Paper review 2026-07-29: 25 candidates #344

[{"paper": {"id": "2607.22759", "title": "Benchmarking LLMs for Verilog Design Flows", "authors": ["Angshuman Chakravertty", "Rahul Koshti", "Buddhi Prakash Sharma", "Vinay Chamola"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.22759", "github": "", "website": ""}}, "category": "hardware", "tags": ["benchmark"], "summary": "Presents a reproducible benchmarking platform for LLM-generated Verilog RTL code across 50 tasks, using iterative refinement, AST-based repair, and formal/simulation verification via Verilator and Icarus Verilog. Evaluates three open-source LLMs on syntax validity and functional correctness.", "reason": "Delivers hardware description (Verilog RTL) code, matching the hardware leaf's definition; benchmark artifact released.", "source": "crawl"}, {"paper": {"id": "2607.23715", "title": "Formally Verified Synthesizable Floating-Point Data Types in ARCH HDL", "authors": ["Shuqing Zhao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-26", "links": {"paper": "https://arxiv.org/abs/2607.23715", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "The paper formally verifies IEEE-754 FP32 and BF16 arithmetic operators for ARCH, a hardware description language designed to be generated by LLMs, using a shared IR rendered to SystemVerilog, SMT-LIB, and Lean proofs. It also re-implements the FMA unit for pipelining while proving bit-identical equivalence to the exact reference.", "reason": "Delivers formally verified hardware description code for an LLM-targeted HDL, matching the hardware artifact leaf.", "source": "crawl"}, {"paper": {"id": "2607.23523", "title": "CircuitWeave: Topology-Behavior Alignment for Executable Multimodal RTL Generation", "authors": ["Jiahao Feng", "Haiyan Qin", "Zhiwei Xie", "Wang Kang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-26", "links": {"paper": "https://arxiv.org/abs/2607.23523", "github": "", "website": ""}}, "category": "hardware", "tags": ["training-data"], "summary": "CircuitWeave generates RTL code from combined schematic and text specifications by extracting and fusing topology and behavior contracts before code generation. It releases a 5,000-sample executable-qualified dataset and improves pass@k on VerilogEval-Human and RTLLM.", "reason": "Task's deliverable is hardware description (RTL) code, matching the hardware leaf.", "source": "crawl"}, {"paper": {"id": "2603.13767", "title": "Retrieve, Schedule, Reflect: LLM Agents for Chip QoR Optimization", "authors": ["Yikang Ouyang", "Yang Luo", "Dongsheng Zuo", "Yuzhe Ma"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2603.13767", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "An agentic LLM framework that schedules chip design optimizations by directly interacting with EDA tools, using retrieval-augmented search-tree guidance and Pareto-driven reflection to improve timing, power, and area.", "reason": "The agent's ultimate purpose is producing an optimized chip design, matching the hardware artifact leaf.", "source": "crawl"}, {"paper": {"id": "2607.24854", "title": "VClare: Resolving Imperfect Specifications in LLM-Based Verilog Generation", "authors": ["Zhuorui Zhao", "Bing Li", "Yu Li", "Zheyu Yan", "Ulf Schlichtmann"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-25", "links": {"paper": "https://arxiv.org/abs/2607.24854", "github": "", "website": ""}}, "category": "hardware", "tags": ["benchmark"], "summary": "VClare is a framework that repairs imperfect natural-language hardware specifications via LLM-driven inconsistency mining and simulation-based clustering, improving downstream Verilog code generation quality, with two new injected-defect benchmarks released.", "reason": "Serves hardware description code generation (Verilog), matching the hardware artifact leaf.", "source": "crawl"}, {"paper": {"id": "2509.06278", "title": "TableMind: An Autonomous Programmatic Agent for Tool-Augmented Table Reasoning", "authors": ["Chuang Jiang", "Mingyue Cheng", "Xiaoyu Tao", "Qingyang Mao", "Jie Ouyang", "Qi Liu"], "venue": "arXiv 2025/09", "category": "", "published": "2025-09-08", "links": {"paper": "https://arxiv.org/abs/2509.06278", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "TableMind is a tuning-based autonomous agent that internalizes multi-turn planning, code-tool use, and reflection via SFT and RL to answer complex table reasoning questions with precise numerical operations.", "reason": "Code is generated as a tool to extract insight from tabular data, matching data-analysis agents under world_research.", "source": "crawl"}, {"paper": {"id": "2607.22642", "title": "CRAFT: Learn the Schema, Execute the Plan", "authors": ["Aakash Kolekar", "Sahika Genc", "Shahriar Shariat", "Bunyamin Sisman", "Tibor Mezi", "Barbara Poblete", "Shree Vandana Kachroo", "Calvin Chi", "Parth Parmar", "Ari Singer", "Prayaas Jain", "Cindy Barker", "Benoit Dumoulin"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-24", "links": {"paper": "https://arxiv.org/abs/2607.22642", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "Presents CRAFT, a two-stage post-training recipe (schema-stripped plan SFT plus execution-shaped RL) for enterprise coding agents that translate natural-language analytical requests into executable code over proprietary schemas and APIs. Evaluated on advertising analytics workflows, improving agent score, consistency, and multi-turn coherence.", "reason": "Code is the tool for data analysis whose deliverable is analytical insight, matching world_research's data-science agent definition.", "source": "crawl"}, {"paper": {"id": "2607.22688", "title": "Co-Harness: Co-Evolving Harnesses and Model Weights for LLM Agents", "authors": ["Zhengyu Chen", "Teng Xiao", "Huaisheng Zhu", "Yige Yuan", "Luan Zhang", "Jingang Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-17", "links": {"paper": "https://arxiv.org/abs/2607.22688", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "Co-Harness jointly optimizes an LLM agent's runtime harness (prompts, tools, memory) and model weights during post-training for automated AI research, using an LLM-based critic to identify and fix harness-level failures from trajectories.", "reason": "Serves automated AI/ML research agents (MLE-engineering style), where code is the instrument for running experiments rather than the deliverable, matching world_research.", "source": "crawl"}, {"paper": {"id": "2607.23286", "title": "TopoFE: topology-aware LLM-guided Automated Feature Engineering", "authors": ["Sha Li", "Naren Ramakrishnan"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-25", "links": {"paper": "https://arxiv.org/abs/2607.23286", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "TopoFE is a topology-aware multi-island evolutionary framework that uses an LLM to synthesize and search feature-engineering programs for tabular learning, improving on prior AutoFE methods across 29 datasets. Generated feature programs are executed and evaluated within the search loop to boost downstream predictive performance.", "reason": "The generated code programs are instruments toward better predictive models, not the deliverable, fitting ML-engineering/AutoML under world_research.", "source": "crawl"}, {"paper": {"id": "2607.24051", "title": "HELIOS: An LLM-Driven Autonomous Indirect Trajectory Optimization Agent", "authors": ["An-yi Huang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24051", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "HELIOS is an LLM-driven agent that autonomously performs symbolic derivation, SymPy verification, C++ shooting-code generation, and numerical solving for low-thrust spacecraft trajectory optimization problems. It is tested across progressively complex deep-space trajectory scenarios with multiple LLM backends.", "reason": "Code is the instrument for solving an aerospace engineering/research problem, not the deliverable itself, fitting world_research.", "source": "crawl"}, {"paper": {"id": "2604.19341", "title": "Structured Scaling of AI Discovery Across Diverse Scientific Domains", "authors": ["Haotian Ye", "Haowei Lin", "Jingyi Tang", "Yizhen Luo", "Rahul Thapa", "Caiyin Yang", "Chang Su", "Rui Yang", "Ruihua Liu", "Rundao Li", "Zeyu Li", "Pengwei Sun", "Chong Gao", "Dachao Ding", "Guangrong He", "Miaolei Zhang", "Lina Sun", "Wenyang Wang", "Yuchen Zhong", "Zhuohao Shen", "Puheng Li", "Pan Lu", "Bianxiao Cui", "Di He", "Jianzhu Ma", "Junfeng Li", "Hexi Baoyin", "Yejin Choi", "Stefano Ermon", "Xiaowen Chu", "Tongyang Li", "Yuzhi Xu", "James Zou"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2604.19341", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "SimpleTES structures test-time scaling of AI-driven scientific discovery loops, organizing evaluator queries across trajectories and refinement to achieve new state-of-the-art solutions across 28 open problems spanning quantum physics, astronomy, biology, and mathematics.", "reason": "An AI-scientist style discovery agent where code and experiments are instruments toward scientific findings, matching world_research.", "source": "crawl"}, {"paper": {"id": "2607.25090", "title": "Matryoshka Agent: Unfolding Sub-Agents for Long-Horizon Machine Learning Engineering", "authors": ["Rushi Qiang", "Changhao Li", "Haotian Sun", "Yuchen Zhuang", "Chao Zhang", "Bo Dai"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.25090", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "Matryoshka Agent is a hierarchical agent framework that decouples a long-horizon Orchestrator from execution Sub-Agents to handle long-horizon machine learning engineering (MLE) tasks like iterative solution debugging and refinement. It shows strong gains on MLE benchmarks, enabling smaller models to match larger ones.", "reason": "Serves machine-learning engineering through iterative experimentation, matching agency/world_research's MLE-agent scope.", "source": "crawl"}, {"paper": {"id": "2607.25145", "title": "Agentic AI for Scientific Reasoning in Autonomous Quantum Sensing Experiments", "authors": ["Takuya Isogawa", "Ryotaro Okabe", "Nutdech Phadetsuwannukun", "Mingda Li", "Paola Cappellaro"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.25145", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "This paper implements an agentic LLM workflow that autonomously conducts NV-center quantum sensing experiments, combining project memory, analysis tools, and deterministic hardware control, and introduces offline benchmarks isolating the agent's scientific reasoning from execution. Reasoning quality varies by task and model, with expected-signal calculations reducing false-positive resonance detections.", "reason": "An LLM agent operating scientific instruments toward discovery is a self-driving-lab scientific-discovery task under agency/world_research.", "source": "crawl"}, {"paper": {"id": "2605.11665", "title": "Nautilus: From One Prompt to Plug-and-Play Robot Learning", "authors": ["Yufeng Jin", "Jianfei Guo", "Xiaogang Jia", "Yu Deng", "Zechu Li", "Han Liu", "Weiran Liao", "Vignesh Prasad", "Mathias Franzius", "Gerhard Neumann", "Georgia Chalvatzaki"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2605.11665", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "NAUTILUS is an agentic harness that turns a single prompt into reproduction, evaluation, fine-tuning, and deployment workflows for robot learning, auto-generating adapters and containers to bridge policies, benchmarks, and robots.", "reason": "Code (adapters/glue) is the tool enabling robot-learning experiment management, matching MLE-engineering style research interaction (world_research).", "source": "crawl"}, {"paper": {"id": "2606.27243", "title": "NOVA: A Verification-Aware Agent Harness for Architecture Evolution in Industrial Recommender Systems", "authors": ["Shaohua Liu", "Liang Fang", "Yilong Sun", "Shudong Huang", "Qingsong Luo", "Shaoxin Liu", "Xiaoyang Chen", "Dongqiang Liu", "Chuangang Ma", "Zhenzhen Chai", "Henghuan Wang", "Shijie Quan", "Changyuan Cui", "Zhangbin Zhu", "Peng Chen", "Wei Xu", "Lei Xiao", "Haijie Gu", "Jie Jiang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2606.27243", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "NOVA is a verification-aware agent harness that searches over concrete architecture modifications for production recommender systems, rejecting semantically invalid candidates and using trajectory memory to guide search, validated via production A/B tests.", "reason": "An ML-engineering agent that generates and evaluates candidate model implementations, matching MLE-style training/tuning/experiment-management interaction (world_research).", "source": "crawl"}, {"paper": {"id": "2607.25834", "title": "Lowering the implementation barrier of neutral-atom quantum computing with agentic workflows", "authors": ["Constantin Dalyac", "Alexandre Dauphin", "Loïc Henriet", "Christophe Jurczak"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25834", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "An agentic workflow automates the pipeline from theoretical quantum protocols to hardware experiments on neutral-atom quantum processors, handling compilation, simulation, and cloud execution with human validation in the loop. Case studies show the agent running overnight QPU campaigns, with domain experts catching two scientifically invalid decisions.", "reason": "Agent operates scientific/quantum computing software and hardware in pursuit of experimental discovery, matching self-driving-lab style world_research.", "source": "crawl"}, {"paper": {"id": "2607.25865", "title": "OmniQEC: discovering practical quantum error-correcting codes by an AI scientist", "authors": ["Ge Yan", "Shanchuan Li", "Pengyue Ma", "Qixin Zhang", "Pingchuan Ma", "Jianping Wang", "Min-Hsiu Hsieh", "Yuxuan Du"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25865", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "OmniQEC is an LLM-orchestrated AI-scientist system that iteratively generates and evaluates quantum error-correcting codes, coordinating code generation, syndrome-extraction synthesis, and circuit-level evaluation to discover hardware-friendly QEC codes outperforming known constructions.", "reason": "An LLM orchestrator writes and evaluates code to drive scientific discovery, matching the AI-scientist pattern under world_research.", "source": "crawl"}, {"paper": {"id": "2607.25886", "title": "RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement", "authors": ["Fanqing Meng", "Lingxiao Du", "Qiguang Chen", "Ziqi Zhao", "Haocheng Lu", "Mengkang Hu", "Michael Qizhe Shieh"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25886", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "RSIBench-Data is a controlled benchmark where LLM agents act as data-centric researchers, iteratively revising post-training data strategies for a fixed model using sandboxed training/evaluation infrastructure, evaluated across software engineering, terminal, QA, and math tasks.", "reason": "Agents conduct ML experimentation/training-strategy research through interaction, matching the MLE-bench-style world_research definition.", "source": "crawl"}, {"paper": {"id": "2607.23147", "title": "False Prophets: On the Security of World Models in Agentic Systems", "authors": ["Erik Imgrund", "Anna Wimbauer", "Klim Kireev", "Konrad Rieck"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-25", "links": {"paper": "https://arxiv.org/abs/2607.23147", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "Identifies security vulnerabilities in world models used to give terminal-based LLM agents predictive capabilities, showing attackers can induce mispredictions leading to malicious code execution, denial of service, or data exfiltration with up to 95% success. Introduces a security benchmark dataset for text-based world models and mitigation recommendations.", "reason": "Security study of terminal-based code-executing agents' world models, fitting world_terminal since the agents act via shell/code in that world.", "source": "crawl"}, {"paper": {"id": "2607.25379", "title": "Cyber-Capable AI Agents: Vulnerabilities, Evaluation Containment, and Defensive Response", "authors": ["Abu Bakar Siddik"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25379", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["survey"], "summary": "A review synthesizing vulnerability classes and containment strategies for cyber-capable AI agents that perform multi-step offensive-security tasks via tool use and execution environments, using a 2026 incident as a case study. It identifies priorities for evaluating capability alongside evaluation-environment security.", "reason": "A field survey focused on one world (offensive-security/terminal agents), so it stays in that leaf with the survey tag rather than in studies.", "source": "crawl"}, {"paper": {"id": "2406.03367", "title": "CLMASP: Coupling Large Language Models with Answer Set Programming for Robotic Task Planning", "authors": ["Xinrui Lin", "Yangfan Wu", "Huanyu Yang", "Yu Zhang", "Yanyong Zhang", "Jianmin Ji"], "venue": "arXiv 2024/06", "category": "", "published": "2024-06-05", "links": {"paper": "https://arxiv.org/abs/2406.03367", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "CLMASP couples an LLM's skeleton plan with Answer Set Programming to generate executable, grounded robot action plans, improving executable-plan rate from under 2% to over 90% on VirtualHome.", "reason": "Generated plans are executed as robot actions, so code/logic-program is the policy language for physical-world control.", "source": "crawl"}, {"paper": {"id": "2607.22832", "title": "MEMENTO: Memory-Guided Memetic Code-as-Policy Evolution", "authors": ["Alkis Sygkounas", "Victor Aregbede", "Amy Loutfi", "Andreas Persson"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22832", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "MEMENTO is a memory-guided memetic framework that evolves executable code-as-policy programs for long-horizon embodied manipulation tasks, using an evolved fitness evaluator to guide program search. It outperforms prior code-as-policy baselines and demonstrates sim-to-real transfer on a physical Franka robot.", "reason": "Code is the executable policy language controlling a physical robot, matching the world_physical code-as-policy example.", "source": "crawl"}, {"paper": {"id": "2607.23784", "title": "A Few Words Go a Long Way: Language Guided Robot Policy Synthesis", "authors": ["Daphne Chen", "Archit Ritesh Jain", "Eric Goossen", "Emma Romig", "Michael Murray", "Nick Walker", "Maya Cakmak"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-26", "links": {"paper": "https://arxiv.org/abs/2607.23784", "github": "", "website": ""}}, "category": "world_physical", "tags": [], "summary": "ARCHITECT uses LLM coding agents to synthesize modular robot programs (perception/control tool calls) as an interactive program-synthesis approach to robot policy acquisition, with human corrections grounded in execution traces. Outperforms VLA baselines on a Franka Panda robot for long-horizon manipulation tasks.", "reason": "Generated programs are executed as robot control policies, matching the world_physical code-as-policy definition.", "source": "crawl"}, {"paper": {"id": "2606.01862", "title": "RadioMaster: Multi-Agent System for Autonomous Radio Signal Generation", "authors": ["Jiazhen Lei", "Yuxin Sha", "Tianze Cao", "Sihan Wang", "Bingbing Wang", "Zeming Yang", "Fengyuan Zhu", "Xiaohua Tian"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2606.01862", "github": "", "website": ""}}, "category": "world_physical", "tags": ["benchmark"], "summary": "RadioMaster is a multi-agent system that turns user intent into verified over-the-air radio signals by chaining protocol planning, baseband synthesis, and hardware configuration, with closed-loop physical-layer verification. The paper also introduces RadioBench, the first benchmark for autonomous radio signal generation.", "reason": "Code/tool actions configure and drive physical radio hardware, making code the instrument for physical-world action rather than a software deliverable.", "source": "crawl"}, {"paper": {"id": "2607.22798", "title": "StateAct: Program State, before Pixels, for Long-Horizon Computer-Use Agents", "authors": ["Yan Yang", "Xiangru Jian", "Ziyang Luo", "Zirui Zhao", "Yutong Dai", "Ziji Shi", "Hanshu Yan", "Jun Hao Liew", "Silvio Savarese", "Junnan Li"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22798", "github": "", "website": ""}}, "category": "world_apps", "tags": [], "summary": "StateAct is a code-first multi-agent harness for computer-use tasks that inspects and modifies program state directly via code rather than relying on screenshots, using a GUI subagent only for rare click-based subgoals. Evaluated on OSWorld 2.0, it improves accuracy and cost efficiency over screenshot-driven baselines.", "reason": "A computer-use agent acting through generated code across OS applications fits world_apps (computer-use agents acting via code).", "source": "crawl"}]