Paper review 2026-07-24: 4 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. OpenForgeRL: Train Harness-native Agents in Any Environment

Xiao Yu, Baolin Peng, Ruize Xu, et al. · arXiv 2026/07 · paper
proposed: world_terminal · tags: none

OpenForgeRL is an RL training framework that lets researchers train harness-native agents (e.g. Claude Code, Codex, OpenClaw) end-to-end by proxying model calls and running rollouts in containerized environments. It is evaluated on both tool/CLI harness benchmarks and GUI browser/computer-use benchmarks.
_{Dominant framing is terminal/CLI coding harnesses (Claude Code, Codex) that act via code/shell commands, so it routes to world_terminal despite spanning GUI benchmarks too.}

2. Training Fast Robot Policies with Slow Foundation Models

Utsav Singh, Pramit Bhattacharyya, Vinay P. Namboodiri, et al. · arXiv 2026/07 · paper
proposed: world_physical · tags: none

VGRS uses LLMs to synthesize executable reward code from instructions, training lightweight robot policies, while a VLM diagnoses failure trajectories to iteratively densify rewards. Foundation models are used only at training time, enabling fast, low-cost deployment on real robots for navigation and manipulation.
_{Generated reward code is the instrument for training a policy that controls physical robots, matching the code-as-policy world_physical boundary.}

3. An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models

Ari Luna Rueda, Eike Cramer, Klaus Hellgardt, et al. · arXiv 2026/07 · paper
proposed: world_physical · tags: none

An LLM-driven workflow decomposes multi-variable process-control design into constrained code-generation steps (plant interface, controller specification, tuning), executing and validating generated artifacts with feedback-based repair. Demonstrated on a gas-preheater plant, improving closed-loop performance via Bayesian optimization.
_{Generated code serves as the control policy operating a physical process, making code the instrument of action rather than the deliverable.}

4. PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning

Alexis Fox, Junlin Wang, Paul Rosu, et al. · arXiv 2026/07 · paper
proposed: world_game · tags: none

PRO-LONG is a context-management framework for LLM agents that maintains a complete structured interaction log and uses coding-agent techniques to search it efficiently for long-horizon exploratory tasks. It improves performance and token efficiency over specialized harnesses on the full ARC-AGI-3 game set.
_{Benchmark routing to ARC-AGI-3's game environments makes this a code-as-tool agent acting in game worlds, per world_game.}

machine payload (do not edit)

[{"paper": {"id": "2607.21557", "title": "OpenForgeRL: Train Harness-native Agents in Any Environment", "authors": ["Xiao Yu", "Baolin Peng", "Ruize Xu", "Hao Zou", "Qianhui Wu", "Hao Cheng", "Wenlin Yao", "Nikhil Singh", "Zhou Yu", "Jianfeng Gao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.21557", "github": "", "website": ""}}, "category": "world_terminal", "tags": [], "summary": "OpenForgeRL is an RL training framework that lets researchers train harness-native agents (e.g. Claude Code, Codex, OpenClaw) end-to-end by proxying model calls and running rollouts in containerized environments. It is evaluated on both tool/CLI harness benchmarks and GUI browser/computer-use benchmarks.", "reason": "Dominant framing is terminal/CLI coding harnesses (Claude Code, Codex) that act via code/shell commands, so it routes to world_terminal despite spanning GUI benchmarks too.", "source": "crawl"}, {"paper": {"id": "2406.05881", "title": "Training Fast Robot Policies with Slow Foundation Models", "authors": ["Utsav Singh", "Pramit Bhattacharyya", "Vinay P. Namboodiri", "Amrit Singh Bedi"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2406.05881", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "VGRS uses LLMs to synthesize executable reward code from instructions, training lightweight robot policies, while a VLM diagnoses failure trajectories to iteratively densify rewards. Foundation models are used only at training time, enabling fast, low-cost deployment on real robots for navigation and manipulation.", "reason": "Generated reward code is the instrument for training a policy that controls physical robots, matching the code-as-policy world_physical boundary.", "source": "crawl"}, {"paper": {"id": "2607.21292", "title": "An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models", "authors": ["Ari Luna Rueda", "Eike Cramer", "Klaus Hellgardt", "Mehmet Mercangöz"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.21292", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "An LLM-driven workflow decomposes multi-variable process-control design into constrained code-generation steps (plant interface, controller specification, tuning), executing and validating generated artifacts with feedback-based repair. Demonstrated on a gas-preheater plant, improving closed-loop performance via Bayesian optimization.", "reason": "Generated code serves as the control policy operating a physical process, making code the instrument of action rather than the deliverable.", "source": "crawl"}, {"paper": {"id": "2607.20064", "title": "PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning", "authors": ["Alexis Fox", "Junlin Wang", "Paul Rosu", "Bhuwan Dhingra"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.20064", "github": "", "website": ""}}, "category": "world_game", "tags": [], "summary": "PRO-LONG is a context-management framework for LLM agents that maintains a complete structured interaction log and uses coding-agent techniques to search it efficiently for long-horizon exploratory tasks. It improves performance and token efficiency over specialized harnesses on the full ARC-AGI-3 game set.", "reason": "Benchmark routing to ARC-AGI-3's game environments makes this a code-as-tool agent acting in game worlds, per world_game.", "source": "crawl"}]