Paper review 2026-08-26: 25 candidates #363

[{"paper": {"id": "2608.21929", "title": "SkillBloat: Token Amplification Attacks via Skill Injection in LLM Coding Agents", "authors": ["Yuanjin Zheng", "Jingbang Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.21929", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "SkillBloat studies token-amplification attacks against LLM coding agents via malicious skill injection, using a two-phase attack-screening and refinement framework to cause resource-abuse token blowups in real skill benchmarks.", "reason": "A security study attacking the skill/instruction ecosystem of code-writing agents, closest fit is software_security's security-assurance activity on code-agent tooling.", "source": "crawl"}, {"paper": {"id": "2608.22345", "title": "Mitigating Error Propagation in Chain-of-Thought: A Tree-of-Thought Framework for Smart Contract Repair", "authors": ["Jingping Zhu", "Hongping Wang", "Xiaoqi Li"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22345", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "A Tree-of-Thoughts framework repairs smart contract vulnerabilities by parsing audit reports, localizing bugs with Slither, and exploring multiple repair paths before verifying patches via compilation. Evaluated on 50 real Code4Rena vulnerabilities, it outperforms ContractTinker in repair success and patch quality.", "reason": "Automated vulnerability repair of smart contract code is a security-assurance task on code, matching software_security.", "source": "crawl"}, {"paper": {"id": "2608.22930", "title": "Concepts for Securing Agentic AI Coding and the Terok Environment", "authors": ["Jiří Vyskočil", "Franz Pöschel", "Andreas Knüpfer"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.22930", "github": "", "website": ""}}, "category": "software_security", "tags": ["position"], "summary": "Assesses IT security risks introduced by agentic AI coding tools and proposes mitigation concepts, implemented in the 'Terok' environment, aiming to enable safer adoption of agentic coding assistants.", "reason": "A position/concept paper on securing the code-production process of agentic coding tools, the closest fit being the security-assurance activity on code.", "source": "crawl"}, {"paper": {"id": "2608.23897", "title": "Names Can Hurt: Spotting Slopsquatting Risks Caused by Package Name Hallucinations in Local Coding LLMs", "authors": ["Akash Raj", "Sargam Sahu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23897", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "Proposes a two-layer detector (PyPI existence check plus a classifier) and a LangGraph-based retry/fallback pipeline to catch and avoid hallucinated package names ('slopsquatting') in LLM-generated code, reducing supply-chain risk. Evaluated on 300 prompts with escalating retries and cross-model fallback.", "reason": "An agentic pipeline that generates and safeguards code against a supply-chain vulnerability class fits software_security.", "source": "crawl"}, {"paper": {"id": "2608.24498", "title": "SeriCrypt: An LLM-Driven Context-Aware Serialization Framework for Cryptographic Protocols", "authors": ["Maosong Chen", "Xi Chen", "Mengcheng Ju", "Dongliang Zhao", "Chunxiang Gu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24498", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "SeriCrypt is an LLM-driven framework that extracts field constraints and cryptographic rules from protocol specifications into a structured DSL, then automates message serialization for protocol conformance and security testing. It uncovers specification violations and improves fuzzing coverage on TLS, IKE, SSH, and TLCP implementations.", "reason": "LLM-driven tool that generates protocol messages/test inputs for vulnerability-targeted fuzzing, matching the software_security lifecycle activity.", "source": "crawl"}, {"paper": {"id": "2603.00489", "title": "Does My README File Need To Be Updated? Exploring LLM-Based README Maintenance", "authors": ["Haoyu Gao", "Hong Yi Lin", "Christoph Treude", "Gregory Gay", "Mansooreh Zahedi"], "venue": "arXiv 2026/02", "category": "", "published": "2026-02-28", "links": {"paper": "https://arxiv.org/abs/2603.00489", "github": "", "website": ""}}, "category": "software_comprehension", "tags": ["empirical"], "summary": "This work formulates README update recommendation as a task and presents an LLM-driven framework that decides whether a pull request requires README updates, locates where changes are needed, and explains the trigger. It is evaluated on 25,511 PRs across 714 repositories plus a developer case study.", "reason": "Documentation maintenance/generation for a codebase is a comprehension-and-documentation lifecycle task under artifact/software.", "source": "crawl"}, {"paper": {"id": "2608.20685", "title": "Temporal Validity on Real Software Histories: Eliminating Stale-Fact Errors in Code-Assistant Memory over GitHub Fixes", "authors": ["Neeraj Yadav"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20685", "github": "", "website": ""}}, "category": "software_comprehension", "tags": [], "summary": "Introduces MemStrata, a deterministic supersession memory that eliminates stale-fact retrieval errors in code-assistant RAG, validated on real GitHub issue fixes from SWE-bench Lite/Verified where facts change over time.", "reason": "A memory resource serving code-assistant question-answering over evolving repository state, i.e. repo comprehension/QA.", "source": "crawl"}, {"paper": {"id": "2608.24221", "title": "DeepRepoQA: Code Repository Question Answering with Deep Agent Exploration", "authors": ["Weihan Peng", "Yuling Shi", "Yingwei Ma", "Longfei Yun", "Beijun Shen", "Xiaodong Gu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24221", "github": "", "website": ""}}, "category": "software_comprehension", "tags": [], "summary": "DeepRepoQA is an agentic repository question-answering framework that uses Monte-Carlo Tree Search over repository structure to enable multi-hop reasoning across files and dependencies. It shows substantial gains over baselines on the SWE-QA benchmark.", "reason": "Agent explores and answers questions about a codebase, matching the comprehension/QA leaf.", "source": "crawl"}, {"paper": {"id": "2605.26851", "title": "LLM-based Low-Level Integration Test Generation for Java", "authors": ["Qinghua Xu", "Guancheng Wang", "Lionel Briand", "Zhaoqiang Guo", "Kui Liu"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-26", "links": {"paper": "https://arxiv.org/abs/2605.26851", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "IntTestGen is an LLM-based approach for generating low-level integration tests for Java by mining dependency usage patterns and applying constraint-enforced repair. It outperforms prior LLM-based and search-based baselines on coverage and mutation score across Defects4J and a new benchmark.", "reason": "Proactive test generation for code correctness, independent of any specific reported bug, fits software_testing.", "source": "crawl"}, {"paper": {"id": "2608.21516", "title": "Neuro-Formal Verification: Agentic Language-Agnostic Formal Program Reasoning", "authors": ["Shuvendu K. Lahiri"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21516", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "Neuro-formal verification uses an AI coding agent to translate mainstream-language programs (e.g., Python) into verification-aware languages like Dafny or CBMC input, then checks correctness with an established verifier, achieving high precision on a bug/correctness dataset.", "reason": "Formal verification of software correctness performed by an agentic pipeline falls under software_testing per the taxonomy's explicit boundary.", "source": "crawl"}, {"paper": {"id": "2608.22630", "title": "VeGo: Direct Deductive Formal Verification of Go Programs for Computer Science Education", "authors": ["Tina Massoudi", "Chris Dutchyn"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22630", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "VeGo is a deductive formal verification system for Go programs using Hoare-style contracts, loop invariants, and SSA-based weakest-precondition calculus, aimed at CS education. It embeds verification annotations directly as Go comments.", "reason": "Formal verification of software correctness falls under software_testing per taxonomy boundary.", "source": "crawl"}, {"paper": {"id": "2608.23402", "title": "EXPLAIN Yourself! Finding Query Planner Stalls Across DBMSes", "authors": ["Geoffrey X. Yu", "Ryan Marcus", "Tim Kraska"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23402", "github": "", "website": ""}}, "category": "software_testing", "tags": ["benchmark"], "summary": "Uses lightweight agentic search to uncover pathological SQL queries that cause query planners across seven DBMSes to stall for minutes, releasing a curated benchmark of these robustness pathologies. The work proactively probes DBMS software for unknown performance/robustness defects.", "reason": "Proactive agentic search for unknown defects (planner stalls) in software independent of a reported bug fits software_testing.", "source": "crawl"}, {"paper": {"id": "2608.24473", "title": "Ockhamareto: Pareto-Gated Segment-Level Credit Assignment for Concise Unit-Test Generation with Reinforcement Learning", "authors": ["Dong Huang", "Mark Harman", "Jie M. Zhang", "Zhijiang Guo", "Mingzhe Du", "See Kiong Ng"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24473", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "Ockhamareto is a GRPO-based reinforcement learning framework for unit-test generation and selection, using Pareto-gated bonuses and token-level segment credit assignment to reward concise, high-mutation-kill test suites. It Pareto-dominates prior RL baselines on multiple benchmarks with fewer, more effective tests.", "reason": "Trains a model to generate/select unit tests, a proactive correctness-assurance lifecycle activity.", "source": "crawl"}, {"paper": {"id": "2608.24650", "title": "Simthesizer: An Agent-Driven Simulation Framework for LLM Serving Systems", "authors": ["Wonung Kim", "Hyunmin Choi", "Minsu Kim", "Jaehong Cho", "Yeongwook Kim", "Jongse Park"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24650", "github": "", "website": ""}}, "category": "software_development", "tags": [], "summary": "Simthesizer/Borg is a framework where a harnessed coding agent (Synthesizer agent) extends a composable LLM-serving simulator by translating natural-language feature requests into code within a shared simulator codebase. It achieves lower error and faster simulation than prior hand-built simulators.", "reason": "A coding agent implements new described functionality inside an existing simulator codebase, matching software_development's feature-implementation scope.", "source": "crawl"}, {"paper": {"id": "2608.24020", "title": "IterCAD: Iterative Program Repair for CAD Code Generation from Orthographic Views", "authors": ["Yuchuan Wu", "Ke Niu", "Haiyang Yu", "Zhuofan Chen", "Xiangyang Xue", "Bin Li"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24020", "github": "", "website": ""}}, "category": "cad", "tags": [], "summary": "IterCAD reframes orthographic-view-to-CAD generation as iterative program repair, letting a model inspect intermediate CAD results and decide to revise or stop, trained via a three-stage pipeline including multi-turn RL.", "reason": "Task's deliverable is executable parametric CAD code, matching the cad artifact leaf.", "source": "crawl"}, {"paper": {"id": "2608.24760", "title": "ExpConCAD: Experience-Guided Text-to-CAD Generation from Shape Descriptions with Implicit Spatial Constraints", "authors": ["Jingyao Liu", "Jinkang Tang", "Chen Huang", "Wenqiang Lei", "See-Kiong Ng"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24760", "github": "", "website": ""}}, "category": "cad", "tags": [], "summary": "ExpConCAD generates executable CadQuery CAD programs from underspecified natural-language shape descriptions by inferring missing spatial constraints using recovered construction structure and retrieved design experience. It targets the Text-to-CAD task of producing valid, executable CAD code.", "reason": "The task's purpose is producing executable CAD programs, matching the cad leaf under artifact.", "source": "crawl"}, {"paper": {"id": "2608.20903", "title": "Generation of Web Apps with Agentic IDEs: An Empirical Assessment", "authors": ["Manuel Marceca", "Maria Teresa Rossi", "Leonardo Mariani"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20903", "github": "", "website": ""}}, "category": "web", "tags": ["empirical"], "summary": "This paper empirically compares three agentic IDEs (Copilot, Cursor, Windsurf) on generating five full-stack web applications from scratch, finding strong results on common patterns like CRUD/auth but more errors on distributed architectures. It concludes agentic IDEs shift developers toward orchestrating LLM agents rather than replacing them.", "reason": "An empirical study focused on one task (web app generation) routes to that artifact leaf plus the empirical tag, since studies is reserved for field-wide/cross-cutting work.", "source": "crawl"}, {"paper": {"id": "2608.23067", "title": "Signal or Noise? A Benchmark Study of Agent Skills in Web Development", "authors": ["Ziyue Yang", "Fan Ding"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23067", "github": "", "website": ""}}, "category": "web", "tags": ["empirical", "benchmark"], "summary": "WebDev-Skills-Bench is a controlled empirical benchmark studying whether injecting reusable 'Agent Skills' into coding-agent prompts helps or hurts performance on 50 Web-Bench web development projects. It finds injected Skills often reduce Pass@2 and increase token cost across four LLMs.", "reason": "An empirical benchmark study confined to the web development activity, so it routes to web plus empirical/benchmark tags rather than the field-wide studies leaf.", "source": "crawl"}, {"paper": {"id": "2608.22390", "title": "SchemaGUI: A Schema-Driven Benchmark for Controllable GUI Generation Evaluation", "authors": ["Jiarui Dong", "Yin Cai", "Zhouhong Gu", "Chenmou Wu", "Ci Tao", "Yiran Chen", "Jialing Li", "Xiaoran Shi", "Juntao Zhang", "Zhijun Fang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22390", "github": "", "website": ""}}, "category": "web", "tags": ["benchmark"], "summary": "SchemaGUI is a schema-driven, template-based benchmark that synthesizes deterministic instruction-to-function-call pairs to evaluate controllable GUI generation by LLMs. It benchmarks several models and identifies geometric layout control as a key bottleneck.", "reason": "Evaluates LLMs generating GUI interface code/specifications from instructions, serving web/UI code production.", "source": "crawl"}, {"paper": {"id": "2608.21470", "title": "Structural Inference in Undocumented Mobile Databases: A Reproducible Benchmark for Evaluating Agentic Reasoning in Digital Forensics", "authors": ["Jeel Piyushkumar Khatiwala", "Divyangkumar Patel", "Weifeng Xu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21470", "github": "", "website": ""}}, "category": "database", "tags": ["benchmark"], "summary": "A reproducible benchmark evaluates agentic LLMs' ability to infer relational structure (table links, join paths) in undocumented mobile SQLite databases like SMS and Snapchat stores, scoring structural correctness against expert-verified SQL ground truth for digital forensics.", "reason": "The agent's task is producing correct SQL queries/joins from raw schemas, matching the database leaf's query-synthesis scope.", "source": "crawl"}, {"paper": {"id": "2608.22063", "title": "From SQL Generation to Tool Selection: A Domain-Oriented Pattern for MCP Servers", "authors": ["Bartolomeo Bogliolo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.22063", "github": "", "website": ""}}, "category": "database", "tags": ["benchmark"], "summary": "Proposes a Domain-Oriented Tooling Pattern for MCP database servers where LLM agents select parameterized query tools instead of generating SQL at query time, evaluated on a Sakila database benchmark.", "reason": "An LLM agent completing database tasks via selected/parameterized SQL tools; database is the deliverable/action domain, evaluated with a released reproducibility benchmark.", "source": "crawl"}, {"paper": {"id": "2608.22551", "title": "DAGSmith: Dependency-Aware Rewriting for dbt-Style SQL Pipelines", "authors": ["Jie Liu", "Lin Ma", "Barzan Mozafari"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22551", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "DAGSmith is a dependency-aware rewriting system for dbt-style SQL pipeline DAGs that uses an LLM to propose pipeline-level refactorings, validates equivalence, and retunes materialization choices, cutting runtime and cost substantially.", "reason": "The system autonomously generates and validates rewritten SQL code, a database-domain artifact-producing task.", "source": "crawl"}, {"paper": {"id": "2608.22651", "title": "Iteration Without Elaboration: A Simple ReAct Architecture Suffices for Text-to-SQL Generation", "authors": ["Jian Lu", "Haiwei Yu", "Raymond M Xiong", "Anru Zhang", "Danyang Zhuo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22651", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "ReAct-SQL is a zero-shot ReAct-style agent that iteratively issues typed DSL calls compiled to SQL, using execution feedback to revise its query construction, matching more elaborate text-to-SQL pipelines while running much faster.", "reason": "An agent that iteratively acts via execution feedback to produce SQL queries, matching the database leaf's text-to-SQL example.", "source": "crawl"}, {"paper": {"id": "2608.22476", "title": "JetStream: Generating Query Accelerators for Existing Database Systems", "authors": ["Akhilesh Balasingam", "Amadou Ngom", "Geoffrey X. Yu", "Tim Kraska"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22476", "github": "", "website": ""}}, "category": "database", "tags": [], "summary": "JetStream is an agentic workflow that generates query-specific accelerator code extending existing DBMSes, including engine-specific adapters and maintenance logic. It achieves large speedups over DuckDB/PostgreSQL on TPC-H while supporting dynamic workloads.", "reason": "An agent generates database-centric accelerator code, so it serves producing database code (artifact/database).", "source": "crawl"}, {"paper": {"id": "2608.23569", "title": "ESQ-Bench: A Multi-Tier Enterprise Oracle Benchmark for Evaluating NL2SQL Dialect Generalization and Silent Semantic Divergence", "authors": ["Sanjay Mishra", "Divya Chukkapalli", "Ganesh R. Naik"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-12", "links": {"paper": "https://arxiv.org/abs/2608.23569", "github": "", "website": ""}}, "category": "database", "tags": ["benchmark"], "summary": "a", "reason": "a", "source": "crawl"}]