Paper review 2026-08-06: 15 candidates #357

[{"paper": {"id": "2505.02012", "title": "Automated Database Testing via LLM-Synthesized SQL Features", "authors": ["Suyang Zhong", "Manuel Rigger"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2505.02012", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "ShQveL augments SQL test-case generators with LLM-synthesized SQL fragments via sketches, broadening covered SQL features and finding 55 new bugs across 5 DBMSs.", "reason": "An automated fuzzing/test-generation approach for general correctness of DBMS software fits software_testing.", "source": "crawl"}, {"paper": {"id": "2509.13656", "title": "Automated Assertion Generation and Regression Testing for Machine Learning Notebooks", "authors": ["Yingao", "Yao", "Vedant Nimje", "Varun Viswanath", "Saikat Dutta"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2509.13656", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "NBTestGen automatically generates cell-scoped regression assertions for ML Jupyter notebooks, integrating with pytest/CI to catch silent regressions in data and model pipelines.", "reason": "Automated test/assertion generation for existing code is a software testing lifecycle activity.", "source": "crawl"}, {"paper": {"id": "2608.04278", "title": "EA-Graph: Artifact-Anchored Verification Memory for Coding Agents under Upstream Drift", "authors": ["Hwai-Jung Hsu", "Cheng-Jan Chi", "Hanna Everett"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.04278", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "EA-Graph is an artifact-anchored verification memory for coding agents that tracks whether prior verification claims about a repository remain valid after upstream code drift, rather than trusting stale prose notes. Evaluated on generated repositories with known ground truth, it improves smaller models' ability to correctly judge claims as unaffected, affected, or unprovable.", "reason": "Serves correctness-verification bookkeeping for coding agents across sessions, closest to software_testing's verification/assurance activity.", "source": "crawl"}, {"paper": {"id": "2608.04459", "title": "AdaptAgent: A Multi-agent, Domain-Guided Reasoning Framework for Code Adaptation", "authors": ["Xiaokai Rong", "Hridya Dhulipala", "Aashish Yadavally", "Tien N. Nguyen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.04459", "github": "", "website": ""}}, "category": "software_development", "tags": [], "summary": "AdaptAgent is a multi-agent framework that adapts external code snippets into a target repository by decomposing the task among intent-summarizing, policy, planning, context-mining, and patch-generating agents with compiler-based verification. It outperforms baselines at producing correct, developer-like integration patches.", "reason": "Integrates described functionality/snippets into an existing codebase, matching software_development's feature-integration definition.", "source": "crawl"}, {"paper": {"id": "2601.05451", "title": "RingSQL: Schema-Independent Synthetic Data Generation for Text-to-SQL Reinforcement Learning", "authors": ["Marko Sterbentz", "Kevin Cushing", "Cameron Barrie", "Kristian J. Hammond"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2601.05451", "github": "", "website": ""}}, "category": "database", "tags": ["training-data"], "summary": "RingSQL is a hybrid synthetic data generation framework combining schema-independent templates with LLM paraphrasing to produce reliable question-SQL pairs for text-to-SQL reinforcement learning.", "reason": "Resource paper whose downstream task is text-to-SQL query generation, routed to database per the served-task rule.", "source": "crawl"}, {"paper": {"id": "2602.11745", "title": "Text2GraphQuery-Bench: A Text to Graph Query Benchmark", "authors": ["Songlin Lyu", "Lujie Ban", "Zihang Wu", "Tianqi Luo", "Jirong Liu", "Ayoub Moussaid", "Oskar van Rest", "Heng Lin", "Chenhao Ma", "Nan Tang", "Shipeng Qi", "Yongchao Liu", "Zhan Qiu", "Juelu Zhang", "Jiajun Zheng"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2602.11745", "github": "", "website": ""}}, "category": "database", "tags": ["benchmark"], "summary": "Text2GraphQuery-Bench is a large benchmark of 267k (question, graph query) pairs spanning Cypher, GQL, and SQL/PGQ across 34 databases, evaluating LLMs' text-to-graph-query translation. It reveals a language-familiarity gap and bottlenecks in schema linking and aggregation logic.", "reason": "Analogous to Spider-style text-to-SQL, the deliverable is a database query, matching the database leaf.", "source": "crawl"}, {"paper": {"id": "2608.04450", "title": "CommBench: Can LLMs Write Correct and Efficient GPU Communication Code?", "authors": ["Shuang Ma", "Yuyi Li", "Yihan Zhang", "Hezhi Xie", "Danyang Chen", "Shuyang Ji", "Ziming Mao", "Cheng Ji", "Ansha Prashanth", "Wenting Yang", "Yiran Wang", "Chihan Cui", "Pei Yu Lin", "Ion Stoica", "Yang Zhou"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.04450", "github": "", "website": ""}}, "category": "systems", "tags": ["benchmark"], "summary": "CommBench is a benchmark of 100+ expert-curated tasks testing whether LLMs can generate correct and efficient GPU communication code (collectives, RDMA, NVLink), with an automated compile-execute-validate evaluation framework. Even top models like GPT-5.5 solve only 30.7% of tasks.", "reason": "Benchmark for generating systems-level (GPU communication) code, routed to systems per benchmark-routing rule.", "source": "crawl"}, {"paper": {"id": "2608.05033", "title": "SparseDitto: Customizing GPU Kernels for Different Sparsity Patterns with LLM-Based Agentic System", "authors": ["Shiyang Li", "Guangyan Sun", "Jinwei Tang", "Yanzhi Wang", "Mingyi Hong", "Caiwen Ding"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.05033", "github": "", "website": ""}}, "category": "systems", "tags": [], "summary": "SparseDitto is an LLM-based agentic system that uses coding and verification agents to construct custom GPU kernels for sparse matrix operations (SpMV/SpMM/SpGEMM) tailored to each matrix and target GPU, achieving large speedups over cuSPARSE.", "reason": "Coding agents generate and refine GPU kernel code, a systems-level artifact, placing it in artifact/systems.", "source": "crawl"}, {"paper": {"id": "2608.04032", "title": "EDATracer: An Agentic Framework for Large-Scale EDA Artifact Analysis", "authors": ["Phat Tieu", "Sayanti Jana", "Matthew DeLorenzo", "Jiawen Wu", "Narendran Srinivasan", "Srinivas Shakkottai", "Jiang Hu", "Jeyavijayan Rajendran"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2608.04032", "github": "", "website": ""}}, "category": "hardware", "tags": ["benchmark"], "summary": "EDATracer is an agentic framework that organizes chip-design EDA artifacts (source files, logs, netlists, reports) into a knowledge graph and vector index for evidence-grounded LLM-agent analysis. It introduces an 18.9GB dataset and 90-question benchmark, outperforming Cursor and Claude Code on debugging/design-flow understanding tasks.", "reason": "Agentic analysis of chip-design artifacts serves the hardware/chip-design domain, closest to the hardware leaf despite no HDL generation.", "source": "crawl"}, {"paper": {"id": "2608.04907", "title": "LLM-Assisted Detection and Repair of Hardware Security Vulnerabilities in Verilog Designs", "authors": ["Ethen Santana", "Gabriel Gyaase", "Hao Zheng"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.04907", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "Proposes an LLM-based methodology to detect Common Weakness Enumeration (CWE) security vulnerabilities directly in Verilog hardware designs and assist with their repair. Evaluated iteratively on single-module Verilog designs to demonstrate automated vulnerability detection.", "reason": "The task targets hardware description (Verilog) code, so it belongs to the hardware leaf despite its security framing.", "source": "crawl"}, {"paper": {"id": "2411.04440", "title": "AutoProteinEngine: A Large Language Model Driven Agent Framework for Multimodal AutoML in Protein Engineering", "authors": ["Yungeng Liu", "Zan Chen", "Yu Guang Wang", "Yiqing Shen"], "venue": "arXiv 2024/11", "category": "", "published": "2024-11-07", "links": {"paper": "https://arxiv.org/abs/2411.04440", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "AutoProteinEngine is an LLM-driven agent framework that lets biologists interact in natural language to automate model selection, hyperparameter optimization, and data retrieval for multimodal protein-engineering AutoML, evaluated on real protein engineering tasks.", "reason": "An MLE-style agent that trains/tunes models through interaction, with the model as product of experimentation, fits world_research.", "source": "crawl"}, {"paper": {"id": "2608.04071", "title": "Monte Carlo Tree Search for Table-to-Multimodal Report Generation", "authors": ["Teng Lin", "Zhiyang Zhang", "Yuyu Luo", "Nan Tang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.04071", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "MCTS-Report uses Monte Carlo Tree Search with LLM-driven actions (chapter planning, chart generation, SQL fact-checking) to construct multimodal data reports, evaluated on the new MMRBench benchmark.", "reason": "Code (SQL, chart generation) is the instrument for producing an insight-bearing report from data, matching world_research's data-exploration/insight purpose.", "source": "crawl"}, {"paper": {"id": "2608.04156", "title": "BrainBench: Benchmarking Large Language Models for Comprehensive EEG Understanding", "authors": ["Yangxuan Zhou", "Sha Zhao", "Yuning Chen", "Chen Wu", "Jiquan Wang", "Shijian Li", "Gang Pan"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-04", "links": {"paper": "https://arxiv.org/abs/2608.04156", "github": "", "website": ""}}, "category": "world_research", "tags": ["benchmark"], "summary": "BrainBench benchmarks LLMs on comprehensive, instruction-conditioned EEG analysis requiring autonomous code execution (CodeAct) and agentic analysis (BrainAgent) to produce scientifically grounded reports.", "reason": "Agents execute code as the instrument to analyze scientific signal data and produce findings, matching world_research's discovery/analysis purpose.", "source": "crawl"}, {"paper": {"id": "2608.04384", "title": "Improving Auto-Design of Neural PDE Solvers with a Domain-Specific Language", "authors": ["Shengxin Kong", "Liwen Xu", "Jingwen Fu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.04384", "github": "", "website": ""}}, "category": "world_research", "tags": [], "summary": "ADSL-PDE introduces a domain-specific language representing neural PDE solver design decisions, compiled deterministically to executable code, letting an evolutionary agent search over solver concepts rather than raw programs. This restructured search space improves auto-design efficiency and stability across PDE benchmarks.", "reason": "An agent iteratively designs and evaluates solver code as the instrument to solve PDE-modeling tasks, matching world_research's MLE/experiment-engineering pattern.", "source": "crawl"}, {"paper": {"id": "2608.04317", "title": "Trident : How to Break Deep Reinforcement Learning Cyber Defenses (Agentic)", "authors": ["Ryozo Masukawa", "Ian Bryant", "Armita Kazeminajafabadi", "Sanggeon Yun", "Hyunwoo Oh", "SungHeon Jeong", "Nathaniel D. Bastian", "Mahdi Imani", "Mohsen Imani"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-05", "links": {"paper": "https://arxiv.org/abs/2608.04317", "github": "", "website": ""}}, "category": "world_terminal", "tags": [], "summary": "Trident is an agentic LLM red-teaming framework where a trainable Planner devises attack strategies from compressed logs and a frozen Coder translates them into executable Python policies deployed against live deep-RL cyber defenders in sandboxed environments. It substantially degrades defender performance and uncovers emergent evasive behaviors undetected by static red-team baselines.", "reason": "Code-as-policy offensive-security agent operating against live systems in simulated shells, matching the world_terminal CTF/pentesting boundary example.", "source": "crawl"}]