Paper review 2026-08-26: 9 candidates #365

[{"paper": {"id": "2608.23084", "title": "ARGUS: MCP-Grounded Root Cause Analysis for Kubernetes Incidents", "authors": ["Ergi Senja", "Seyed Mohammad Reza Razavi Zadegan", "Philipp Leitner"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23084", "github": "", "website": ""}}, "category": "world_terminal", "tags": [], "summary": "ARGUS is an MCP-grounded LLM assistant that connects to live Kubernetes observability tools (Prometheus, Loki, NATS) to perform automated root-cause analysis and deliver diagnostic summaries for on-call engineers.", "reason": "An incident-mitigation/SRE agent operating a live system falls under world_terminal per the operations boundary.", "source": "crawl"}, {"paper": {"id": "2608.23179", "title": "NetConfArena: An Executable Benchmark for LLM Agents in Closed-Loop Network Configuration", "authors": ["Chang Liu", "Xiaohui Xie", "Xinyi Chen", "Yong Cui"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23179", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "NetConfArena is an executable benchmark placing LLM agents in emulated multi-device networks to perform closed-loop network configuration, evaluated via hidden executable tests. It reveals agent failures beyond command errors, including specification adherence and planning gaps.", "reason": "Agents act via commands to operate and configure networked systems, matching world_terminal's system administration/DevOps scope.", "source": "crawl"}, {"paper": {"id": "2608.21031", "title": "PhysCaP: Grounding Code-as-Policy Agent with Physics-Informed Exploration", "authors": ["Chen-Yu Lin", "Jing-Wen Chen", "Hsueh-En Chang", "Hung-An Chen", "Sheng-Hsun Chang", "Chi-Pin Huang", "Fu-En Yang", "Min-Hung Chen", "Yi-Ting Chen", "Yu-Chiang Frank Wang", "Shao-Hua Sun"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21031", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "PhysCaP augments code-as-policy robotic manipulation agents with a physics-informed exploration layer that infers hidden object properties like mass and stiffness through active interaction. A dual-agent Planner/Prioritizer design balances exploration cost against information gain on real-world and simulated manipulation tasks.", "reason": "Code-as-policy agent controlling a physical robot; code is the action language for physical manipulation.", "source": "crawl"}, {"paper": {"id": "2608.21417", "title": "Retrieval-grounded robot program generation and simulation-based correction via Model Context Protocol", "authors": ["Zhichao Zhou", "Siyuan Chen", "Omkar Salunkhe", "Ebru Turanoglu Bekar", "Johan Stahre", "Anders Skoogh"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-13", "links": {"paper": "https://arxiv.org/abs/2608.21417", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "A retrieval-augmented LLM workflow generates, validates, and iteratively corrects ABB RAPID robot programs from natural-language task descriptions, using a Model Context Protocol server linked to RobotStudio simulation for diagnostic feedback and correction.", "reason": "world_physical: code-as-policy robot control via MCP-simulator feedback loop.", "source": "crawl"}, {"paper": {"id": "2608.22149", "title": "Meta-Ctrl: Guaranteed Plan Generation by Decoupling Syntactic and Semantic Constraints", "authors": ["Gwen Yidou-Weng", "Edward Sun", "Tianyi Ma", "Metin Alp Dogan", "Benjie Wang", "Allen Peng", "Guy Van den Broeck", "Yuchen Cui"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22149", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "Meta-Ctrl is a constrained-decoding framework using compact 'meta-tokens' to guarantee syntactic and semantic validity of LLM-generated robot action plans while preserving plan quality, demonstrated on embodied benchmarks and a real tabletop robot. It closes the gap between soft grounding methods and rigid symbolic planners.", "reason": "Generated action plans are executed as robot control policies, matching code-as-policy physical-world action.", "source": "crawl"}, {"paper": {"id": "2608.22055", "title": "GenCoord: Skill-Path Commitments under Private Information", "authors": ["Peng He", "Junning Zhu", "Haohan Yuan", "Jianpeng Liang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.22055", "github": "", "website": ""}}, "category": "world_game", "tags": [], "summary": "GenCoord studies multi-agent coordination under private information, where local models emit plans compiled into executable Mineflayer skills and verified via handoff and terminal state in a Minecraft-like world.", "reason": "Agent plans are compiled to Mineflayer (Minecraft) skill code that is executed to act in a game world, matching Voyager-style code-as-policy play.", "source": "crawl"}, {"paper": {"id": "2608.20381", "title": "EditPPT: Faithful Long-Deck Slide Editing via Structured Tool-Using Multi-Agent with Dual-Modal Validators", "authors": ["Jiheon Kim", "Kyudan Jung", "Jaegul Choo"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-29", "links": {"paper": "https://arxiv.org/abs/2608.20381", "github": "", "website": ""}}, "category": "world_apps", "tags": ["benchmark"], "summary": "EditPPT is a multi-agent system that edits long PowerPoint decks by executing localized shape-level operations through the native PowerPoint COM interface, validated with dual-modal (text and visual) checks, alongside the new DeckEdit-Bench benchmark. Code (COM calls) is the language of action for manipulating the existing application artifact, not the deliverable.", "reason": "Agent operates an office application via generated control code to edit existing files, matching world_apps' computer-use-via-code definition.", "source": "crawl"}, {"paper": {"id": "2608.22533", "title": "CONTRAMEM: Learning Self-Evolving Procedural Memory from Contrasting Multi-Model Trajectories", "authors": ["Zheyuan Deng", "Binghang Lu", "Hanqi Feng", "Shirley Huang", "Dianzhuo Wang", "Yuanda Xu", "Zhiwei Zhang", "Yige Sun", "Changhong Mou", "Runyu Zhang", "Yuexing Hao", "Barnabas Poczos", "Xiaomin Li"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22533", "github": "", "website": ""}}, "category": "world_apps", "tags": [], "summary": "CONTRAMEM is a training-free framework that builds self-evolving procedural memory (Function and Skill Cards) for autonomous computer-use agents from contrasting multi-model trajectories. It substantially improves success on GAIA2/ARE and AppWorld computer-use benchmarks across multiple LLM backbones.", "reason": "Serves computer-use agents that act on applications (e.g. AppWorld's API-call code execution), placing it in world_apps.", "source": "crawl"}, {"paper": {"id": "2608.23282", "title": "From Natural Language Policies to Executable Obligations: A Verification Harness for Dependable In-Car LLM Agents", "authors": ["Radouane Bouchekir", "Damir Safin", "Tomas Bueno Momcilovic"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23282", "github": "", "website": ""}}, "category": "world_apps", "tags": [], "summary": "Presents AgentGuardUtil, a verify-and-revise harness for in-car LLM agents that compiles natural-language operating policies into executable obligation rules and deterministic gates to enforce compliant tool calls. The agent's actions are tool/API calls executed against a simulated in-car system as part of CAR-bench.", "reason": "An LLM agent acts through generated/executed tool calls to operate an in-vehicle software system, matching world_apps' code-as-action-in-an-application scope.", "source": "crawl"}]