Paper review 2026-08-26: 25 candidates #364

[{"paper": {"id": "2512.15798", "title": "Bridging Business Intent and Data: A Benchmark for Automatic Relational Data Product Generation", "authors": ["Faisal Chowdhury", "Sola Shirai", "Sarthak Dash", "Nandana Mihindukulasooriya", "Horst Samulowitz"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2512.15798", "github": "", "website": ""}}, "category": "database", "tags": ["benchmark"], "summary": "DP-Bench is a benchmark for automatically generating relational data products (curated database assets) from high-level business requests, combining elements of text-to-SQL and ELT pipeline datasets. The paper also proposes LLM-based baseline methods for this generation task.", "reason": "Benchmark and baselines for LLMs producing database/query artifacts from natural-language intent, matching the database leaf.", "source": "crawl"}, {"paper": {"id": "2608.21157", "title": "HIERA: Workload-Aware Planning Across Implementation Spaces for GPU Kernel Optimization", "authors": ["Jinghao Wang", "Qiqi Gu", "Chenpeng Wu", "Jianguo Yao", "Haibing Guan", "Xijun Li"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21157", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "HIERA is a hierarchical search-space planning framework that uses LLMs to generate and optimize GPU kernels across PyTorch, CUDA library, and custom CUDA implementation spaces, guided by profiling feedback. Evaluated on KernelBench, it improves validity and optimization performance over training-free baselines.", "reason": "The task is producing optimized systems-level GPU kernel code, matching the systems leaf.", "source": "crawl"}, {"paper": {"id": "2608.20711", "title": "AsmEvo: Agentic Assembly-Level Optimization of AMD GPU Kernels with Functional Equivalence Verification", "authors": ["Ji Liu", "Puyuan Yang", "Rongzhang Zheng", "Fan Wang", "Jinglin Wang", "Muhammad A. Awad", "Mortis Huang", "Andy Chang", "Zekai Li", "Zeping Li", "Zihao An", "Yue Liu", "Yuchen Yang", "Jianghui Wang", "Chushi Chen", "Ziqiong Liu", "Fuwei Yang", "Dong Li", "Wen Heng Chung", "Shengcai Liu", "Emad Barsoum"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20711", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "AsmEvo is an agentic assembly-level optimizer that edits compiled AMD GPU kernel binaries directly, using differential verification against the original object to preserve functional equivalence. It achieves substantial speedups on KernelBench and production AITer/vLLM/SGLang kernels on MI300X/MI308X hardware.", "reason": "Agent produces low-level kernel/systems code (GPU assembly) as the deliverable, matching the systems leaf per KernelBench routing precedent.", "source": "crawl"}, {"paper": {"id": "2608.21584", "title": "Data-Driven Dynamic Algorithm Dispatch with Large Language Models", "authors": ["Rushil Shah", "Emmanuel Lujan", "Rabab Alomairy", "Alan Edelman"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21584", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "An LLM-driven approach using LLaMA 3 and a performance database to synthesize dynamic algorithm-dispatch heuristics for high-performance linear algebra, demonstrated on LU factorization. Aims to replicate expert-designed selection strategies for adaptive numerical software.", "reason": "The task is generating systems-level performance/dispatch code for numerical libraries, matching the systems artifact leaf.", "source": "crawl"}, {"paper": {"id": "2608.21833", "title": "GameXpert-Bench: How Far Are Coding Agents from Expert Game Development?", "authors": ["Kun Chen", "Haorong Hong", "Peizhong Gao", "Jianfeng Lin", "Tongxu Luo", "Yuxuan Xie", "Chenxu Liu", "Jieling He", "Zhongyuan Liu", "Zeno Zeng"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.21833", "github": "", "website": ""}}, "category": "game", "tags": ["benchmark"], "summary": "GameXpert-Bench evaluates coding agents across the full game development lifecycle—initial generation, bug diagnosis/repair, and multi-turn optimization—via three tracks with playable-game and behavioral tests. Results show agents excel at initial implementation but struggle with defect discovery and preserving functionality across changes.", "reason": "Benchmark for building and evolving a game's codebase serves the game artifact leaf per benchmark routing.", "source": "crawl"}, {"paper": {"id": "2606.30949", "title": "AgRefactor: Self-Evolving Agentic Workflow for HLS Compatibility and Performance", "authors": ["Yang Zou", "Zijian Ding", "Yizhou Sun", "Jason Cong"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-29", "links": {"paper": "https://arxiv.org/abs/2606.30949", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "AgRefactor is a self-evolving multi-agent LLM workflow that refactors real-world software into HLS-compatible, synthesizable hardware code, combining LLM rewrites with automated refactoring tools and pragma tuning. It outperforms state-of-the-art tools on long real-world benchmarks and yields large speedups via agentic performance optimization.", "reason": "The task's deliverable is hardware-synthesizable code (HLS), placing it in the hardware artifact domain rather than generic software.", "source": "crawl"}, {"paper": {"id": "2608.21356", "title": "AI with Authority, from Application to Silicon", "authors": ["Jason Hickey"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21356", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "A researcher directs a fleet of AI agents through a fully machine-verified pipeline—from application code through a verified compiler and executive to RTL—resulting in a RISC-V processor taped out with no human-written RTL or human-reviewed proofs. Verification chains Lean 4 kernel proofs to SAT-checked equivalence at the silicon boundary.", "reason": "The ultimate deliverable is taped-out hardware (RTL/RISC-V processor) produced by AI agents, matching the hardware leaf.", "source": "crawl"}, {"paper": {"id": "2608.22938", "title": "Execution-Anchored Hallucination Calibration Reranking for Verilog Code Generation", "authors": ["Guang Yang", "Xing Hu", "Xiang Chen", "Terry Yue Zhuo", "Xin Xia"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.22938", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "EAHC is a reranking framework for LLM-generated Verilog code that fuses execution-based testbench signals with a hallucination-calibrated reasoning discriminator to select the best candidate. It targets the low-resource HDL code generation setting, improving selection over pure execution or LLM-judge reranking.", "reason": "Serves hardware description code generation, the produced artifact is Verilog/RTL code.", "source": "crawl"}, {"paper": {"id": "2602.00704", "title": "QiMeng-ChipV-RTL: Exploiting Information Locality for IP-level Verilog Generation", "authors": ["Hanqi Lyu", "Di Huang", "Yaoyu Zhu", "Kangcheng Liu", "Bohan Dou", "Chongxiao Li", "Pengwei Jin", "Shuyao Cheng", "Rui Zhang", "Zidong Du", "Qi Guo", "Xing Hu", "Yunji Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2602.00704", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "ChipV-RTL is a multi-agent framework that decomposes long IP-level hardware specifications into shorter document and code tasks, using hierarchical partitioning, localized generation, and AST-guided debugging to generate Verilog RTL code, achieving state-of-the-art pass rates on RealBench.", "reason": "The agent produces hardware description (Verilog/RTL) code, matching the hardware artifact leaf.", "source": "crawl"}, {"paper": {"id": "2608.23602", "title": "PACT: Post-route Agentic Checkpoint Tuning for FPGA Timing Closure", "authors": ["Huan Lin", "Kunlong Li", "Lingli Wang", "Zhiang Zhang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-19", "links": {"paper": "https://arxiv.org/abs/2608.23602", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "a", "reason": "a", "source": "crawl"}, {"paper": {"id": "2608.23630", "title": "FPGAgent: An LLM-Assisted Framework for Autonomous HLS Code Generation and Verification in FPGA Environments", "authors": ["Tianyu Wang", "Wenjie Wang", "Jianguo Yao", "Haibing Guan", "Xijun Li"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.23630", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "FPGAgent is a multi-agent LLM framework that autonomously generates, verifies, and deploys HLS FPGA designs from natural-language specifications, including validation programs and board-level executability checks. Evaluated on the HLS-Eval benchmark, it substantially improves synthesizable rate, executability, and functional correctness over baselines.", "reason": "Produces hardware description/HLS code as the deliverable, matching the hardware leaf.", "source": "crawl"}, {"paper": {"id": "2606.01160", "title": "Expected Value Alignment for Generative Reward Modeling in Formal Mathematics Verification", "authors": ["Shihao Ji", "Haotao Tan", "Zihui Song", "Mingyu Li"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-31", "links": {"paper": "https://arxiv.org/abs/2606.01160", "github": "", "website": ""}}, "category": "world_research", "tags": ["model"], "summary": "Introduces EVA, a reward-modeling technique that extracts continuous scores from discrete token distributions, instantiated as the Leibniz reward model for evaluating Lean 4 formal proof steps.", "reason": "Resource paper serving RL/search-based Lean theorem proving, which is pure mathematical discovery under world_research.", "source": "crawl"}, {"paper": {"id": "2608.20434", "title": "An LLM agent for end-to-end computational materials discovery", "authors": ["Chen Yuntong", "Huang Ju", "Liu Yu", "Zhao Dan", "Sun Mingqi", "Ju Chentian", "Liu Yanbing", "Huang Lijiang", "Zhao Guobin"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-20", "links": {"paper": "https://arxiv.org/abs/2608.20434", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "MAESTRO is an LLM agent system that runs an end-to-end computational materials discovery pipeline, linking literature to crystal structures and screening metal-organic frameworks for gas separation. It autonomously coordinates multi-scale computational tools to identify promising candidate materials.", "reason": "Code/tools are instruments for scientific discovery whose product is a material finding, matching world_research's AI-scientist style discovery agents.", "source": "crawl"}, {"paper": {"id": "2608.09855", "title": "The Greatness of Science Cannot Be Planned: Agentic Auto-Research is Fuzz Testing", "authors": ["Yifeng He", "Jicheng Wang", "Yinzhe Zhao", "Chengyang Shi", "Jiachen Liu", "Hao Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-10", "links": {"paper": "https://arxiv.org/abs/2608.09855", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "This paper argues agentic auto-research should mimic fuzz testing by using dense intermediate progress signals to guide exploration rather than optimizing only a final benchmark score, demonstrating a simulated physics-law discovery agent that outperforms optimization-driven baselines.", "reason": "Proposes a research agent whose goal is scientific discovery through interaction, matching world_research's AI-scientist style discovery agents.", "source": "crawl"}, {"paper": {"id": "2608.21601", "title": "K-Bench: measuring model performance on real scientific agent requests", "authors": ["Aubrey Brueckner", "Darshil Patel", "Yuhuan He", "Timothy Kassis"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21601", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "K-Bench evaluates nine frontier LLM agents end-to-end on real, underspecified scientific requests sampled from live user traffic, scoring 1,602 sandboxed agent runs with an eight-dimension rubric via blinded judges.", "reason": "Benchmarks agents performing open-ended scientific research tasks in sandboxes, matching world_research's discovery-through-interaction scope.", "source": "crawl"}, {"paper": {"id": "2608.21976", "title": "Closed-loop AI achieves certifiable engineering design", "authors": ["Tianyi Yu", "Chengxing Tao", "Haoxuan Shen", "Huiyang Li", "Rugang Chen", "Long Teng", "Lilin Wang", "Yan Li", "Qingbin Chen", "Chaogang Xu", "Lizhong Wang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.21976", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "The AI Engineer couples LLMs to deterministic engineering solvers (CalculiX, Zwind) in a closed loop with topology and particle-swarm optimization to design certifiable offshore floating-wind structures. An automated reviewer gates candidates until a certified design outperforms a human-optimized baseline and passes classification-society approval.", "reason": "Agentic closed-loop design-and-simulation discovery where code/solvers are the instrument and the product is a certified physical engineering design, matching world_research.", "source": "crawl"}, {"paper": {"id": "2511.22884", "title": "InsightEval: An Expert-Curated Benchmark for Assessing Insight Discovery in LLM-Driven Data Agents", "authors": ["Zhenghao Zhu", "Yuanfeng Song", "Xin Chen", "Chengzhong Liu", "Yakun Cui", "Caleb Chen Cao", "Sirui Han", "Yike Guo"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-28", "links": {"paper": "https://arxiv.org/abs/2511.22884", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "InsightEval is a newly curated benchmark for evaluating LLM-driven data agents' ability to discover insights via exploratory data analysis, addressing flaws in the prior InsightBench dataset with a new metric for exploratory performance.", "reason": "Benchmarks data-analysis agents whose deliverable is insight through interaction, matching world_research.", "source": "crawl"}, {"paper": {"id": "2608.23622", "title": "LLM Agents Perform Controlled Experiments Using Simulation Models", "authors": ["Yuchen Xia", "Michael Weyrich", "Nasser Jazdi", "Johannes Stümpfle", "Johannes Sigel", "Akshay Narla", "Gavin K. Reynolds", "Anna Jawor-Baczynska", "Pol Llopart"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.23622", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "A multi-agent LLM framework that designs, executes, and interprets controlled experiments using scientific simulation models for pharmaceutical process design, producing evidence-based optimization recommendations. Evaluated in an industrial setting, it outperforms language-only reasoning in specificity and correctness.", "reason": "Agents operate simulators to drive scientific/engineering discovery, matching world_research's simulation-for-discovery examples.", "source": "crawl"}, {"paper": {"id": "2608.24086", "title": "EMRB: A Multi-Level Benchmark for Evaluating LLM Reasoning over Raw Electromagnetic Signals", "authors": ["Mingxu Zhang", "Ying Sun", "Yuhan Li", "Yang Ji", "Dazhong Shen", "Ke Zhang", "Shan Huang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24086", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "EMRB is a benchmark testing whether LLM code agents can analyze raw electromagnetic I/Q signal data by writing and executing code, covering 200 problems across five difficulty levels; the authors also propose ReconPilot, a reconnaissance-then-analysis method that improves scores.", "reason": "Agents write/execute code as an instrument to derive scientific/engineering insight from raw data, matching world_research's data-analysis/discovery scope.", "source": "crawl"}, {"paper": {"id": "2605.19743", "title": "EngiAI: Capability-Based Evaluation of Tool-Connected LLM Agents for Engineering Design", "authors": ["Gioele Molinari", "Florian Felten", "Soheyl Massoudi", "Mark Fuge"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2605.19743", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "EngiAI is a capability-based evaluation framework for tool-connected LLM agents on engineering design tasks (beam and photonics optimization), assessing workflow execution, HPC orchestration, and training-code authoring via a LangGraph implementation on EngiBench. It benchmarks multiple LLM backends across these engineering-agent capabilities.", "reason": "Agents orchestrate HPC and author training code to drive engineering experiments, matching world_research's MLE/engineering-discovery-through-interaction scope.", "source": "crawl"}, {"paper": {"id": "2608.20485", "title": "Terminal Agents: A Survey of AI Agents in Command-Line Environments", "authors": ["Yi Bin", "Xiaoyang Yuan", "Haoxi Zeng", "Wencheng Ye", "Wenqi Shao", "Chen Qian", "Wei Ye", "Yujuan Ding", "Zheng Wang", "Pengpeng Zeng", "Jingkuan Song", "Heng Tao Shen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-20", "links": {"paper": "https://arxiv.org/abs/2608.20485", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["survey"], "summary": "A survey unifying terminal-mediated AI agent research across software engineering, tool use, and computer-use, proposing a seven-dimensional terminal competence profile connecting architecture, learning, and evaluation. It highlights uneven evaluation of process quality, recovery, and governance in terminal agents.", "reason": "A survey scoped to a single world (terminal/command-line agents), so it routes to world_terminal plus the survey tag rather than the field-wide studies leaf.", "source": "crawl"}, {"paper": {"id": "2608.21101", "title": "ClawSentry: A Progressive Multi-Tier Security Monitor for Safeguarding Autonomous LLM Agents", "authors": ["Kai Wang", "Zeming Wei", "BiaoJie Zeng", "Chang Jin", "An Wang", "Xiaokun Luan", "Zhixiao Lin", "Jingjing Qu", "Xia Hu", "Xingcheng Xu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21101", "github": "", "website": ""}}, "category": "world_terminal", "tags": [], "summary": "ClawSentry is a multi-tier security gateway that supervises autonomous LLM agents (Codex, Claude Code, Kimi CLI, Gemini CLI) as they execute code, read files, and invoke tools, guarding against malicious skill packages across the agent's action loop. It substantially reduces attack success rates on skill-injection and safety benchmarks while preserving task success.", "reason": "Secures CLI/terminal-based agents whose language of action is code, so it serves the terminal/OS world rather than a specific software artifact task.", "source": "crawl"}, {"paper": {"id": "2608.21423", "title": "Agentic Security: A Systematization of Tools, Failure Modes, and Design Laws for LLM-Driven Penetration Testing", "authors": ["Israt Moyeen Noumi", "Tarannum Ahmed Nowshin", "Md. Mehedi Hasan Nipu", "Mohammad Sakib Mahmood", "Md. Jakir Hossain", "M. F. Mridha"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-16", "links": {"paper": "https://arxiv.org/abs/2608.21423", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["survey"], "summary": "A systematization of failure modes in LLM-driven agentic penetration testing across ten security tools, deriving quantitative design laws (evidence loss, verdict cascades, execution caps) and introducing the Inspectra platform as a worked instantiation. Focuses on offensive-security agents operating tools in unattended pipelines.", "reason": "Field-wide systematization of agentic pentesting tools/failure modes is a survey of one world (offensive security in terminals/CLI pipelines).", "source": "crawl"}, {"paper": {"id": "2608.22103", "title": "Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks", "authors": ["Amit Roth", "Ivan Bercovich", "Yonathan Efroni"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.22103", "github": "", "website": ""}}, "category": "world_terminal", "tags": ["benchmark"], "summary": "Introduces Hack-Verifiable Terminal Bench, adapting the hack-verifiable-environments methodology to Terminal Bench to automatically detect reward hacking in real-world terminal and coding tasks. Measures reward-hacking rates across frontier models and tests whether prompting mitigates known and unknown exploits.", "reason": "Benchmark of agents completing real-world terminal tasks, the terminal is the agent's world.", "source": "crawl"}, {"paper": {"id": "2608.22631", "title": "Learning Generalizable Behaviors for Terminal Agents", "authors": ["Yihang Yao", "Bo Pang", "Xuan Phi Nguyen", "Ding Zhao", "Shafiq Joty", "Semih Yavuz"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22631", "github": "", "website": ""}}, "category": "world_terminal", "tags": [], "summary": "Studies how RL shapes terminal-agent behavior, proposing the Agentic Compositional Generalization hypothesis and River, a training recipe filtering low-quality environments and adding process-level rewards. River improves RL efficiency across terminal-agent benchmarks like Terminal-Bench.", "reason": "The agent acts in a terminal/shell to complete jobs, making the terminal the world of action per world_terminal.", "source": "crawl"}]