Paper review 2026-08-26: 25 candidates #362

[{"paper": {"id": "2608.24271", "title": "Observability and Fault Injection for LLM-Based Multi-Agent Systems in Software Engineering", "authors": ["Zahra Seyedghorban", "Egor Klimov", "Arie van Deursen", "Annibale Panichella", "Burcu Kulahcioglu Ozkan"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24271", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "llmmas-otel is a framework-agnostic observability and fault-injection tool that instruments LLM-based multi-agent software engineering systems with distributed tracing, enabling reproducible comparison of baseline versus faulty agent executions. It aims to make multi-agent SE workflows inspectable and debuggable.", "reason": "Tooling to inspect and debug agentic SE systems, analogous to the AgentLens precedent, routes to software_debugging rather than studies.", "source": "crawl"}, {"paper": {"id": "2507.20923", "title": "Pareto-Grid-Guided Large Language Models for Fast and High-Quality Heuristics Design in Multi-Objective Combinatorial Optimization", "authors": ["Minh Hieu Ha", "Hung Phan", "Tung Duy Doan", "Tung Dao", "Dao Tran", "Huynh Thi Thanh Binh"], "venue": "arXiv 2025/07", "category": "", "published": "2025-07-28", "links": {"paper": "https://arxiv.org/abs/2507.20923", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "MPaGE combines LLM-driven code synthesis with a Pareto-Front-Grid-guided evolutionary framework to automatically generate diverse heuristics for multi-objective combinatorial optimization problems.", "reason": "The LLM synthesizes heuristic programs as the deliverable, so the task's purpose is producing code (program synthesis).", "source": "crawl"}, {"paper": {"id": "2608.21483", "title": "SLICE: Specification-Level Isolation of Contract Enforcement", "authors": ["Soohan Lim", "Hyundong Jin", "Yo-Sub Han"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21483", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "SLICE is a code generation framework that separately identifies functional requirements and input contracts from specifications, then generates function bodies and contract-enforcing assertions in distinct stages. It improves joint satisfaction of functionality and input validation over baselines on ContractEval.", "reason": "A method for generating function-level code from natural-language specifications, matching code generation & completion.", "source": "crawl"}, {"paper": {"id": "2608.22529", "title": "Benchmarking the Titans: A Multi-Dimensional Empirical Evaluation of LLM Code Generation Quality in the .NET Ecosystem", "authors": ["Seyed Mohammad Mahdi Ghalandarian", "Majid Bazargani", "Masoumeh Taromirad"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22529", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["empirical", "benchmark"], "summary": "A multi-dimensional benchmark evaluates LLM (GPT, Gemini, Claude, Grok) code generation quality in C#/.NET across correctness, static code quality, and runtime efficiency on 85 HumanEval-derived tasks. It finds weak correlation between functional correctness and code quality metrics.", "reason": "A benchmark/empirical study of code generation quality for a single activity (code generation) in a specific ecosystem.", "source": "crawl"}, {"paper": {"id": "2608.23552", "title": "Prime Agent: A Self-Improving RLM Harness", "authors": ["Seth Karten", "Alex L. Zhang", "Kevin Thomas", "Sebastian Müller", "Elie Bakouch", "Daniel Auras", "Mika Senghaas", "Fares Obeid", "Konstantin Dunas", "Johannes Hagemann", "Sami Jaghouar"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23552", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "Prime Agent is an open-source, self-improving harness with a persistent REPL, recursive subagents, and continual memory for long-horizon coding-agent workflows, evaluated on long-context coding, GPU-kernel generation, emulator construction, ML training speedruns, and game environments.", "reason": "A generalist coding-agent harness positioned primarily around coding-agent workflows; routed by its dominant coding-oriented benchmarks to software_code_generation.", "source": "crawl"}, {"paper": {"id": "2608.23632", "title": "Function-Level Execution Feedback for Code Preference Optimization", "authors": ["Idris Nechnech", "Sehwan Kim", "Jimin Seo", "Yeongoon Kim", "Minhae Oh", "Sangwoo Hong", "Jungwoo Lee"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.23632", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "STEP-KTODER is a code-specific preference optimization framework that defines steps as module-level functions in decomposed programs, using unit-test-derived binary correctness labels for stepwise KTO training. It improves over outcome-only KTO/DPO baselines on HumanEval(+), MBPP(+), BigCodeBench, and LiveCodeBench.", "reason": "A code-specific (not domain-agnostic) training method whose downstream task is code generation, per the artifact/software_code_generation leaf.", "source": "crawl"}, {"paper": {"id": "2608.23653", "title": "Beyond Executable Models: The Pufibara Agent Harness and the Modelica Agent Workflow Benchmark for Physical System Modeling", "authors": ["Zizhe Wang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23653", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark"], "summary": "Introduces Pufibara, an agent harness with persistent engineering state for Modelica physical-system modeling, and a 232-task benchmark spanning model repair, generation, and tuning, outperforming Claude Code under matched LLM backends.", "reason": "Serves producing correct Modelica code artifacts across repair/generation/tuning; treated as code-generation-centric since revision-tracked model production is the core loop.", "source": "crawl"}, {"paper": {"id": "2608.23740", "title": "AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace", "authors": ["Seonglae Cho", "Donghyun Lee"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23740", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "AgentRoom introduces a CRDT-backed realtime collaborative editing protocol letting multiple coding agents concurrently claim files and coordinate via MCP tools, reducing task abandonment versus solo and phase-sequenced multi-agent baselines.", "reason": "A multi-agent system whose product is collaboratively generated code across coding tasks, matching the MetaGPT/ChatDev-style software_code_generation example.", "source": "crawl"}, {"paper": {"id": "2608.23842", "title": "Automated Synthesis of Cloud Emulators", "authors": ["Archit Bhatnagar", "Zhenning Yang", "Sarah McClure", "Yiming Qiu", "Sylvia Ratnasamy", "Ang Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23842", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "CloudEmu automatically synthesizes cloud service emulators from documentation using neurosymbolic code synthesis, combining LLM code generation with symbolic abstractions and real-cloud-oracle-based testing and repair. It outperforms the manually built LocalStack on AWS/GCP coverage and accuracy.", "reason": "The task's ultimate purpose is producing code (emulator implementations) via an LLM-driven synthesis-test-repair loop, routing to software_code_generation.", "source": "crawl"}, {"paper": {"id": "2608.23918", "title": "MARS: Multi-Specialist LLM Relay System for Competitive Programming", "authors": ["Andrei Mikhailov", "Mikhail Burtsev", "Alsu Sagirova"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23918", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "MARS is a prompt-only multi-agent relay system where topic-specialist LLMs (grounded via RAG over algorithm theory) collaboratively write and refine C++ solutions to competitive programming problems, tested against public examples in a sandbox. It improves pass rate on CodeContests over direct prompting at lower cost than a strong baseline.", "reason": "Competitive programming code generation via a multi-agent pipeline serves producing code, per software_code_generation examples.", "source": "crawl"}, {"paper": {"id": "2608.24135", "title": "Robust Code RL via Faulty-Code-Driven Test case Synthesis and Dense Reward Shaping", "authors": ["Yiwen Zhang", "Xiaodong Yan", "Zhenyu Huang", "Deng Zhao", "Liang Jiang", "Qing Cui", "Zujie Wen", "Zhiqiang Zhang", "Jun Zhou"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24135", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["training-data"], "summary": "RobustTests improves RL training for LLM code generation by synthesizing faulty-code-driven test cases and applying dense, pass-rate-based reward shaping to reduce reward hacking, augmenting CodeContests and boosting LiveCodeBench performance.", "reason": "Downstream task served is competitive-programming code generation via RLVR, so it routes to software_code_generation with a training-data resource tag.", "source": "crawl"}, {"paper": {"id": "2608.24644", "title": "A Literate Programming Environment for Human and Machine Agents", "authors": ["Adam T. Burke"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24644", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "This paper introduces a literate programming environment combining executable program essays, a name-aware parser, and a name-graph to co-locate code with natural language context for both human and LLM coding agents. It aims to give coding agents IDE-like symbol search and usage capabilities to better produce code.", "reason": "An environment enabling machine agents to write and organize code, serving code production, fits software_code_generation.", "source": "crawl"}, {"paper": {"id": "2608.20525", "title": "Bolo: Verified Model Hub for Next-Generation AI Databases", "authors": ["Yunqi Li", "Ila Petrovic", "Yongjoo Park"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-20", "links": {"paper": "https://arxiv.org/abs/2608.20525", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": [], "summary": "Bolo is an agentic system that repairs broken model-inference pipelines and synthesizes runnable pipelines from documentation for models on hubs like Hugging Face, verifying them via multi-stage structural and semantic checks. It aims to build a large-scale verified model hub of ready-to-use code pipelines.", "reason": "The agent repairs/synthesizes code so pipelines become buildable and runnable, matching software_infrastructure's build/environment-repair scope.", "source": "crawl"}, {"paper": {"id": "2606.05518", "title": "Latent Reasoning Guidance for Parallel Code Translation", "authors": ["Tomer Bitan", "Erel Kaplan", "Roee Bar-Yadin", "Lian Ghrayeb", "Le Chen", "Samyak Jhaveri", "Niranjan Hasabnis", "Gal Oren"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-03", "links": {"paper": "https://arxiv.org/abs/2606.05518", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "Proposes a test-time latent-reasoning guidance method that trains a Process Reward Model over hidden-state prefixes to select promising continuations before code is decoded, improving executable parallel-code translation validation rates.", "reason": "Serves code translation between languages, a behavior-preserving evolution task under software_maintenance.", "source": "crawl"}, {"paper": {"id": "2608.20776", "title": "An Extensive Empirical Study on Code Translation Technique", "authors": ["Ruihang Fan", "Jiajun Jiang", "Xinpeng Wang", "Jiateng Fu", "Fengjie Li", "Jiasi Shen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20776", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["empirical"], "summary": "A large-scale empirical study comparing learning-based, LLM-based, and general-purpose LLM code translation techniques across languages and granularities, analyzing correctness, similarity, and failure patterns.", "reason": "Single-activity empirical study of code migration/translation techniques, per studies boundary routing single-activity empirical work to that leaf.", "source": "crawl"}, {"paper": {"id": "2608.21208", "title": "Specification Portability Across LLM Development Agents: Cross-Agent Compatibility in Specification-Driven Software Migration", "authors": ["Oleg Grynets", "Oleksii Ilchuk", "Dariia Zatulna", "Vasyl Lyashkevych"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21208", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["empirical"], "summary": "Empirically evaluates cross-agent portability of migration specifications using Oracle-to-PostgreSQL transformation across LLM development agents (Kiro, Gemini, Copilot, Claude Code, Cursor). Finds that specifications are not agent-neutral and cross-agent transfer can severely degrade output quality.", "reason": "An empirical study of a single lifecycle activity, behavior-preserving code migration, routes to software_maintenance plus empirical tag.", "source": "crawl"}, {"paper": {"id": "2608.23564", "title": "SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration?", "authors": ["Deyao Hong", "Yizhe Chi", "Wenyi Li", "Xiaoqiu Wang", "Mingju Gao", "Kaisen Yang", "Bingxiang He", "Youjie Zheng", "Calvin Xiao", "Qinhuai Na"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23564", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["benchmark"], "summary": "Introduces SWE Refactor Bench, a benchmark of 20 whole-repository technical-debt migrations evaluated via a three-stage protocol checking that migration actually occurred and behavior is preserved. Frontier coding agents perform poorly, especially on language rewrites.", "reason": "Benchmark tests behavior-preserving repository migration by coding agents, which is software_maintenance.", "source": "crawl"}, {"paper": {"id": "2608.23118", "title": "From Metrics to Improvement: A Lifecycle-Aware LLM Feedback Framework for Research Software Quality", "authors": ["Nafis Tanveer Islam", "Nafiseh Soveizi", "Yutong Li", "Zhiming Zhao"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23118", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "Proposes a lifecycle-aware quality model plus an iterative LLM-based code refinement framework that improves research software along maintainability and structural-quality metrics.", "reason": "Iterative LLM code refinement targeting code quality/duplication/maintainability is behavior-preserving evolution, i.e. software_maintenance.", "source": "crawl"}, {"paper": {"id": "2608.23146", "title": "An AI-Assisted Migration Framework for Transforming Legacy Scientific Applications into Reusable Cloud-Based Workflows", "authors": ["Nafiseh Soveizi", "Sven Tesselaar", "Hero Robinson Brouwer", "Zhiming Zhao"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23146", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "Proposes an AI-assisted framework combining RM-ODP analysis, LLMs, and DSM refinement to migrate legacy scientific applications into reusable, containerized cloud-native workflows while preserving functionality. It is evaluated on two legacy scientific applications from different domains.", "reason": "Behavior-preserving transformation of legacy code into a new execution environment fits software_maintenance's migration scope.", "source": "crawl"}, {"paper": {"id": "2608.23611", "title": "REFINE: A Multi-Agent LLM Approach for Evidence-Guided Code Refactoring", "authors": ["Muhammad Waseem", "Aakash Ahmad", "Pekka Abrahamsson"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.23611", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "REFINE is a tool-agnostic multi-agent LLM system that performs evidence-guided, file-level Java refactoring by combining static-analysis smell detection, planning, transformation, and preservation checks. Evaluated on 450 files across 15 open-source systems, it substantially reduces code smells while flagging residual behavioral risks.", "reason": "Behavior-preserving code transformation (refactoring) matches the software_maintenance leaf.", "source": "crawl"}, {"paper": {"id": "2608.24153", "title": "pigzpp: Fast, Parallel, Portable Compression for the Whole Stack", "authors": ["Thamme Gowda"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24153", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "pigzpp is a from-scratch, thread-safe C++23 rewrite of the pigz parallel gzip utility exposing one compression core to multiple languages, developed as a case study in AI-assisted modernization validated by automated compatibility tests.", "reason": "A behavior-preserving rewrite/migration of an existing library using AI-assisted development matches software_maintenance's migration and translation scope.", "source": "crawl"}, {"paper": {"id": "2608.21311", "title": "AI-to-AI Code Reviews of GitHub Pull Requests", "authors": ["Niruthiha Selvanayagam", "Taher A. Ghaleb"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21311", "github": "", "website": ""}}, "category": "software_review", "tags": ["empirical"], "summary": "An empirical, large-scale study of AI coding agents reviewing pull requests authored by other AI agents on GitHub, comparing cross-product versus same-product review behavior, comment types, and latency. Built from a dataset of over 248,000 AI-authored, AI-reviewed PRs.", "reason": "An empirical study focused on a single activity, automated code review, routes to software_review plus empirical tag.", "source": "crawl"}, {"paper": {"id": "2608.20637", "title": "ARQ: Agentic CodeQL Query Refinement for C/C++ Vulnerability Detection", "authors": ["Chunyi Wang", "Yunfei Ke", "Junfeng Yang", "Yun-Yun Tsai", "Penghui Li"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20637", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "ARQ is an agentic framework that refines CodeQL vulnerability-detection queries for C/C++ using execution-grounded disagreements from synthesized programs, reducing false positives/negatives without labeled data. It fixed long-open GitHub CodeQL issues and found new bugs in libpng and zlib.", "reason": "Agentic refinement of vulnerability-detection queries serves security-assurance work on code, matching software_security.", "source": "crawl"}, {"paper": {"id": "2608.20628", "title": "Behavior Specification-Guided Program Synthesis for Binary Deobfuscation", "authors": ["Kangchen Zhu", "Shangwen Wang", "Zhiliang Tian", "Zhouyang Jia", "Xiaoling Li", "Jun Ma", "Jie Yu", "Xiaoguang Mao"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20628", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "BinMirror reframes binary deobfuscation as behavior-specification-guided program synthesis, using dynamic execution traces to synthesize and validate source code that preserves obfuscated binaries' runtime behavior for security/reverse-engineering analysis.", "reason": "Produces code to restore analyzability of obfuscated binaries in service of security analysis, matching software_security.", "source": "crawl"}, {"paper": {"id": "2608.20963", "title": "Vibe Coding and Web Application Security: A Twin-Prompt Study", "authors": ["Darko Andročec"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20963", "github": "", "website": ""}}, "category": "software_security", "tags": ["empirical"], "summary": "A twin-prompt study comparing web applications generated by an agentic coding assistant with and without explicit security-requirement prompts, finding the security-aware variant yields far fewer and less severe confirmed vulnerabilities.", "reason": "Single-activity empirical study of AI-generated code security routes to software_security plus empirical tag, not the field-wide studies leaf.", "source": "crawl"}]