["build CogniGraph", "temporal search", "rerank"], "summary": "Agentic memory framework using CogniGraph hierarchical graph for efficient long-term dialogue reasoning."} {"id": 16495, "in_agent_scope": true, "types": [{"area": "search-research", "role": "scaffold", "why": "Self-evolving agent iteratively creates query rewriting rules to enhance BM25"}], "pattern": "vary-scaffold", "fixes": ["BM25", "LLM params"], "varies": ["rewriting rules"], "benchmarks": [{"name": "LeCaRD-v2", "purpose": "evaluate legal case retrieval", "is_new": false}], "results": [], "workflow": ["generate rules", "validate combinations", "prune rules"], "summary": "Training-free self-evolving LLM agent that creates and prunes query-rewriting rules for legal case retrieval."} {"id": 18296, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "train-agent", "why": "Fine-tunes SpeechLLM with rubric reasoning and conformal calibration"}], "pattern": "vary-both", "fixes": ["base model"], "varies": ["fine-tuning"], "benchmarks": [], "results": [], "workflow": ["rubric reasoning", "fine-tune", "conformal calibration"], "summary": "Rubric-guided fine-tuning of Qwen2-Audio with conformal calibration for L2 speech assessment."} {"id": 6901, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "method", "why": "Interactive video virtual try-on with spatial-semantic guidance"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Introduces interactive video virtual try-on task and guidance-based method."} {"id": 18518, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "scaffold", "why": "Modular architecture blocking execution until missing info is clarified"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["architecture"], "benchmarks": [], "results": [], "workflow": ["track task structure", "detect missing info", "block, clarify"], "summary": "Modular architecture enforcing missing information as an execution precondition for instruction-following agents."} {"id": 16061, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "benchmark", "why": "ActorMindBench for speech role-playing"}, {"area": "social-sim", "role": "scaffold", "why": "Multi-agent CoT reasoning framework for speech role-playing"}], "pattern": "vary-both", "fixes": ["base models"], "varies": ["tasks", "reasoning framework"], "benchmarks": [{"name": "ActorMindBench", "purpose": "speech role-playing evaluation", "is_new": true}], "results": [], "workflow": ["multi-agent reasoning", "personalized response"], "summary": "Speech role-playing benchmark and multi-agent chain-of-thought framework emulating actor reasoning."} {"id": 14284, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "data", "why": "User simulator synthesizes and selects dialogue data"}, {"area": "domain-agents", "role": "train-agent", "why": "Train recruitment proactive dialogue agents"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training data selection"], "benchmarks": [], "results": [], "workflow": ["simulate users", "evaluate with CoI", "select data", "train"], "summary": "Simulator-based data synthesis and CoI-based selection to train recruitment proactive dialogue agents."} {"id": 12886, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "environment", "why": "Hierarchical demographic tree for topic-adaptive agent population generation"}], "pattern": "fix-agent-vary-env", "fixes": ["simulation agents"], "varies": ["population generation"], "benchmarks": [], "results": [], "workflow": ["build topic tree", "ground real data", "sample agents"], "summary": "Framework generating realistic topic-adaptive agent populations for agent-based simulation."} {"id": 13998, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "application", "why": "UI for offline evaluation and refinement of multi-agent workflows"}], "pattern": "vary-scaffold", "fixes": ["workflow"], "varies": ["agent prompts"], "benchmarks": [], "results": [], "workflow": ["run workflow", "score nodes", "backward evaluation"], "summary": "Demo UI that scores intermediate artifacts, localizes bottlenecks and refines multi-agent LLM workflows."} {"id": 10451, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "LLM designs communication protocol for cooperative MARL"}], "pattern": "fix-env-vary-agent", "fixes": ["MARL benchmarks"], "varies": ["communication protocol"], "benchmarks": [], "results": [], "workflow": ["LLM proposes protocol", "state-awareness check", "refine"], "summary": "LLM-driven iterative design of MARL communication protocols for state reconstruction."} {"id": 14530, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "benchmark", "why": "Multimodal financial RAG benchmark"}, {"area": "domain-agents", "role": "train-agent", "why": "ReAct agent trained on agentic trajectories"}], "pattern": "vary-both", "fixes": ["base models"], "varies": ["tasks", "agent training"], "benchmarks": [{"name": "FinMRAGBench", "purpose": "financial multimodal RAG", "is_new": true}], "results": [], "workflow": ["collect trajectories", "train ReAct agent"], "summary": "Financial multimodal RAG benchmark of 887 QA pairs plus a trained ReAct agent."} {"id": 18192, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "defense", "why": "Cross-lingual multilingual safety classifier for 100 languages"}], "pattern": "vary-data", "fixes": ["threat model"], "varies": ["guardrail model"], "benchmarks": [], "results": [], "workflow": ["pick 13 languages", "train classifier", "transfer to 100"], "summary": "0.5B multilingual safety guardrail with cluster-guided cross-lingual transfer from 13 languages."} {"id": 14468, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Multi-track comparative music QA benchmark"}], "pattern": "fix-agent-vary-env", "fixes": ["audio-language models"], "varies": ["tasks"], "benchmarks": [{"name": "Jamendo-MT-QA", "purpose": "comparative music QA", "is_new": true}, {"name": "Jamendo-QA", "purpose": "base dataset", "is_new": false}], "results": [], "workflow": [], "summary": "Benchmark of 36,519 comparative QA items over music track pairs."} {"id": 11811, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Multi-agent RL for planning, coding, answering on tables"}], "pattern": "fix-env-vary-agent", "fixes": ["table benchmarks"], "varies": ["multi-agent RL"], "benchmarks": [], "results": [], "workflow": ["plan", "code", "answer", "RL"], "summary": "Multi-agent framework with planner, coder, answerer trained by RL for table reasoning."} {"id": 2527, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "method", "why": "Semantic-guided watermark for reasoning LLMs"}], "pattern": "n/a", "fixes": ["models"], "varies": ["watermark method"], "benchmarks": [], "results": [], "workflow": ["thinking phase", "criticality score", "adaptive watermark"], "summary": "ReasonMark watermarks only the answering phase using semantic vector from reasoning trace."} {"id": 13218, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Knowledge-grounded memory evolution for remote sensing agents"}, {"area": "science-data", "role": "benchmark", "why": "Systematic evaluation of RS agents"}], "pattern": "vary-scaffold", "fixes": ["base LLM"], "varies": ["memory mechanism"], "benchmarks": [], "results": [], "workflow": ["ground knowledge", "distill failures", "reuse experience"], "summary": "Knowledge-enhanced memory evolution using hierarchical knowledge and failure-aware experience for remote sensing agents."} {"id": 14254, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Deployed multi-agent credit risk model refreshing with DPO"}], "pattern": "vary-scaffold", "fixes": ["credit workflow"], "varies": ["agent orchestration"], "benchmarks": [], "results": [], "workflow": ["orchestrate specialists", "retrieve workflows", "DPO align"], "summary": "Deployed multi-agent framework automating credit model refresh with guardrails, retrieval and DPO."} {"id": 15854, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "analysis", "why": "Effect of reviewer guideline design on automated peer review"}], "pattern": "vary-conditions", "fixes": ["LLM reviewers"], "varies": ["guidelines"], "benchmarks": [], "results": [], "workflow": [], "summary": "Analyzes how official versus reviewer-imitating guidelines affect LLM peer review quality."} {"id": 12866, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Vision-grounded RL with MLLM critic for front-end coding"}], "pattern": "fix-env-vary-agent", "fixes": ["web coding tasks"], "varies": ["RL reward", "critic"], "benchmarks": [], "results": [], "workflow": ["generate", "visual critique", "refine", "RL"], "summary": "Agentic RL for web coding using MLLM visual critic, zero reward for invalid renders and forced optimization."} {"id": 13375, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Self-reflective generative reward model"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["reward model"], "benchmarks": [], "results": [], "workflow": ["joint train", "self-reflect", "select analysis"], "summary": "Generative reward model using self-reflection to assess analysis quality and improve preference modeling."} {"id": 16998, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Plan-Critic guided decoding for autoregressive text-to-audio"}], "pattern": "n/a", "fixes": ["generator"], "varies": ["decoding"], "benchmarks": [], "results": [], "workflow": ["score prefixes", "prune", "reallocate compute"], "summary": "Critic predicts final quality from prefixes to prune and guide AR audio generation."} {"id": 16132, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Evidence-contract multi-agent debate for fact-checking"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["debate protocol"], "benchmarks": [], "results": [], "workflow": ["debate subclaims", "scrutinize", "judge"], "summary": "Multi-agent debate with evidence contracts and mechanized scrutiny for provenance-grounded fact-checking."} {"id": 7142, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "data", "why": "Automated CAPTCHA trajectory generation"}, {"area": "web-gui", "role": "train-agent", "why": "Self-corrective training of native GUI agent"}], "pattern": "vary-data", "fixes": ["VLM base"], "varies": ["training data"], "benchmarks": [], "results": [], "workflow": ["build CAPTCHAs", "generate trajectories", "self-correct training"], "summary": "CAPTCHA-capable native GUI agent via synthetic CAPTCHA system, trajectory pipeline and self-corrective training."} {"id": 14980, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Uncertainty-driven adaptive planning and MCTS reasoning for web agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["planning", "reasoning"], "benchmarks": [], "results": [], "workflow": ["adaptive plan", "MCTS", "uncertainty"], "summary": "Web agent framework using task and action uncertainty for adaptive planning and MCTS."} {"id": 2324, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "analysis", "why": "Bias of AI programming assistants toward mainstream languages"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["languages", "frameworks"], "benchmarks": [], "results": [], "workflow": [], "summary": "Large-scale study showing AI coding assistants perform better on mainstream languages, suggesting a Matthew effect."} {"id": 9331, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Evolutionary search for multi-agent workflows"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["workflow design"], "benchmarks": [], "results": [], "workflow": ["evolve workflows", "role evolution", "curriculum"], "summary": "Evolutionary framework with strategies, role evolution and curriculum to design agentic workflows."} {"id": 16911, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "analysis", "why": "Empirical realism of LLM agents imitating social network users"}, {"area": "social-sim", "role": "eval-method", "why": "Formal framework and validation for social simulation"}], "pattern": "vary-conditions", "fixes": ["LLMs"], "varies": ["imitation approaches", "language"], "benchmarks": [], "results": [], "workflow": [], "summary": "Tests approaches to imitate social network users in English and German; argues for empirical realism validation."} {"id": 6912, "in_agent_scope": false, "types": [{"area": "safety/privacy", "role": "defense", "why": "Latent-space unlearning unifying deletion and refusal"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["unlearning method"], "benchmarks": [], "results": [], "workflow": [], "summary": "D2 unlearning restricts latent response distribution to erase knowledge while distinguishing retained knowledge."} {"id": 13763, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "benchmark", "why": "Decision-aware compositional mobile GUI benchmark"}, {"area": "web-gui", "role": "eval-method", "why": "Weighted LCS and decision accuracy metrics"}], "pattern": "fix-agent-vary-env", "fixes": ["7 agents"], "varies": ["tasks", "metrics"], "benchmarks": [{"name": "DAC-Bench", "purpose": "compositional mobile GUI tasks", "is_new": true}], "results": [], "workflow": [], "summary": "Benchmark of 830 mobile episodes with branching structures and new progress metrics."} {"id": 18644, "in_agent_scope": false, "types": [{"area": "other/ml", "role": "method", "why": "Emergent communication of grounded vocabularies among agents"}], "pattern": "n/a", "fixes": ["datasets"], "varies": ["method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Methodology for agent populations to converge on grounded linguistic conventions in continuous feature spaces."} {"id": 7620, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Training-free hierarchical memory guardrail for agent safety"}], "pattern": "vary-scaffold", "fixes": ["threat model"], "varies": ["guardrail"], "benchmarks": [], "results": [], "workflow": ["generate adversarial", "extract rules", "inject via memory"], "summary": "Memory-augmented guardrail extracting context-aware defense rules to balance safety and over-refusal."} {"id": 5604, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "environment", "why": "Synthetic training tasks for coding agents"}, {"area": "code-swe", "role": "data", "why": "Principles for auxiliary task design"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["training tasks"], "benchmarks": [{"name": "SWE-Bench", "purpose": "contrast with training tasks", "is_new": false}], "results": [], "workflow": ["decompose trajectories", "design tasks", "train"], "summary": "Hybrid-Gym: scalable synthetic tasks teaching transferable skills to coding agents."} {"id": 13544, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Attention-guided process supervision for efficient reasoning"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["RL reward"], "benchmarks": [], "results": [], "workflow": ["find heads", "step credit", "RL"], "summary": "Uses attention heads for step-level credit assignment to shorten reasoning in RLVR."} {"id": 18675, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "survey", "why": "Survey of instructional text for AI agents"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Systematic review of 181 papers on instructional text representations, tasks and challenges."} {"id": 7818, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "Agentic RL tool-calling video reasoning model"}, {"area": "domain-agents", "role": "data", "why": "ClinVideoSuite supervision data"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["training"], "benchmarks": [{"name": "ClinVideoSuite", "purpose": "supervision data", "is_new": true}], "results": [], "workflow": ["coarse search", "targeted tool calls", "verify"], "summary": "Tool-using clinical video reasoning model with coarse-to-fine evidence seeking trained via agentic RL."} {"id": 107, "in_agent_scope": false, "types": [{"area": "train/peft", "role": "method", "why": "Proxy trained on textual knowledge from frozen LLM"}], "pattern": "n/a", "fixes": ["frozen LLM"], "varies": ["proxy"], "benchmarks": [], "results": [], "workflow": ["elicit knowledge", "train proxy"], "summary": "Knowledge-guided proxy tuning without access to LLM probability distributions."} {"id": 1837, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Non-autoregressive multi-turn agentic dialogue synthesis"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["data generation"], "benchmarks": [], "results": [], "workflow": ["init skeleton", "refine", "verify"], "summary": "Generates multi-turn function-calling dialogues via coarse initialization, iterative refinement and offline verification."} {"id": 13948, "in_agent_scope": false, "types": [{"area": "train/pretrain", "role": "method", "why": "Pretraining with language learning tasks"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["pretraining data"], "benchmarks": [], "results": [], "workflow": [], "summary": "L2T adds structured language-learning tasks to pretraining to improve linguistic competence."} {"id": 18530, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Coding framework for collaborative problem-solving dialogue"}], "pattern": "n/a", "fixes": ["datasets"], "varies": ["coding scheme"], "benchmarks": [], "results": [], "workflow": [], "summary": "Two-layer coding scheme for analyzing human-AI and multi-agent collaborative dialogue across nine datasets."} {"id": 12440, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "World model of guideline execution for rule following"}], "pattern": "fix-env-vary-agent", "fixes": ["guidelines"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": [], "summary": "Models guideline execution as environment with explicit rule state to improve rule following."} {"id": 15060, "in_agent_scope": false, "types": [{"area": "safety/privacy", "role": "defense", "why": "Representation-guided LoRA unlearning"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["unlearning method"], "benchmarks": [], "results": [], "workflow": [], "summary": "ReGLU uses representation geometry for LoRA initialization to improve forget-retain trade-off."} {"id": 5747, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "eval-method", "why": "Sensitivity of tool-calling evaluation to implementation choices"}, {"area": "tool-use", "role": "analysis", "why": "Efficiency of RL training for tool calling"}], "pattern": "vary-conditions", "fixes": ["models"], "varies": ["eval settings", "RL setups"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows tool-calling results are sensitive to seeds and templates and studies RL training efficiency."} {"id": 18500, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Real-time Hinglish spoken dialogue system with parallel TTS"}], "pattern": "n/a", "fixes": ["ASR-LLM-TTS"], "varies": ["pipeline"], "benchmarks": [], "results": [], "workflow": ["stream TTS", "chunk boundaries", "abort on barge-in"], "summary": "Production voice system with parallel TTS streaming, code-switch chunking and barge-in handling."} {"id": 8612, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Anchor-shaped exploration replacing KL regularization"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["RLVR objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "SAGE shapes anchors for exploration in RLVR to improve pass@k."} {"id": 15437, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Multi-agent debate video retrieval"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval design"], "benchmarks": [], "results": [], "workflow": ["parse videos", "nominate candidates", "debate, verify"], "summary": "Multi-agent video retrieval with structured semantic library, planner and logic-aware debate."} {"id": 16866, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "VLM document navigation trained with Episodic GRPO"}, {"area": "web-gui", "role": "data", "why": "Chain of Scroll trajectories"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["training"], "benchmarks": [], "results": [], "workflow": ["generate trajectories", "scroll selectively", "EGRPO"], "summary": "VLM navigates documents via Chain of Scroll, trained on synthetic trajectories with episodic GRPO."} {"id": 11578, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Confidence-guided leader-supporter adaptive debate"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["debate"], "benchmarks": [], "results": [], "workflow": [], "summary": "LASE selectively engages debate in non-neutral regimes with leader-supporter structure."} {"id": 12383, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Training-free evidence-grounded agents for motion time-series"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["distill knowledge", "retrieve evidence", "iterate cues"], "summary": "Training-free agentic framework using distilled textual knowledge and retrieval for activity recognition."} {"id": 14139, "in_agent_scope": false, "types": [{"area": "multiling/lowres", "role": "method", "why": "Thesis on uncertainty-guided code-switched ASR"}], "pattern": "n/a", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Thesis proposal for self-adaptive, uncertainty-guided ASR on African code-switched speech."} {"id": 245, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "train-agent", "why": "RL-trained multi-agent medical collaboration"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["RL training"], "benchmarks": [], "results": [], "workflow": ["triage", "specialists", "RL attending"], "summary": "RL-trained triage and attending agents for dynamic multi-agent medical VQA."} {"id": 12203, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Error-localized policy optimization for tool-integrated reasoning"}, {"area": "planning-decision", "role": "train-agent", "why": "Step credit assignment"}], "pattern": "fix-env-vary-agent", "fixes": ["TIR benchmarks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": ["binary-search rollouts", "advantage attribution", "adaptive clipping"], "summary": "Localizes first irrecoverable step via rollout trees for fine-grained credit in TIR RL."} {"id": 15551, "in_agent_scope": false, "types": [{"area": "reason/planning", "role": "method", "why": "Regret-based rollback for temporal KGQA reasoning"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["reasoning framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Energy-inspired framework evaluating step potential and rolling back errors in temporal KGQA."} {"id": 16248, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "eval-method", "why": "Sentient agent judge for social cognition"}, {"area": "social-sim", "role": "benchmark", "why": "Supportive-dialogue scenarios"}], "pattern": "vary-evaluation", "fixes": ["tested LLMs"], "varies": ["evaluation protocol"], "benchmarks": [{"name": "BLRI", "purpose": "validation of emotion score", "is_new": false}], "results": [], "workflow": [], "summary": "Simulated-human agent tracks emotion to evaluate LLM social cognition in multi-turn dialogue."} {"id": 11213, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "train-agent", "why": "Procedure-aware mobile agent with meta reasoning"}], "pattern": "fix-env-vary-agent", "fixes": ["mobile tasks"], "varies": ["procedural knowledge", "training"], "benchmarks": [], "results": [], "workflow": ["obtain procedure", "condition actions"], "summary": "Trains grounded mobile agent conditioned on natural-language procedures from LLM or search."} {"id": 2800, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Adversarial summarization and quiz agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": ["generate summary", "quiz", "review"], "summary": "Multi-agent summarization with quiz generators and examinee validating summary coverage."} {"id": 11774, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "scaffold", "why": "Grammar-based search over multi-agent systems"}], "pattern": "vary-scaffold", "fixes": ["backbone LLMs"], "varies": ["search method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Structured composable-component search for multi-agent systems outperforming free-form LLM search."} {"id": 9080, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "train-agent", "why": "Self-evolving tool-integrated vision-language agent"}, {"area": "memory-selfevolve", "role": "train-agent", "why": "Self-evolution with tool-grounded self-repair"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["self-evolution training"], "benchmarks": [], "results": [], "workflow": ["reason with tools", "self-evaluate", "self-repair"], "summary": "Self-evolving VL agent using tools in reasoning, self-evaluation and self-repair."} {"id": 10312, "in_agent_scope": false, "types": [{"area": "safety/security", "role": "method", "why": "Decoding-guided watermark for diffusion LMs"}], "pattern": "n/a", "fixes": ["dLLMs"], "varies": ["unmasking order"], "benchmarks": [], "results": [], "workflow": [], "summary": "dgMARK steers unmasking order toward parity-satisfying tokens for watermarking diffusion LMs."}