Paper review 2026-08-26: 25 candidates #361

[{"paper": {"id": "2601.19100", "title": "Reward Engineering for Software Tasks: A Survey of Reinforcement Learning Approaches", "authors": ["Md Rayhanul Masud", "Azmine Toushik Wasi", "Salman Rahman", "Md Rizwan Parvez"], "venue": "arXiv 2026/01", "category": "", "published": "2026-01-27", "links": {"paper": "https://arxiv.org/abs/2601.19100", "github": "", "website": ""}}, "category": "studies", "tags": ["survey"], "summary": "A systematic survey of reward engineering for reinforcement learning applied to code-centric software engineering tasks, organizing reward sources, granularity, and aggregation across generation, repair, testing, and optimization.", "reason": "Cross-cutting survey spanning multiple software lifecycle activities with no single dominant leaf, so it belongs to studies rather than any one activity.", "source": "crawl"}, {"paper": {"id": "2608.20446", "title": "Vibe Coding: Practice, Performance, Productivity, and Risk -A State-of-the-Art Review", "authors": ["Dominik L. Michels", "Mutaz Abu Ghazaleh", "Francois Lazzari", "Nabil Kassem", "Jonathan Klein"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-20", "links": {"paper": "https://arxiv.org/abs/2608.20446", "github": "", "website": ""}}, "category": "studies", "tags": ["survey"], "summary": "A cross-disciplinary state-of-the-art review synthesizing empirical evidence on 'vibe coding' practice, covering productivity effects, code quality, security, and skill atrophy. It reconciles contradictory productivity findings across field experiments, trials, and telemetry studies.", "reason": "Field-wide empirical/survey study whose object is the practice of AI-assisted coding itself, not a proposed agent or task.", "source": "crawl"}, {"paper": {"id": "2608.20622", "title": "Applying Anthropic Primitives at Large Enterprises: Harness Paradigm for Knowledge Work", "authors": ["George Juraj Salapa"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-20", "links": {"paper": "https://arxiv.org/abs/2608.20622", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "Proposes a harness paradigm for deploying coding-agent harnesses as governed enterprise infrastructure, arguing harness choice dominates agent performance and adoption requires closing a governance gap, with mechanisms like credential-scoped tooling and unified deployment across cron, chat, and terminal.", "reason": "A vision/position paper arguing for a new architecture and paradigm around coding-agent harnesses rather than proposing an agent for a specific task, matching the studies position-paper definition.", "source": "crawl"}, {"paper": {"id": "2608.20341", "title": "SDAD: Spec-Driven Agentic Development for the AI-Native SDLC", "authors": ["Vu Hung Nguyen", "Thanh Nguyen"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-05", "links": {"paper": "https://arxiv.org/abs/2608.20341", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "A position/vision report formalizing Spec-Driven Agentic Development (SDAD) as a new SDLC paradigm built around frontier coding agents, comparing it to Agile and proposing governance metrics and adoption roadmaps. It does not propose or evaluate a specific agent or benchmark.", "reason": "Object of study is the practice/paradigm of agentic software development itself, not a task-performing agent, so it is a field-wide position paper.", "source": "crawl"}, {"paper": {"id": "2608.21107", "title": "Large Language Models at the Intersection of Software Engineering and Software Security:An Evidence-Centered Structured Survey and Research Agenda", "authors": ["Wei Lin", "Tao Zhou", "Zhaofei Xie", "Changgui Hong"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21107", "github": "", "website": ""}}, "category": "studies", "tags": ["survey"], "summary": "A structured, evidence-centered survey synthesizing research on LLM agents at the intersection of software engineering and software security, spanning task taxonomies, adaptation mechanisms, and evaluation validity threats. Proposes an assurance framework and research agenda for jointly secure-and-functional agent evaluation.", "reason": "A field-wide, cross-activity survey whose object is the agents and evaluation practices themselves, not one task, so it belongs in studies.", "source": "crawl"}, {"paper": {"id": "2608.21747", "title": "Architecture as Capability Equalizer for Coding Agents", "authors": ["Arquimedes Canedo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.21747", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A controlled experiment measuring how five architecture-specification formats (prose, diagrams, OpenAPI, DSL, typed contracts) affect the quality of code generated by coding agents across six LLMs, finding format matters far more for weaker models.", "reason": "Its object of study is coding agents' behavior under varying inputs, not a proposed agent/method for a task, so it is an empirical study.", "source": "crawl"}, {"paper": {"id": "2608.21884", "title": "Loop Engineering: Building Blocks, Adoption, and Impact", "authors": ["Jai Lal Lulla", "Vahram Nersesyan", "Seyedmoein Mohsenimofidi", "Christoph Treude", "Sebastian Baltes"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.21884", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An exploratory empirical study of 'loop engineering,' the emerging practice of designing systems that autonomously trigger and stop agentic coding tool runs. Mines 36,710 repositories to characterize adoption of agent loops and their traceability in version control.", "reason": "The paper's object of study is the practice and adoption of agentic coding tools themselves, matching the empirical-study definition of studies.", "source": "crawl"}, {"paper": {"id": "2608.22960", "title": "What Process Evaluation of Coding Agents Actually Measures: Action, Task, and Step Are Three Different Levels", "authors": ["Jiawei He", "Mengyu Shi", "Jie jia", "Xikai Yang", "Dong Sun"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.22960", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "Proposes a measurement framework distinguishing action, task, and step levels in process evaluation of coding agents, instantiated as SCAE, a replay-based causal attribution estimator. Experiments on file-localization episodes reveal biases like collider bias in full-trace judges.", "reason": "The paper's object of study is the methodology for evaluating coding agents rather than performing a coding task itself, fitting studies.", "source": "crawl"}, {"paper": {"id": "2608.23550", "title": "When \"Do Not\" Is Not Deny: Security Rules in CLAUDE.md vs Built-In Controls", "authors": ["Ting Yan"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23550", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "Empirically measures the gap between natural-language security rules written in Claude Code's CLAUDE.md files and the tool's built-in enforceable controls across 481 public files, finding only 4-16% of rules have a matching enforced control.", "reason": "An empirical/behavioral study of how a coding agent's security configuration practice works, not a method that performs a task.", "source": "crawl"}, {"paper": {"id": "2608.23610", "title": "From Traceability to Justifiability: Accountability Structures in Agentic Software Engineering", "authors": ["Rashid Azarang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.23610", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical documentation audit of 47 CI/CD and agent-serving platforms assessing whether their published records can substantiate claims of behavioral identity between evaluated and deployed AI systems. It finds systemic gaps in identity binding and verifiable evidence for agentic software engineering pipelines.", "reason": "Object of study is accountability practices in agentic SE platforms, not a proposed task-performing agent, so it fits studies/empirical.", "source": "crawl"}, {"paper": {"id": "2608.23651", "title": "Feedback That Backfires: Why Small Language Model Agents Repeat the Call They Just Watched Fail", "authors": ["Esmail Gumaan"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23651", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "Empirically shows small LM agents are more likely to repeat a tool call after seeing it fail, across tool-calling and MBPP program-repair environments, and identifies harness fixes that mitigate the effect.", "reason": "An empirical/behavioral study of agent harness design and failure feedback, not a task-performing method, so it belongs to studies per the off-axis rule.", "source": "crawl"}, {"paper": {"id": "2608.23953", "title": "The Empire, Long Divided, Must Unite: Architectural Convergence in Three LLM Agent Harnesses", "authors": ["Dai Jiahong"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.23953", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A source-level, multi-case empirical study comparing the architecture of three open coding-agent harnesses (LangChain deepagents, Earendil pi, DeepSeek dsh), finding convergence toward a common set of five architectural elements despite opposing design philosophies. Notes a shared gap in external verifiability across all three.", "reason": "The object of study is the agent harnesses themselves (architecture/design convergence), matching the studies/empirical definition.", "source": "crawl"}, {"paper": {"id": "2608.24358", "title": "The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents", "authors": ["Roy Ganz", "Mor Shpigel Nacson", "Adi Kalyanpur", "Ron Litman"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24358", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical study of long-running coding agents examining how switching between low-cost and high-cost LLMs mid-trajectory affects quality and cost, quantifying a 'handoff tax' when receivers continue another model's non-native trajectory. Tests full-trajectory transfer, compaction, and trajectory removal across Claude and GPT model families.", "reason": "An empirical/behavioral study of coding agent handoff dynamics, not a proposed method for a specific task, fits studies.", "source": "crawl"}, {"paper": {"id": "2608.24720", "title": "''You Can't Open an LLM With a Screwdriver'': The De-Democratization of Software", "authors": ["Zixuan Feng", "Italo Santos", "Kostadin Damevski", "Anita Sarma"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24720", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "A vision paper arguing that AI-driven code generation broadens access to producing code but concentrates control (inspection, governance, integration) among fewer experts, shifting software engineering expertise toward intent specification. It identifies research directions for education, tooling, and policy in response.", "reason": "Position/vision paper whose object is the practice and future of AI-assisted software engineering, matching the studies leaf.", "source": "crawl"}, {"paper": {"id": "2601.19066", "title": "Dynamic Cogeneration of Bug Reproduction Test in Agentic Program Repair", "authors": ["Runxiang Cheng", "Michele Tufano", "José Cambronero", "Renyao Wei", "Sherry Shi", "Grant Uy", "Pat Rondon", "Franjo Ivančić"], "venue": "arXiv 2026/01", "category": "", "published": "2026-01-27", "links": {"paper": "https://arxiv.org/abs/2601.19066", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "Studies agentic automated program repair where the agent cogenerates a bug reproduction test alongside the fix in the same patch, evaluated on 120 Google bugs. Develops patch selectors and analyzes failure modes of cogeneration strategies.", "reason": "Bug reproduction test generation combined with fix generation is core issue-resolution/APR work under software_debugging.", "source": "crawl"}, {"paper": {"id": "2602.18537", "title": "CRANE-LLM: Runtime-Augmented LLMs for Crash Prediction and Diagnosis in ML Notebooks", "authors": ["Yiran Wang", "José Antonio Hernández López", "Ulf Nilsson", "Dániel Varró"], "venue": "arXiv 2026/02", "category": "", "published": "2026-02-20", "links": {"paper": "https://arxiv.org/abs/2602.18537", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "CRANE-LLM augments LLMs with runtime information extracted from Jupyter notebook kernels to predict and diagnose crashes in ML notebooks before cell execution, improving accuracy over source-code-only baselines.", "reason": "Matches software_debugging's inclusion of debugging assistants using dynamic analysis for fault diagnosis.", "source": "crawl"}, {"paper": {"id": "2608.20664", "title": "DreamBench-SWE: A Multi-Session Memory-Hygiene Benchmark for Software Agents", "authors": ["Sarthak Singh"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20664", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["benchmark"], "summary": "DreamBench-SWE is a multi-session benchmark testing whether software agents maintain 'memory hygiene' across sessions, scoring software tasks with executable hidden oracles. A preregistered audit finds discriminating power but no clear evidence for external memory mechanisms outperforming baselines.", "reason": "Benchmark of executable, oracle-scored software tasks for coding agents routes to SWE-bench-style software_debugging per benchmark-routing rule.", "source": "crawl"}, {"paper": {"id": "2608.21310", "title": "Beyond Fault Localization: A Trajectory-Level Study of LLM Agents for Microservice Root Cause Analysis", "authors": ["Qisheng Lu", "Aoyang Fang", "Junjielong Xu", "Jin'ao Shang", "Songhan Zhang", "Yifan Yang", "Xiaochuan Yan", "Pinjia He"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.21310", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "Proposes a trajectory-level evaluation framework for LLM agents performing microservice root cause analysis, revealing gaps between answer correctness and diagnostic quality, and introduces DiagGuard to improve localization accuracy.", "reason": "Root-cause analysis and fault localization of reported system failures falls under software_debugging.", "source": "crawl"}, {"paper": {"id": "2608.20851", "title": "BC-Bench: Evaluating Agentic Engineering in a Domain-Specific Language for ERP", "authors": ["Haoran Sun", "Klaus Marius Hansen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-21", "links": {"paper": "https://arxiv.org/abs/2608.20851", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["benchmark"], "summary": "BC-Bench is a SWE-bench-style benchmark of 101 real-world tasks in AL, the DSL for Microsoft Dynamics 365 Business Central ERP systems, evaluating agentic bug fixing and test generation. Frontier models are evaluated across two agent harnesses, showing gains on general benchmarks do not transfer to this domain.", "reason": "Benchmark adapts SWE-bench methodology for bug fixing/issue resolution in an ERP DSL, routing to software_debugging.", "source": "crawl"}, {"paper": {"id": "2608.21718", "title": "XRFix: Exploring Performance Bug Repair of Extended Reality Applications with Large Language Models", "authors": ["Jingwen Wu", "Hanyang Guo", "Hong-Ning Dai", "Xiapu Luo"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.21718", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["benchmark"], "summary": "XRFix is an LLM-based automated program repair framework targeting performance bugs in Extended Reality applications, built on a new codebase and dataset of 104 real-world XR bugs with custom static analysis tools. It fixes bugs at single-line, function, and class levels, outperforming prior APR baselines.", "reason": "Repairing reported performance bugs in code is program repair, matching software_debugging.", "source": "crawl"}, {"paper": {"id": "2608.21867", "title": "MemGuard: Persisting Verifier Signals for LLM-Agent Memory Governance", "authors": ["Haoyu Wang", "Guangyuan Dong", "He Liang", "Zijing Zhang", "Jiachen Luo", "Chuang Liu", "Chao Xue", "Hao Tang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-22", "links": {"paper": "https://arxiv.org/abs/2608.21867", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "MemGuard is a memory-governance framework for LLM agents that persists verifier signals as lifecycle metadata to filter unreliable admissions and repair memory drift across long task streams. Evaluated on Terminal-Bench, SWE-Bench Verified, WebArena, and Mind2Web, it improves success rates and reduces steps over prior memory baselines.", "reason": "A generic agent-memory method evaluated on SWE-bench plus non-code-action benchmarks is routed to artifact per the cross-branch evaluation rule, with issue-resolution (SWE-Bench Verified) as its representative code task.", "source": "crawl"}, {"paper": {"id": "2608.22191", "title": "Disagree to Explore, Agree to Commit: Routing-Guided Test-Time Scaling for Software Agents", "authors": ["Kang Chen", "Junjie Nian", "Yixin Cao", "Yugang Jiang"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-23", "links": {"paper": "https://arxiv.org/abs/2608.22191", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "Risa uses native MoE router traces from software-engineering agents to guide exploration and steer patch-candidate arbitration across sampled trajectories, improving resolved rates on SWE-bench Verified without external judges or test execution. It transfers across model families including gpt-oss and Qwen3.6.", "reason": "Evaluated on SWE-bench Verified issue-resolution task, routing benchmark to software_debugging.", "source": "crawl"}, {"paper": {"id": "2608.23341", "title": "DPIAgent: Divide, Protocol, Isolate for Agentic Reproduction Test Generation", "authors": ["Hao Liu", "Steven Liu", "Xin Zhang", "Jane Luo", "Yu Kang", "Jie Wu", "Fangkai Yang", "Yangyu Huang", "Pengfei Gao", "Scarlett Li", "Yan Lu"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-24", "links": {"paper": "https://arxiv.org/abs/2608.23341", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "DPIAgent splits agentic reproduction test generation into separated diagnosis and test-writing phases with a handoff protocol and isolated toolsets, reducing goal drift. It achieves state-of-the-art results on SWT-Bench Verified across multiple backbone LLMs.", "reason": "Reproduction test generation from bug reports is explicitly software_debugging per the taxonomy's issue-reproduction inclusion.", "source": "crawl"}, {"paper": {"id": "2506.13182", "title": "From Empirical Evaluation to Context-Aware Enhancement: Repairing Regression Errors with LLMs", "authors": ["Anh Ho", "Thanh Le-Cong", "Bach Le", "Christine Rizkallah"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2506.13182", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["benchmark", "empirical"], "summary": "An empirical study evaluating classical and LLM-based automated program repair techniques on a new benchmark of 200 real-world regression bugs (RegressionBug4APR), showing LLM-based APR benefits substantially from bug-inducing change context. Ablations confirm generalizability across Java and Python.", "reason": "Serves automated program repair of reported regression bugs, i.e. issue resolution.", "source": "crawl"}, {"paper": {"id": "2608.24188", "title": "Paritok-4B: Intent-Conditioned Context Compression for Coding Agents", "authors": ["Jiayu Shi", "Luzhuo Chen"], "venue": "arXiv 2026/08", "category": "", "published": "2026-08-25", "links": {"paper": "https://arxiv.org/abs/2608.24188", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["model"], "summary": "Paritok-4B is a 4B extractive, intent-conditioned context compressor distilled from real OpenHands trajectories to shrink coding-agent context while preserving edit-relevant spans. Evaluated on SWE-bench Lite, it compresses context to ~26% of original size with minimal solve-rate loss.", "reason": "Resource paper whose downstream task is issue-resolution agents (SWE-bench), routed by the task it serves.", "source": "crawl"}]