Paper review 2026-08-07: 6 candidates

[github-actions]

Reply with commands, one per line:
/approve all · /approve 1,3-5 · /reject 2 · /edit 3 category=world_terminal tags=benchmark (edit implies approve; tags=- clears tags). Valid category keys: see taxonomy.json.

1. DoctorAgents: an agentic framework to iteratively refine AutoML pipeline for small clinical temporal data

Ruilin Wang, Bo-Hong Wang, Elizabeth Kourbatski, et al. · arXiv 2026/08 · paper
proposed: world_research · tags: none

DoctorAgents is an agentic LLM framework with specialized agents for generating, validating, and refining AutoML pipelines on small clinical temporal datasets, using textual gradient descent instead of exhaustive search. It outperforms established AutoML baselines while yielding more interpretable pipelines.
_{An ML-engineering agent that builds and tunes training pipelines through iterative interaction fits world_research's MLE-bench-style inclusion.}

2. Towards a Risk Assessment of Malicious Skill Files in Coding Agents

Rui Yang, Michael Fu, Kla Tantithamthavorn, et al. · arXiv 2026/08 · paper
proposed: world_terminal · tags: benchmark

Studies how malicious 'skill files' can hide shell commands that hijack terminal-native coding agents like Gemini CLI and Qwen Code, releasing an adversarial benchmark of 2,826 skills mapped to MITRE ATT&CK tactics. Large-scale evaluation shows both agents are exploited in the majority of runs with rare safety recognition.
_{Offensive-security exploitation of shell/terminal-acting coding agents falls under world_terminal per its CTF/pentest boundary.}

3. Recursive Synthesis for Long-Horizon Terminal Tasks

Zhongzhi Li, Yucheng Shi, Zongxia Li, et al. · arXiv 2026/08 · paper
proposed: world_terminal · tags: training-data

Presents RST, a recursive verified synthesis framework that scales up long-horizon terminal-agent training tasks by iteratively extending reference solutions and verifiers, producing 37K tasks used to fine-tune and RL-train terminal agents on Terminal-Bench.
_{Downstream task served is terminal-agent training/evaluation, so it routes to world_terminal with a training-data tag.}

4. Matching Matters: A Fair Quality-Efficiency Benchmark for Command-Line Agents

Han Chi, Jiaxin Qi, Yan Cui, et al. · arXiv 2026/08 · paper
proposed: world_terminal · tags: benchmark

AgentMeter is a quality-efficiency benchmark for command-line-interface agents that jointly measures task quality, cost, and resource usage across language-model and CLI pairings, plus a fairer optimization variant for evaluation.
_{The agents act via CLI shells to complete tasks, making the terminal the agent's world per world_terminal.}

5. CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks

Fanzhe Meng, Guoxin Chen, Jiale Zhao, et al. · arXiv 2026/08 · paper
proposed: world_terminal · tags: training-data

CalibForge is an autonomous terminal-task synthesis system that uses adversarial solver calibration to construct appropriately challenging, verified training tasks for terminal agents. Models trained on its 5,431 calibrated tasks show large gains on Terminal-Bench 2.0, SWE-bench Pro, and Doc2Repo.
_{A resource paper generating training data whose downstream task is terminal-agent operation, matching the world_terminal leaf per the CLI-agent precedent.}

6. As You Wish: Mission Planning with Formal Verification using LLMs in Precision Agriculture

Marcos Abel Zuzuárregui, Stefano Carpin · arXiv 2026/08 · paper
proposed: world_physical · tags: none

Extends an LLM-based mission planner for precision-agriculture robots with feedback loops using linear temporal logic to verify natural-language mission plans against user specifications. Two LLMs handle specification and verification to reduce ambiguity and bias in autonomous robot task planning.
_{Generated mission plans are executed as robot control policies, so code/plans are the language of physical-world action.}

machine payload (do not edit)

[{"paper": {"id": "2608.05375", "title": "DoctorAgents: an agentic framework to iteratively refine AutoML pipeline for small clinical temporal data", "authors": ["Ruilin Wang", "Bo-Hong Wang", "Elizabeth Kourbatski", "Jun Bai", "Hegang Chen", "Ziyang Song", "Gilles Boire", "Marie Hudson", "Yue Li"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.05375", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "DoctorAgents is an agentic LLM framework with specialized agents for generating, validating, and refining AutoML pipelines on small clinical temporal datasets, using textual gradient descent instead of exhaustive search. It outperforms established AutoML baselines while yielding more interpretable pipelines.", "reason": "An ML-engineering agent that builds and tunes training pipelines through iterative interaction fits world_research's MLE-bench-style inclusion.", "source": "crawl"}, {"paper": {"id": "2608.05223", "title": "Towards a Risk Assessment of Malicious Skill Files in Coding Agents", "authors": ["Rui Yang", "Michael Fu", "Kla Tantithamthavorn", "Chetan Arora", "Joey Chua"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.05223", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "Studies how malicious 'skill files' can hide shell commands that hijack terminal-native coding agents like Gemini CLI and Qwen Code, releasing an adversarial benchmark of 2,826 skills mapped to MITRE ATT&CK tactics. Large-scale evaluation shows both agents are exploited in the majority of runs with rare safety recognition.", "reason": "Offensive-security exploitation of shell/terminal-acting coding agents falls under world_terminal per its CTF/pentest boundary.", "source": "crawl"}, {"paper": {"id": "2608.05466", "title": "Recursive Synthesis for Long-Horizon Terminal Tasks", "authors": ["Zhongzhi Li", "Yucheng Shi", "Zongxia Li", "Ruhan Wang", "Anhao Li", "Zixun Huang", "Junyao Yang", "Lei Ke", "Ninghao Liu", "Haitao Mi", "Leowei Liang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.05466", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["training-data"], "summary": "Presents RST, a recursive verified synthesis framework that scales up long-horizon terminal-agent training tasks by iteratively extending reference solutions and verifiers, producing 37K tasks used to fine-tune and RL-train terminal agents on Terminal-Bench.", "reason": "Downstream task served is terminal-agent training/evaluation, so it routes to world_terminal with a training-data tag.", "source": "crawl"}, {"paper": {"id": "2606.21140", "title": "Matching Matters: A Fair Quality-Efficiency Benchmark for Command-Line Agents", "authors": ["Han Chi", "Jiaxin Qi", "Yan Cui", "Baisheng Lai", "Jianqiang Huang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-06", "links": {"paper": "https://arxiv.org/abs/2606.21140", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "AgentMeter is a quality-efficiency benchmark for command-line-interface agents that jointly measures task quality, cost, and resource usage across language-model and CLI pairings, plus a fairer optimization variant for evaluation.", "reason": "The agents act via CLI shells to complete tasks, making the terminal the agent's world per world_terminal.", "source": "crawl"}, {"paper": {"id": "2608.06352", "title": "CalibForge: Adversarial Solver Calibration for Scaling Learnable Terminal Tasks", "authors": ["Fanzhe Meng", "Guoxin Chen", "Jiale Zhao", "Shuang Sun", "Zhiyu Lin", "Wayne Xin Zhao", "Ruihua Song", "Ji-Rong Wen", "Kai Jia"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-06", "links": {"paper": "https://arxiv.org/abs/2608.06352", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["training-data"], "summary": "CalibForge is an autonomous terminal-task synthesis system that uses adversarial solver calibration to construct appropriately challenging, verified training tasks for terminal agents. Models trained on its 5,431 calibrated tasks show large gains on Terminal-Bench 2.0, SWE-bench Pro, and Doc2Repo.", "reason": "A resource paper generating training data whose downstream task is terminal-agent operation, matching the world_terminal leaf per the CLI-agent precedent.", "source": "crawl"}, {"paper": {"id": "2606.18519", "title": "As You Wish: Mission Planning with Formal Verification using LLMs in Precision Agriculture", "authors": ["Marcos Abel Zuzuárregui", "Stefano Carpin"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2606.18519", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "Extends an LLM-based mission planner for precision-agriculture robots with feedback loops using linear temporal logic to verify natural-language mission plans against user specifications. Two LLMs handle specification and verification to reduce ambiguity and bias in autonomous robot task planning.", "reason": "Generated mission plans are executed as robot control policies, so code/plans are the language of physical-world action.", "source": "crawl"}]