Files
ai-taxonomy-literature-anal…/data/labels/o12.jsonl
T
zzyang d6fc7c1dfa Add types-of-work classification, benchmarks and score pages
- Work type = area x role (fixed vs varied), vocabulary in scripts/worktypes.json
- LLM-read labels for 2,678 agent-related papers (data/labels), rule labels for the rest
- Per-type synthesis notes (why-same, sub-variants, benchmarks, workflow)
- Pages: /work-types matrix and detail (timeline, benchmarks, workflow, score charts), /benchmarks
- build_worktypes.py adds work-type tables to data/papers.db

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

61 lines
24 KiB
JSON

{"id": 5932, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "analysis", "why": "Studies VLM visual preferences via controlled image perturbations in choice tasks"}], "pattern": "vary-conditions", "fixes": ["VLMs"], "varies": ["image edits", "choice conditions"], "benchmarks": [], "results": [], "workflow": [], "summary": "Infers VLM visual utility through revealed preference over systematically edited images."}
{"id": 13258, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Training-free iterative prompt compression for multi-hop QA"}], "pattern": "vary-scaffold", "fixes": ["LLM", "benchmarks"], "varies": ["compression method"], "benchmarks": [], "results": [], "workflow": ["Decompose documents into segments", "Evaluate question answerability", "Generate follow-up questions", "Iterate compression"], "summary": "Training-free iterative prompt compression using answerability checks and follow-up questions for multi-hop QA."}
{"id": 14582, "in_agent_scope": false, "types": [{"area": "efficiency/training", "role": "method", "why": "Hybrid zeroth- and first-order fine-tuning with importance-guided tensor selection"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["optimizer"], "benchmarks": [], "results": [], "workflow": [], "summary": "Hybrid ZO/FO LLM fine-tuning that selects important tensors for first-order updates."}
{"id": 11929, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "RL-trained agents with history rendered as compact image"}, {"area": "planning-decision", "role": "train-agent", "why": "Optical caching for scalable multi-turn RL rollouts"}], "pattern": "vary-scaffold", "fixes": ["environments"], "varies": ["history representation"], "benchmarks": [], "results": [], "workflow": [], "summary": "Compresses multi-turn agent history into rendered images with segment optical caching for RL agents."}
{"id": 14789, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "method", "why": "PCA sweep to choose components in semantic differential analysis"}], "pattern": "n/a", "fixes": [], "varies": ["dimensionality selection"], "benchmarks": [], "results": [], "workflow": [], "summary": "PCA sweep procedure for selecting retained components in Supervised Semantic Differential, with AI discourse case study."}
{"id": 1651, "in_agent_scope": false, "types": [{"area": "align/pref", "role": "method", "why": "Scalable oversight using partitioned complementary-label human supervision"}], "pattern": "n/a", "fixes": [], "varies": ["supervision signal"], "benchmarks": [], "results": [], "workflow": [], "summary": "Scalable oversight framework using weak complementary labels from narrow-domain human experts."}
{"id": 10564, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "Trajectory-level uncertainty metric for tool-agent-user interaction"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["uncertainty metric"], "benchmarks": [], "results": [], "workflow": [], "summary": "TRACER aggregates trajectory risk signals to estimate uncertainty of tool-using agents."}
{"id": 5606, "in_agent_scope": false, "types": [{"area": "arch/multimodal", "role": "train-agent", "why": "Vision-aware self-improvement training for MLLM reasoning"}], "pattern": "vary-data", "fixes": ["MLLM"], "varies": ["self-improvement training"], "benchmarks": [], "results": [], "workflow": [], "summary": "Vision-aware self-improvement training addressing data imbalance and language prior bias in MLLMs."}
{"id": 16894, "in_agent_scope": false, "types": [{"area": "apps/dialogue", "role": "analysis", "why": "Tests linguistic convergence of models to users across dialogue corpora"}], "pattern": "vary-conditions", "fixes": ["methods"], "varies": ["models", "corpora"], "benchmarks": [], "results": [], "workflow": [], "summary": "Studies whether sixteen LLMs converge to user linguistic patterns compared to humans."}
{"id": 15772, "in_agent_scope": false, "types": [{"area": "apps/science", "role": "application", "why": "Metadata-aware RAG pipeline for cancer patient QA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["retrieval pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "CoMeta, hybrid Boolean and semantic retrieval pipeline for cancer patient question answering."}
{"id": 16310, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "analysis", "why": "Profiles token flows and latency across MCP workflow stages"}], "pattern": "vary-conditions", "fixes": ["MCP agents"], "varies": ["pipeline stage"], "benchmarks": [], "results": [], "workflow": [], "summary": "ProMCP profiling framework attributing token and latency costs in MCP-based agents."}
{"id": 12042, "in_agent_scope": false, "types": [{"area": "align/rlvr", "role": "method", "why": "Memory-efficient RL for diffusion LLMs via boundary-guided optimization"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "Boundary-guided policy optimization reducing memory for RL on diffusion language models."}
{"id": 10003, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Crafts messages evading embedding-based defenses in multi-agent systems"}, {"area": "agent-safety", "role": "defense", "why": "Rethinks defenses for multi-agent safety"}], "pattern": "vary-conditions", "fixes": ["MAS"], "varies": ["attack messages", "defense"], "benchmarks": [], "results": [], "workflow": [], "summary": "Shows embedding-based MAS defenses can be bypassed and rethinks multi-agent safety."}
{"id": 16263, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "train-agent", "why": "Aligns proactive team agents toward long-term team sustainability"}], "pattern": "fix-env-vary-agent", "fixes": ["team setting"], "varies": ["task allocation objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "Addresses collaborative myopia in proactive multi-agent scientific teams via strategic equilibrium."}
{"id": 5040, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "attack", "why": "Adaptive red-teaming agent with plug-and-play attacks on VLMs"}], "pattern": "vary-conditions", "fixes": ["target VLM"], "varies": ["attack strategies"], "benchmarks": [], "results": [], "workflow": [], "summary": "ARMs automatically optimizes multimodal attack strategies to red-team VLMs."}
{"id": 9415, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "environment", "why": "Code-generated curriculum of environments in open-ended worlds"}], "pattern": "fix-agent-vary-env", "fixes": ["agent"], "varies": ["generated environment curriculum"], "benchmarks": [], "results": [], "workflow": [], "summary": "DiCode uses foundation models to generate code environments forming a learnable curriculum."}
{"id": 18157, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "benchmark", "why": "New Swedish professional speech corpus"}], "pattern": "n/a", "fixes": [], "varies": ["corpus"], "benchmarks": [], "results": [], "workflow": [], "summary": "Storspigg-TBI, legally cleared parallel Swedish read-speech corpus from 99 narrators."}
{"id": 3864, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Closed-loop multi-agent system for antimicrobial peptide design"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["multi-agent workflow"], "benchmarks": [], "results": [], "workflow": [], "summary": "MAC-AMP multi-agent LLM system for multi-objective antimicrobial peptide design."}
{"id": 13614, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "eval-method", "why": "Simulation framework evaluating multimodal safety of generative agents"}], "pattern": "vary-evaluation", "fixes": ["agents"], "varies": ["safety metrics"], "benchmarks": [], "results": [], "workflow": [], "summary": "Reproducible social simulation framework to evaluate multimodal generative agent safety."}
{"id": 379, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "benchmark", "why": "Benchmark for complex agentic feature development with executable evaluation"}], "pattern": "fix-agent-vary-env", "fixes": ["agents"], "varies": ["tasks"], "benchmarks": [{"name": "FeatureBench", "purpose": "feature development evaluation", "is_new": true}], "results": [], "workflow": [], "summary": "FeatureBench evaluates coding agents on complex feature development with automated, updatable tasks."}
{"id": 15638, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Multilingual RAG legal assistant for Bangladesh"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "Mina, Bengali legal assistant using RAG chain-of-tools for drafts and citations."}
{"id": 16157, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL for agentic search that recognizes reasoning boundaries"}], "pattern": "fix-env-vary-agent", "fixes": ["search environment"], "varies": ["RL objective"], "benchmarks": [], "results": [], "workflow": [], "summary": "BAPO trains search agents to admit 'I don't know' when evidence is insufficient."}
{"id": 16852, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "Distills multi-agent coding into single 7B model"}], "pattern": "fix-env-vary-agent", "fixes": ["benchmarks"], "varies": ["distillation"], "benchmarks": [], "results": [], "workflow": [], "summary": "MapCoder-Lite distills multi-agent coding reasoning into one small LLM."}
{"id": 17464, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "analysis", "why": "Simulates uncooperative behaviors destabilizing LLM multi-agent systems"}], "pattern": "vary-conditions", "fixes": ["MAS"], "varies": ["uncooperative behaviors"], "benchmarks": [], "results": [], "workflow": [], "summary": "Taxonomy and simulation of uncooperative agent behaviors in collaborative resource management."}
{"id": 1117, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Bayesian adversarial multi-agent low-code platform for scientific code"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["multi-agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Bayesian adversarial multi-agent framework for reliable AI4S code generation."}
{"id": 15251, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "train-agent", "why": "Parameter fusion via directional consensus for agent continual learning"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["continual learning method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Agent-Dice merges knowledge updates to avoid forgetting in continually learning agents."}
{"id": 5510, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Proactive agent clarifying ambiguity for text-to-CAD generation"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agentic framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "ProCAD resolves under-specified CAD descriptions via clarification before generating CadQuery."}
{"id": 5874, "in_agent_scope": true, "types": [{"area": "tool-use", "role": "data", "why": "Diverse verifiable tool-use task synthesis by executing real tools first"}], "pattern": "vary-data", "fixes": ["training recipe"], "varies": ["synthesized tasks"], "benchmarks": [], "results": [], "workflow": [], "summary": "DIVE synthesizes diverse executable tool-use tasks to improve generalization."}
{"id": 12794, "in_agent_scope": false, "types": [{"area": "reason/cot", "role": "method", "why": "Accumulates reusable metacognitive skills across instances"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["meta-reasoning"], "benchmarks": [], "results": [], "workflow": [], "summary": "Metacognitive consolidation lets LLMs accumulate reusable meta-reasoning skills."}
{"id": 14395, "in_agent_scope": true, "types": [{"area": "science-data", "role": "scaffold", "why": "Composes chemistry tools as building-block agents"}], "pattern": "vary-scaffold", "fixes": ["tools", "LLM"], "varies": ["tool coordination"], "benchmarks": [], "results": [], "workflow": [], "summary": "ChemAmp dynamically coordinates chemistry tools as composable agents."}
{"id": 5937, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "train-agent", "why": "Hierarchical agent over heterogeneous action spaces with chain-of-action"}], "pattern": "fix-env-vary-agent", "fixes": ["environments"], "varies": ["agent architecture"], "benchmarks": [], "results": [], "workflow": [], "summary": "DeepHA unifies semantic skills and motor controls via chain-of-action reasoning."}
{"id": 5903, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "defense", "why": "Adversarial co-evolution MARL internalizing safety in multi-agent systems"}], "pattern": "vary-conditions", "fixes": ["threat model"], "varies": ["safety training"], "benchmarks": [], "results": [], "workflow": [], "summary": "AdvEvo-MARL co-evolves attackers and defenders to internalize safety in LLM multi-agent systems."}
{"id": 10269, "in_agent_scope": false, "types": [{"area": "eval/bench", "role": "benchmark", "why": "Synthetic in-context induction evaluation with controllable complexity"}], "pattern": "fix-agent-vary-env", "fixes": ["models"], "varies": ["task complexity axes"], "benchmarks": [], "results": [], "workflow": [], "summary": "Modular synthetic evaluation of inductive generalization with controllable complexity."}
{"id": 12789, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Semantic tree table serialization for table QA"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["table serialization"], "benchmarks": [], "results": [], "workflow": [], "summary": "ASTRA builds adaptive semantic trees for complex table question answering."}
{"id": 9654, "in_agent_scope": false, "types": [{"area": "efficiency/training", "role": "method", "why": "Heavy-tail guided layerwise learning rates"}], "pattern": "n/a", "fixes": ["benchmarks"], "varies": ["learning rates"], "benchmarks": [], "results": [], "workflow": [], "summary": "Layerwise learning rates assigned using heavy-tailed self-regularization spectral analysis."}
{"id": 15095, "in_agent_scope": false, "types": [{"area": "apps/domain", "role": "analysis", "why": "Studies factors affecting LLMs and RLLMs in financial QA"}], "pattern": "vary-conditions", "fixes": ["methods"], "varies": ["models", "methods"], "benchmarks": [], "results": [], "workflow": [], "summary": "Systematic assessment of methods across five LLMs and four reasoning LLMs in financial QA."}
{"id": 9585, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "LLM agents construct clinical scoring systems"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent workflow"], "benchmarks": [], "results": [], "workflow": [], "summary": "LLM agents automatically build interpretable clinical scoring systems."}
{"id": 14350, "in_agent_scope": false, "types": [{"area": "train/data", "role": "method", "why": "Multi-agent LLM GAN for synthetic tabular data"}], "pattern": "n/a", "fixes": [], "varies": ["generator"], "benchmarks": [], "results": [], "workflow": [], "summary": "MALLM-GAN emulates GAN with LLM agents to synthesize tabular data."}
{"id": 14012, "in_agent_scope": true, "types": [{"area": "multi-agent", "role": "environment", "why": "Toolkit for probing emergent risks in multi-agent systems"}, {"area": "agent-safety", "role": "eval-method", "why": "Controlled measurement of system-level risks"}], "pattern": "vary-conditions", "fixes": ["agents"], "varies": ["topology", "protocol"], "benchmarks": [], "results": [], "workflow": [], "summary": "RiskLab toolkit for controlled studies of emergent risks in LLM multi-agent systems."}
{"id": 15906, "in_agent_scope": true, "types": [{"area": "social-sim", "role": "scaffold", "why": "Adaptive multi-agent framework with scene manager for role-playing"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["interaction framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "AdaMARP uses immersive message format and Scene Manager for multi-character role-playing."}
{"id": 10223, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "RL-trained context folding for long-horizon agents"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["context management", "RL"], "benchmarks": [], "results": [], "workflow": [], "summary": "Context Folding with FoldGRPO trains agents to branch and fold sub-trajectories."}
{"id": 16641, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "DeBERTa token classification for MWE identification"}], "pattern": "n/a", "fixes": [], "varies": ["model"], "benchmarks": [{"name": "CoAM", "purpose": "MWE identification evaluation", "is_new": false}], "results": [], "workflow": [], "summary": "Binary token-level DeBERTa approach for multiword expression identification."}
{"id": 13635, "in_agent_scope": false, "types": [{"area": "apps/classic", "role": "method", "why": "Confidence-routed expert debate for sexism detection"}], "pattern": "n/a", "fixes": [], "varies": ["detection method"], "benchmarks": [], "results": [], "workflow": [], "summary": "Two-stage training and expert debate routing for sexism detection."}
{"id": 9882, "in_agent_scope": true, "types": [{"area": "embodied-nav", "role": "scaffold", "why": "Runtime-efficient multi-agent embodied planning with semantic memory"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["planning framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "Mosaic reduces failed actions in LLM multi-agent embodied planning."}
{"id": 11364, "in_agent_scope": false, "types": [{"area": "train/sft", "role": "data", "why": "Diagnostic-guided instruction data curation"}], "pattern": "vary-data", "fixes": ["SFT recipe"], "varies": ["selected data"], "benchmarks": [], "results": [], "workflow": [], "summary": "Cure-SFT separates redundancy from quality defects when curating instruction data."}
{"id": 12517, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Hierarchical multimodal retrieval for document QA"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["retrieval"], "benchmarks": [], "results": [], "workflow": [], "summary": "HiKEY hierarchical multimodal retrieval for open-domain document QA."}
{"id": 14911, "in_agent_scope": true, "types": [{"area": "web-gui", "role": "scaffold", "why": "Dynamic task decomposition and re-planning for web tasks"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["planning scaffold"], "benchmarks": [], "results": [], "workflow": [], "summary": "WebDART decomposes web objectives into subtasks and continuously re-plans."}
{"id": 14520, "in_agent_scope": false, "types": [{"area": "knowledge/rag", "role": "method", "why": "Subgraph-guided logical form generation for KBQA"}], "pattern": "n/a", "fixes": ["LLM"], "varies": ["retrieval", "generation"], "benchmarks": [], "results": [], "workflow": [], "summary": "SELF-KBQA conditions logical form generation on structurally aligned subgraphs."}
{"id": 12431, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "scaffold", "why": "Profiling-guided workflow for parallel code generation agent"}], "pattern": "vary-scaffold", "fixes": ["Codex agent"], "varies": ["workflow"], "benchmarks": [], "results": [], "workflow": [], "summary": "ParaCodex turns a coding agent into an OpenMP GPU offload system with profiling-guided refinement."}
{"id": 15540, "in_agent_scope": true, "types": [{"area": "search-research", "role": "benchmark", "why": "Long-form multimodal report generation tasks"}], "pattern": "fix-agent-vary-env", "fixes": ["LLMs", "coding agents"], "varies": ["tasks"], "benchmarks": [{"name": "AnalystBench", "purpose": "report generation", "is_new": true}], "results": [], "workflow": [], "summary": "AnalystBench evaluates LLMs and coding agents on professional report generation."}
{"id": 12342, "in_agent_scope": true, "types": [{"area": "domain-agents", "role": "application", "why": "Two collaborative LLM agents for explainable ICD coding"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["agent framework"], "benchmarks": [], "results": [], "workflow": [], "summary": "ICDAGENT uses collaborating agents for explainable medical coding."}
{"id": 13979, "in_agent_scope": false, "types": [{"area": "eval/judge", "role": "eval-method", "why": "Metric-human alignment for QA in Indic languages"}], "pattern": "vary-evaluation", "fixes": ["systems"], "varies": ["metrics"], "benchmarks": [], "results": [], "workflow": [], "summary": "Revisits QA evaluation metrics for low-resource Indic languages against human judgment."}
{"id": 16646, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "survey", "why": "Position arguing for process-level agent evaluation"}, {"area": "agent-safety", "role": "eval-method", "why": "Evaluate steps not just outcomes"}], "pattern": "vary-evaluation", "fixes": [], "varies": [], "benchmarks": [], "results": [], "workflow": [], "summary": "Position paper arguing agents should be evaluated on process, not only outcomes."}
{"id": 16335, "in_agent_scope": true, "types": [{"area": "science-data", "role": "application", "why": "Fine-tuned multi-agent framework for safer lipid design"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["domain pipeline"], "benchmarks": [], "results": [], "workflow": [], "summary": "LipoAgent coordinates fine-tuned agents with toxicity-first prediction for lipid discovery."}
{"id": 12453, "in_agent_scope": true, "types": [{"area": "memory-selfevolve", "role": "scaffold", "why": "Quasi-symbolic abstractions for experience reuse in evolving agents"}], "pattern": "vary-scaffold", "fixes": ["LLM"], "varies": ["abstraction memory"], "benchmarks": [], "results": [], "workflow": [], "summary": "EvA learns and reuses quasi-symbolic reasoning abstractions for agent adaptation."}
{"id": 11802, "in_agent_scope": true, "types": [{"area": "search-research", "role": "train-agent", "why": "RL for conversational reasoning and retrieval"}], "pattern": "fix-env-vary-agent", "fixes": ["CQA benchmarks"], "varies": ["RL reward"], "benchmarks": [], "results": [], "workflow": [], "summary": "ChatR1 interleaves search and reasoning across dialogue turns trained by RL."}
{"id": 3084, "in_agent_scope": true, "types": [{"area": "planning-decision", "role": "train-agent", "why": "Tree-search-based grouped RL for agents"}], "pattern": "fix-env-vary-agent", "fixes": ["tasks"], "varies": ["RL algorithm"], "benchmarks": [], "results": [], "workflow": [], "summary": "Tree-GRPO shares prefixes via tree search to densify supervision in agent RL."}
{"id": 6568, "in_agent_scope": true, "types": [{"area": "code-swe", "role": "train-agent", "why": "RL for Triton kernel generation with reward hacking checks"}, {"area": "code-swe", "role": "environment", "why": "KernelGYM distributed GPU environment"}], "pattern": "fix-env-vary-agent", "fixes": ["environment"], "varies": ["RL training"], "benchmarks": [], "results": [], "workflow": [], "summary": "Dr. Kernel studies RL for kernel generation using KernelGYM and anti-hacking checks."}
{"id": 5869, "in_agent_scope": true, "types": [{"area": "agent-safety", "role": "benchmark", "why": "Cyber threat investigation benchmark in Azure SQL environment"}], "pattern": "fix-agent-vary-env", "fixes": ["LLM agents"], "varies": ["tasks"], "benchmarks": [{"name": "ExCyTIn-Bench", "purpose": "cyber investigation", "is_new": true}], "results": [], "workflow": [], "summary": "ExCyTIn-Bench evaluates LLM agents on security investigation questions from logs."}
{"id": 12365, "in_agent_scope": false, "types": [{"area": "train/distill", "role": "data", "why": "Router-guided personalized teacher selection for synthetic data"}], "pattern": "vary-data", "fixes": ["student training"], "varies": ["teacher data"], "benchmarks": [], "results": [], "workflow": [], "summary": "PerSyn routes each prompt to optimal teacher for student-tailored synthetic data."}