Paper review 2026-08-04: 25 candidates #351

[{"paper": {"id": "2608.00122", "title": "Shared Organizational Memory for Enterprise Coding Agents: System Design and Deployment Snapshot", "authors": ["Harsh Rao Dhanyamraju", "Leonidas Raghav"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-31", "links": {"paper": "https://arxiv.org/abs/2608.00122", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "Describes a deployed shared organizational memory system that captures task-adjacent developer experience into curated question-answer memories, gated for security/privacy, to help enterprise coding agents reuse tacit internal-platform knowledge. Reports the system design and an operational deployment snapshot rather than evaluation results.", "reason": "A resource/system paper whose downstream task is helping coding agents produce better code using enterprise context, closest to code generation support.", "source": "crawl"}, {"paper": {"id": "2608.00494", "title": "TaPR: Test-Aware Policy Refinement for Feedback-Conditioned Code Generation", "authors": ["Aofan Liu", "Jingxiang Meng", "Fangxin Liu", "Yongbiao Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-01", "links": {"paper": "https://arxiv.org/abs/2608.00494", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "TaPR is a reinforcement learning framework that converts execution feedback into dense per-turn test-pass-ratio rewards to train multi-turn code agents' self-repair capability, evaluated on LiveCodeBench.", "reason": "A multi-turn agent repairs generated programs using execution feedback on general competitive-programming problems, serving code generation/completion rather than real-world issue resolution.", "source": "crawl"}, {"paper": {"id": "2608.01927", "title": "Effective and Efficient Context Retrieval via Partial Dependency Graph for Repository-Level Code Generation", "authors": ["Zhongxin Liu", "Zhonghao Jiang", "Zhifan Ye", "Haoye Wang", "Jiakun Liu", "Xiaoxue Ren"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.01927", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "DyRetriever performs repository-level context retrieval for code generation by dynamically building partial code dependency graphs via LLM-guided multi-hop reasoning, avoiding costly static global graphs. Combined into DyCoder, it improves Pass@1 on CoderEval and DevEval while being much faster than static graph baselines.", "reason": "Serves repository-level code generation/completion, matching software_code_generation.", "source": "crawl"}, {"paper": {"id": "2507.20402", "title": "CIgrate: Automating CI Service Migration with Large Language Models", "authors": ["Md Nazmul Hossain", "Taher A. Ghaleb"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2507.20402", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": [], "summary": "CIgrate is an LLM-based framework that automates migration of CI configurations between services (e.g., Travis CI to GitHub Actions), using zero-shot/few-shot prompting and fine-tuning, validated by submitting real pull requests. It substantially outperforms the rule-based baseline CIMig in accuracy and YAML validity.", "reason": "The task is CI/CD configuration migration, an enabling activity that makes code buildable/runnable, matching software_infrastructure.", "source": "crawl"}, {"paper": {"id": "2608.00947", "title": "Claim Plane: Reliability Gains and the Limits of Selective Concurrency for Parallel Coding Agents: A 30-Pair, Three-Seed Confirmatory Study of Deterministic Pre-Write Admission", "authors": ["Maxim Nikolaev"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2608.00947", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["empirical"], "summary": "Claim Plane is a deterministic pre-write admission mechanism that coordinates parallel coding agents editing the same codebase to prevent conflicting changes. A confirmatory study on CooperBench shows it substantially raises pair-pass and integration success rates, though largely by serializing execution.", "reason": "Coordinates concurrent code edits to avoid conflicts, a version-control-style enabling task under software_infrastructure.", "source": "crawl"}, {"paper": {"id": "2608.01189", "title": "MADE: Belief-Driven Dual-Agent Coordination for Autonomous Model Deployment", "authors": ["Yicheng Liu", "Bolin Zhang", "Weiran Liu", "Yakun Zhang", "Yangqin Jiang", "Zhiying Tu", "Dianhui Chu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2608.01189", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["benchmark"], "summary": "MADE is a dual-agent system that iteratively builds and validates deployment artifacts to turn heterogeneous open-source models into ready-to-call APIs, using belief updates from execution feedback. The authors also release M2ABench, a 122-model benchmark, showing MADE outperforms SWE-agent and OpenHands on deployment success rate.", "reason": "Agent writes/executes deployment code to make a model runnable as an API, an enabling/infrastructure task in service of usable software.", "source": "crawl"}, {"paper": {"id": "2608.01451", "title": "Doc2CI: A Multi-Service Study of CI Configuration Generation Using Large Language Models", "authors": ["Taher A. Ghaleb"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2608.01451", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["benchmark", "empirical"], "summary": "Doc2CI introduces a benchmark of 3,363 description-to-YAML CI configuration pairs and evaluates 16 LLMs on generating valid CI configs, analyzing failure modes and proposing schema-guided repair. It finds structural validity and reference similarity are distinct objectives requiring schema-aware tooling.", "reason": "A benchmark/resource paper whose downstream task is CI/CD configuration generation, matching software_infrastructure per resource-paper routing.", "source": "crawl"}, {"paper": {"id": "2608.01955", "title": "Agentic Self-Healing for Data and AI Pipelines: An Affordable Vendor-Agnostic Architecture using Open-Source Software", "authors": ["Solomon Eshun", "Dennis Murage", "Sharleen Muoki", "Chih-Chun Chen", "Stephen Adjignon", "Matteo Staar", "Oliver Angélil"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.01955", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["position"], "summary": "Proposes an affordable, vendor-agnostic open-source reference architecture for agentic self-healing of data, ML, and software delivery pipelines, combining monitoring, AI-assisted diagnosis, and controlled automated remediation. It compares existing commercial ZeroOps/AIOps tools and argues the gap is architectural rather than technological.", "reason": "Serves enabling-work of keeping build/data/CI pipelines running via automated diagnosis and repair, matching software_infrastructure.", "source": "crawl"}, {"paper": {"id": "2607.28835", "title": "From C to Idiomatic Rust: A Ship-of-Theseus Agentic Translation", "authors": ["Vasily A. Sartakov"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28835", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "Presents an agentic methodology for migrating mature C codebases to idiomatic Rust via a two-stage process: generating a semantics-preserving baseline then incrementally rewriting toward idiomatic Rust, validated by compilation and behavioral testing. Demonstrated on the 12.5k-SLOC iodine codebase.", "reason": "Behavior-preserving cross-language code migration is software_maintenance.", "source": "crawl"}, {"paper": {"id": "2608.00924", "title": "RefactorAssist: Agentic Refinement for Reliable Code Refactoring", "authors": ["Jonathan Cordeiro", "Shayan Noei", "Ying Zou"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2608.00924", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "RefactorAssist is an agentic system that iteratively repairs LLM-generated code refactorings by analyzing unit-test failures, error logs, and code diffs. It combines static fixes with context-aware repair, achieving up to 94.2% cumulative test pass rate on real Java projects.", "reason": "Refactoring is behavior-preserving code evolution, matching software_maintenance.", "source": "crawl"}, {"paper": {"id": "2607.29516", "title": "From Code Review to Code Critique: Intent, Drift, and Spotlight for AI-Generated Diffs at Scale", "authors": ["Chandra Maddila", "Mashrur Rashik", "Euna Mehnaz Khan", "Smriti Jha", "James Saindon", "Nachi Nagappan", "Peter C. Rigby"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-31", "links": {"paper": "https://arxiv.org/abs/2607.29516", "github": "", "website": ""}}, "category": "software_review", "tags": [], "summary": "ARCTIC is an AI code critique system that reframes review of AI-generated diffs around intent prediction, drift detection between developer intent and agent output, and spotlighting diff regions warranting scrutiny. It shows strong offline metrics and reduced code misalignment in production rollout.", "reason": "The task is judging AI-generated code changes for review purposes, matching software_review's definition.", "source": "crawl"}, {"paper": {"id": "2509.05372", "title": "Adversarial Bug Reports as a Security Risk in Language Model-Based Automated Program Repair", "authors": ["Piotr Przymus", "Andreas Happe", "Jürgen Cito"], "venue": "arXiv 2026/01", "category": "", "published": "2026-01-26", "links": {"paper": "https://arxiv.org/abs/2509.05372", "github": "", "website": ""}}, "category": "software_security", "tags": ["empirical"], "summary": "This empirical study shows that adversarial bug reports can mislead LLM-based Automated Program Repair systems into producing insecure or malicious code patches, with 90% attack success and inadequate pre/post-repair defenses. The authors build a threat model and a prototype framework for generating such adversarial reports at scale.", "reason": "Studies a security vulnerability in the code-patch-producing pipeline of APR agents, fitting the security-assurance activity on code.", "source": "crawl"}, {"paper": {"id": "2607.29422", "title": "AgenticRepair: Multi-Faceted Program Context Engineering for Agentic Vulnerability Repair", "authors": ["Michael Fu", "Qiyue Mei", "Patanamon Thongtanunam", "Kla Tantithamthavorn"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-31", "links": {"paper": "https://arxiv.org/abs/2607.29422", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "AgenticRepair is a multi-agent framework that engineers code-structure, runtime-execution, and commit-history context to condition patch synthesis for automated vulnerability repair. Evaluated on SEC-Bench's 300 real-world CVE instances, it achieves 73% repair success, beating the strongest baseline by 29%.", "reason": "Task is vulnerability patch repair, which falls under software_security per the boundary that CVE repair belongs there.", "source": "crawl"}, {"paper": {"id": "2608.00692", "title": "Vul4Py: Benchmarking Automated Vulnerability Repair in Python with Paired Exploit and Functional Oracles", "authors": ["Tan Bui", "Ting Zhang", "Ferdian Thung", "Yunpeng Xiong", "Penghao Jiang", "Xin Zhou", "David Lo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-01", "links": {"paper": "https://arxiv.org/abs/2608.00692", "github": "", "website": ""}}, "category": "software_security", "tags": ["benchmark"], "summary": "Introduces Vul4Py, a Python automated vulnerability repair benchmark with 100 real CVEs each carrying paired exploit and functional pytest oracles, addressing false-accept issues in prior benchmarks. Evaluation shows SWE-agent OpenHands substantially outperforms directly prompted LLMs and specialized repair tools.", "reason": "Benchmark serving CVE/vulnerability repair by code agents falls under software_security per boundary rules.", "source": "crawl"}, {"paper": {"id": "2608.00739", "title": "EduPluginBench: Executable Assurance for AI-Generated Educational Plugins", "authors": ["Nizam Kadir"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-01", "links": {"paper": "https://arxiv.org/abs/2608.00739", "github": "", "website": ""}}, "category": "software_security", "tags": ["benchmark"], "summary": "EduPluginBench is an executable benchmark and staged admission method assessing AI-generated educational plugins for compliance with privilege, consent, provenance and safety constraints. It reveals that current coding models' generated plugins largely fail security/governance conformance checks.", "reason": "The benchmark assesses security/compliance properties of AI-generated code, serving the vulnerability/assurance activity (software_security).", "source": "crawl"}, {"paper": {"id": "2608.02001", "title": "VulnGym: Benchmarking Coding Agents for Repository-Level Vulnerability Detection", "authors": ["Kexing Ji", "Jiachen Liu", "Enze Hu", "Cuiyun Gao", "Keke Lian", "Yongheng Liu", "Lei Zhang", "Tian Dong", "Hao Chen", "Wang Bin"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.02001", "github": "", "website": ""}}, "category": "software_security", "tags": ["benchmark"], "summary": "VulnGym is a repository-level benchmark that evaluates coding agents on autonomously exploring codebases to locate and diagnose vulnerabilities, using fine-grained ground-truth traces from real GitHub advisories. It reveals current agents' limitations in end-to-end detection and evidence construction.", "reason": "Task is vulnerability detection on repository code, served by coding agents; routes to software_security per benchmark routing.", "source": "crawl"}, {"paper": {"id": "2608.02407", "title": "Antares: Foundation Models for Agentic Vulnerability Localization", "authors": ["Supriti Vijay", "Aman Priyanshu", "Didier Chapoteau", "Arthur Goldblatt", "Jianliang He", "Kimia Majd", "Fraser Burch", "Baturay Saglam", "Takahiro Matsumoto", "Zhuoran Yang", "Amin Karbasi"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2608.02407", "github": "", "website": ""}}, "category": "software_security", "tags": ["model"], "summary": "Antares is a family of compact (350M-3B) foundation models trained via SFT and RL for agentic vulnerability localization, iteratively exploring repositories to find vulnerable code, matching much larger models at a fraction of cost.", "reason": "A model specialized and evaluated for vulnerability localization, a security-assurance task on code, routes to software_security plus model tag.", "source": "crawl"}, {"paper": {"id": "2608.01507", "title": "Deep Agentic Search for Repository-Level Code Question Answering: An Empirical Study", "authors": ["Amirkia Rafiei Oskooei", "Bora Ilci", "Alperen Kayim", "Mehmet Egemen Uzun", "Berat Can", "Kaan Emre Kara", "Ozan Orhan", "Mehmet S. Aktas"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2608.01507", "github": "", "website": ""}}, "category": "software_comprehension", "tags": ["empirical"], "summary": "An empirical study comparing semantic search versus deep agentic (subagent-delegated) search for repository-level code question answering, finding semantic search more accurate and cheaper on SWE-QA. It taxonomizes failure modes, showing agentic hand-offs introduce silent errors.", "reason": "Narrow empirical study of one activity (repository code QA/comprehension), not field-wide, so routes to software_comprehension rather than studies.", "source": "crawl"}, {"paper": {"id": "2605.01660", "title": "Testing, Credible Compilation, and Verification in the Axon Verified Compiler in Lean and Claude Code", "authors": ["Martin Rinard"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-31", "links": {"paper": "https://arxiv.org/abs/2605.01660", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "This paper describes using Claude Code as a coding agent to write Lean code and machine-checked correctness proofs for the Axon verified compiler, combining testing, translation validation, and formal verification. The proofs eliminate the need to audit generated code, demonstrating an agent-driven formal verification workflow.", "reason": "Formal, machine-checked verification of a software compiler by a coding agent falls under software_testing's formal-verification scope.", "source": "crawl"}, {"paper": {"id": "2607.28775", "title": "Repository-Aware Metamorphic Relation Generation for Augmented Reality Applications using Large Language Models", "authors": ["Dibyendu Brinto Bose", "Jiawei Qin", "Chris Brown"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-30", "links": {"paper": "https://arxiv.org/abs/2607.28775", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "Introduces an LLM-based pipeline that generates and agentically refines metamorphic relations as test oracles for augmented reality applications using repository-level context, evaluated on 142 mobile AR repositories. Refined relations were validated as logically sound and convertible into executable tests that detect mutations.", "reason": "Generates test oracles/executable tests proactively (no specific reported defect), matching software_testing.", "source": "crawl"}, {"paper": {"id": "2607.28928", "title": "Automated Testing and Repair for Verified Compilers Generated by a Coding Agent", "authors": ["Martin Rinard"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-31", "links": {"paper": "https://arxiv.org/abs/2607.28928", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "Presents an agent-based automated testing and repair system for verified compilers containing verified, checked, unverified code and specifications, using specialized defect-detection techniques and a coding agent to repair surfaced defects. Evaluated on the Axon compiler, checking for reward hacking in both generation and repair.", "reason": "Proactive defect detection plus repair of compiler code independent of any single reported bug fits software_testing's inclusion of formal verification and correctness assurance.", "source": "crawl"}, {"paper": {"id": "2604.23509", "title": "Uncovering Business Logic Bugs via Semantics-Driven Unit Test Generation", "authors": ["Chen Yang", "Junjie Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-03", "links": {"paper": "https://arxiv.org/abs/2604.23509", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "SeGa is a semantics-driven unit test generation technique that builds a knowledge base from requirement documents to derive business scenarios guiding LLM-based test generation, uncovering business logic bugs missed by code-centric approaches. Evaluated on industrial Go projects, it outperforms state-of-the-art baselines and finds new confirmed bugs in production.", "reason": "Unit test generation for proactive correctness assurance falls under software_testing.", "source": "crawl"}, {"paper": {"id": "2605.03356", "title": "POSTCONDBENCH: Benchmarking Correctness and Completeness in Formal Postcondition Inference", "authors": ["Gehao Zhang", "Juan Zhai"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-05", "links": {"paper": "https://arxiv.org/abs/2605.03356", "github": "", "website": ""}}, "category": "software_testing", "tags": ["benchmark"], "summary": "POSTCONDBENCH is a multilingual benchmark of 420 real-world Python and Java tasks for evaluating LLM-generated formal postconditions, measuring both correctness and completeness via a defect-discrimination execution environment. It reveals a substantial correctness-completeness gap across five SOTA LLMs.", "reason": "A benchmark serving formal verification/postcondition generation, a software_testing lifecycle activity.", "source": "crawl"}, {"paper": {"id": "2605.03822", "title": "KVerus: Scalable and Resilient Formal Verification Proof Generation for Rust Code", "authors": ["Yuwei Liu", "Xinyi Wan", "Yanhao Wang", "Minghua Wang", "Lin Huang", "Tao Wei"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2605.03822", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "KVerus is a retrieval-augmented, self-adaptive system for generating and repairing Verus-based Rust formal verification proofs, bridging the semantic-structural gap between LLM code patterns and rigid verification dependencies. It substantially outperforms prior tools on single-file and repository-level benchmarks and produces upstream-accepted proofs for the Asterinas OS kernel.", "reason": "An agentic system performing formal verification proof generation and repair, which is software_testing (formal verification).", "source": "crawl"}, {"paper": {"id": "2605.21930", "title": "PITMuS: A Tool for Automated Bug Dataset Generation via Source-Level Mutant Reconstruction", "authors": ["Tasfia Tasnim", "Soneya Binta Hossain"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-02", "links": {"paper": "https://arxiv.org/abs/2605.21930", "github": "", "website": ""}}, "category": "software_testing", "tags": ["training-data"], "summary": "PITMuS reconstructs PIT's bytecode-level Java mutants into inspectable source-level edits, producing reusable mutation-based bug datasets with high coverage across open-source systems. These datasets serve mutation testing, program repair, and learning-based software engineering research.", "reason": "A deterministic tool generating reusable mutation-testing bug datasets serves proactive correctness-testing research, matching software_testing plus training-data.", "source": "crawl"}]