Paper review 2026-07-24: 25 candidates

Reply with commands, one per line:
`/approve all` · `/approve 1,3-5` · `/reject 2` · `/edit 3 category=world_terminal tags=benchmark` (edit implies approve; `tags=-` clears tags). Valid category keys: see taxonomy.json.

_Auto-skipped 229 out-of-scope; 0 failed (retried next run)._

### 1. Is Agentic Code Review Helpful? Mining Developers' Feedback to CodeRabbit Reviews in the Wild
_Hong Yi Lin, Mingzhao Liang, Patanamon Thongtanunam, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.03316)
**proposed: `studies`** · tags: `empirical`
> An empirical study mining over 31,000 CodeRabbit agentic code review comments and developer responses across 10,191 GitHub pull requests, analyzing acceptance/rejection patterns and predictors of rejection. It measures real-world developer reception of an existing agentic code review tool rather than proposing a new reviewing method.
Behavioral/empirical study of real-world usage of an existing code-review agent, not a proposed agent or method, fitting the studies/empirical precedent (SWE-chat).

### 2. Maintenance Signals in AI-Assisted GitHub Repositories: Evidence from GenAI Adopters
_Rikuto Tsuchida, Youmei Fan, Kazumasa Shimari, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.21079)
**proposed: `studies`** · tags: `empirical`
> An empirical study analyzing GitHub repositories and issues from GenAI adopters, finding that AI-assisted repos show distinct documentation patterns and maintenance issues tied to external AI-provider dependencies. The study suggests GenAI shifts developer effort toward validation and maintenance rather than eliminating it.
A behavioral/empirical study of how GenAI adoption affects real-world software maintenance practices, matching the studies leaf's empirical-study definition.

### 3. Beyond Fail-to-Pass: Iterative Hardening of Co-Generated Bug Reproduction Tests and Fixes
_Yuhao Tan, Zhibang Yang, Fangkai Yang, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.19843)
**proposed: `software_debugging`** · tags: none
> CoHarden is a co-generation framework that iteratively hardens bug reproduction tests and patches together, using a 'Lax' mutation-survival signal beyond fail-to-pass to ensure tests actually validate correct fixes for automated program repair.
Targets bug reproduction and repair (APR) evaluated on SWE-bench Verified, the core of software_debugging.

### 4. NVIDIA-labs OO Agents: Native Python Object-Oriented Agents
_Paul Furgale, Severin Klingler, James Nolan, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.20709)
**proposed: `software_debugging`** · tags: none
> NOOA is a model-agnostic Python framework where agents are objects whose methods (bodies of '...') are completed by an LLM agent loop, unifying prompts, tools, and state into standard Python. It is evaluated on agentic benchmarks including SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3.
A generalist agent-programming SDK that performs tasks must route to its dominant evaluated activity rather than studies; SWE-bench Verified is its lead benchmark.

### 5. Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction
_Tencent WorkBuddy Bench Team, Siqi Cai, Shaopeng Chen, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.20911)
**proposed: `software_debugging`** · tags: `benchmark`
> Tencent WorkBuddy Bench is a contamination-resistant, multi-domain coding-agent benchmark spanning Code, Web, Office, and Security tasks, each reverse-engineered from real commits, PRs, or business scenarios and rewritten as role-played requests. It reports a cross-model leaderboard using two agent harnesses under a uniform, reproducible task-directory protocol.
A multi-domain coding-agent evaluation suite led by repository-level engineering tasks, routed like SWE-Compass to the debugging/issue-resolution activity despite broader framing.

### 6. FormulaSPIN: Self-Play Fine-Tuning for Natural Language to Spreadsheet Formula Generation
_Cy Xie_ · arXiv 2026/05 · [paper](https://arxiv.org/abs/2607.19354)
**proposed: `software_code_generation`** · tags: none
> FormulaSPIN is a self-play fine-tuning framework that uses execution feedback to train models for natural-language-to-spreadsheet-formula generation, resolving contradictory gradients from execution-equivalent formulas. It achieves state-of-the-art exact-match and execution accuracy on NL2FORMULA benchmarks.
A training method whose downstream task is generating executable code (formulas) places it in artifact/software_code_generation.

### 7. Bridging Behavior and Implementation: Automated Java Glue Code Generation for Behavior-Driven Development
_Xinyu Shi, Zhou Yang, An Ran Chen_ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.19703)
**proposed: `software_code_generation`** · tags: none
> AutoGlue is a hierarchical multi-agent framework that generates Java glue code linking BDD natural-language scenarios to project implementations, using behavior interpretation and context retrieval agents. It substantially improves API F1 and CodeBLEU over few-shot prompting on eight open-source projects.
A multi-agent system that produces executable code from specifications and project context, serving code generation.

### 8. PennySynth: RAG-Driven Data Synthesis for Automated Quantum Code Generation
_Minghao Shao, Nouhaila Innan, Hariharan Janardhanan, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2605.25572)
**proposed: `software_code_generation`** · tags: `training-data`
> PennySynth is a retrieval-augmented generation framework that conditions LLM code generation on a curated corpus of PennyLane instruction-code pairs to improve quantum circuit code correctness. It boosts pass@5 on QHack coding challenges substantially over non-retrieval baselines.
The task's ultimate purpose is producing correct quantum program code from specifications, matching software_code_generation.

### 9. Learn2Zinc: Fine-tuning Small Language Models for Text-to-Model Translation in MiniZinc
_Serdar Kadioglu, Karthik Uppuluri_ · arXiv 2026/05 · [paper](https://arxiv.org/abs/2607.20456)
**proposed: `software_code_generation`** · tags: `model` `training-data`
> Fine-tunes small language models to translate natural language into MiniZinc constraint models, using cross-model error bootstrapping to fix syntax failures. Achieves up to 98% execution accuracy though constraint-reasoning solution accuracy remains lower at 35%.
Text-to-code generation for a domain-specific language, analogous to text-to-SQL, serving the code-generation task with released models/data.

### 10. ExecuGraph: A Multi-Agent, Execution-Grounded Framework for Reliable Backend Code Synthesis with Large Language Models
_Sai Deekshith Lekkala, Jothi Prabha Appadurai, Rohith Reddy Bellibatlu, et al._ · arXiv 2026/06 · [paper](https://arxiv.org/abs/2607.20499)
**proposed: `software_code_generation`** · tags: none
> ExecuGraph is a six-agent workflow (Planner, Generator, Reviewer, Evaluator, Optimizer, Explainer) that grounds backend code synthesis in sandboxed execution-based validation with bounded retries. Evaluated on HumanEval, APPS, and a custom DSA suite, it isolates the marginal contribution of multi-agent decomposition versus single-agent retry baselines.
The agent's ultimate purpose is producing correct backend code, matching software_code_generation.

### 11. Delivery, Not Storage: Cue-Anchored Working Memory as a Harness Property for Coding Agents
_Swapnanil Saha_ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.20972)
**proposed: `software_code_generation`** · tags: none
> Argues that long-running coding agents need a harness-level 'cue-anchored' working memory tier, distinct from deliberately authored documents, to reliably surface situational facts during coding sessions. A controlled evaluation shows deterministic memory injection outperforms voluntary agent memory use and survives compaction where conversational memory does not.
Proposes a harness/memory design for coding agents performing general coding tasks, not tied to one SWE lifecycle activity, so it defaults to the general code-production leaf.

### 12. ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders
_Zhongyuan Peng, Dan Huang, Chuyu Zhang, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.21217)
**proposed: `software_code_generation`** · tags: `benchmark`
> ICAE-Bench evaluates coding agents that must transform fuzzy product requirements into working software via planning, requirement clarification with a simulated user agent, and repository-level construction. Tasks derive from real open-source repositories with multi-dimensional correctness diagnostics.
The served task is building software from an underspecified intent, matching whole-project/repository code generation.

### 13. AgentCgroup: Understanding and Controlling OS Resources of AI Agents
_Yusheng Zheng, Jiakun Fan, Quanzhi Fu, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2602.09345)
**proposed: `software_infrastructure`** · tags: `empirical`
> AgentCgroup characterizes OS-level resource dynamics (CPU, memory, latency) of sandboxed AI coding agents performing SWE tasks and proposes an eBPF-based, tool-call-aware cgroup controller to improve multi-tenant isolation and reduce resource waste.
Enabling infrastructure work (sandbox/container resource management) that supports running code agents, matching software_infrastructure's environment-setup scope.

### 14. Verifier-First Evaluation of Agentic LLMs for Infrastructure-as-Code Generation
_Mohamed Jouini_ · arXiv 2026/05 · [paper](https://arxiv.org/abs/2607.20478)
**proposed: `software_infrastructure`** · tags: `empirical` `benchmark`
> An empirical, verifier-first study of seven agentic strategies (RAG, iterative refinement, prompt optimization) for generating Terraform Infrastructure-as-Code, evaluated on the IaC-Eval v2 benchmark with staged validate/plan/policy verifiers. Finds active retrieval and verifier-guided refinement substantially raise pass@1 rates.
Agentic generation of infrastructure/configuration code serves the environment-setup and provisioning activity, matching software_infrastructure.

### 15. EditLord: Learning Code Transformation Rules for Code Editing
_Weichen Li, Albert Jan, Baishakhi Ray, et al._ · arXiv 2025/07 · [paper](https://arxiv.org/abs/2504.15284)
**proposed: `software_maintenance`** · tags: none
> EditLord extracts explicit code transformation rules via an LM to guide behavior-preserving code editing, improving performance, robustness, and functional correctness across software engineering and security editing tasks.
Task is behavior-preserving code transformation (property changes without altering functionality), matching software_maintenance.

### 16. PerfAgent: Profiler-Guided Iterative Refinement for Repository-Level Code Optimization
_Ryan Deng, Yuanzhe Liu, Bastian Lipka, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.19653)
**proposed: `software_maintenance`** · tags: none
> PerfAgent is a profiler-guided, verifier-in-the-loop workflow that augments coding agents to perform repository-level performance optimization, using profiler evidence to find real bottlenecks while preserving behavior. It roughly doubles the rate of expert-matching optimization patches on GSO and SWE-fficiency-Lite benchmarks.
Task is behavior-preserving runtime performance optimization of existing repositories, matching the software_maintenance leaf.

### 17. Multi-Source and Cross-Scenario Strategy-Guided Code Optimization
_Yuwei Zhao, Qianyu Xiao, Ye Cui, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.20353)
**proposed: `software_maintenance`** · tags: none
> MoST is an LLM-based code optimization framework that unifies strategies from multiple knowledge sources and transfers them across programming-language scenarios to generate performance-optimization patches. It outperforms prior strategy-guided baselines on C/C++, Python, and Rust optimization benchmarks and real-world projects.
The task is performance optimization of existing code while preserving behavior, matching software_maintenance.

### 18. Learning to Detect UI Principle Violations via Reinforcement Learning
_Nishi Mehta, Swathi Alse, Himani Kumavat, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.20690)
**proposed: `software_review`** · tags: `model` `training-data`
> This work trains a lightweight vision-language critic via reinforcement learning to detect UI principle violations (accessibility, dark patterns, poor hierarchy) in LLM-generated web front-end code, using a synthetic dataset of ~10,000 injected-violation web pages. The critic can audit generated interfaces and provide reward signals for design-aware code generation.
The task judges the quality of generated code artifacts, matching the code-review lifecycle activity rather than code production itself.

### 19. An Agentic Approach Towards Replication Package Quality Evaluation
_Maximilian Alexander Amougou Mbida, Florian Angermeir_ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2606.02006)
**proposed: `software_review`** · tags: none
> Proposes a multi-agent prototype that operationalizes open-science reproducibility guidelines into machine-verifiable criteria to automatically inspect and score replication package quality, producing evidence-grounded improvement reports. Preliminary evaluation shows high consistency and researcher-perceived usefulness.
The agent judges the quality of research code/artifacts and drafts improvement reports, analogous to automated review and change-quality judgment.

### 20. zkCraft: Prompt-Guided LLM as a Zero-Shot Mutation Pattern Oracle for TCCT-Powered ZK Fuzzing
_Rong Fu, Jia Yee Tan, Ziyu Kong, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2602.00667)
**proposed: `software_security`** · tags: none
> zkCraft combines LLM-guided mutation templates with proof-bearing algebraic verification to detect under- and over-constrained faults in zero-knowledge circuits. It reduces solver interaction while achieving low false positives on real Circom code.
Detects and localizes security-critical constraint bugs in ZK circuits, matching software_security's vulnerability detection/repair scope.

### 21. Rewriting the Response Path: Silent Tampering and Provider-Signed Defense in BYOK LLM Agents
_Mingyu Luo, Zihan Zhang, Zesen Liu, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2605.02187)
**proposed: `software_security`** · tags: none
> This paper identifies a response-path integrity gap in Bring-Your-Own-Key LLM agent setups where a relay can silently tamper with execution-bearing fields of responses, causing malicious code changes to pass tests while defeating security checks; it proposes a provider-signed defense (sign-c) that rejects tampered responses with negligible overhead.
Centers on protecting the integrity of agent-produced code changes against tampering that introduces vulnerabilities, matching the security-assurance activity on code.

### 22. Sifting the Noise: A Comparative Study of LLM Agents in Vulnerability False Positive Filtering
_Yunpeng Xiong, Ting Zhang_ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2601.22952)
**proposed: `software_security`** · tags: `empirical`
> A comparative study evaluating three LLM-based agent frameworks (Aider, OpenHands, SWE-agent) for filtering false-positive static-analysis vulnerability alerts across OWASP, real-world Java, and post-cutoff C/C++ datasets.
An empirical study focused on a single security activity (vulnerability FP filtering) routes to software_security plus empirical tag, not to the field-wide studies leaf.

### 23. Buzz to Boom: Detecting Message Progression Vulnerabilities in Electron Applications via Segmented Directed Fuzzing
_Jianjia Yu, Zhengyu Liu, Ziyang Li, et al._ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.20698)
**proposed: `software_security`** · tags: none
> Proton is a segmented directed fuzzing framework that decomposes cross-process message-passing in Electron apps to detect Message Progression Vulnerabilities, synthesizing end-to-end exploit inputs. Evaluated on 589 real-world apps, it found 23 zero-day vulnerabilities leading to CVEs and bug bounties.
Vulnerability-targeted fuzzing to find and validate security flaws in code matches the software_security leaf.

### 24. IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests
_Ankur Singh, Jinqiu Yang, Tse-Hsun Chen_ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2607.20759)
**proposed: `software_security`** · tags: `benchmark`
> IssueTrojanBench benchmarks state-of-the-art coding agents (Cursor, Claude Code, Codex Desktop) against malicious issue requests using four attack categories and six delivery vectors, finding 66.5% penetrate all guardrails. It shows current agent- and LLM-level defenses are insufficient against prompt-injected malicious instructions embedded in development artifacts.
Evaluates vulnerability/attack surface of code-writing agents themselves, fitting software_security's auditing and vulnerability-benchmark scope.

### 25. AI Security Policy Should Assess Systems, Not Only Models
_Michael A. Riegler, Inga Strümke_ · arXiv 2026/07 · [paper](https://arxiv.org/abs/2605.09504)
**proposed: `software_security`** · tags: none
> Presents swarm-attack, a multi-LLM-agent adversarial testing framework combining jailbreak attacks and automated vulnerability discovery via source-code analysis and binary fuzzing. It recovers all 9 planted CWEs in a vulnerable C application within minutes, arguing security policy must assess whole systems.
Fuzzing and source-code analysis aimed at discovering CVEs/CWEs is vulnerability detection, per the software_testing/software_security boundary.

<details><summary>machine payload (do not edit)</summary>

```json
[{"paper": {"id": "2607.03316", "title": "Is Agentic Code Review Helpful? Mining Developers' Feedback to CodeRabbit Reviews in the Wild", "authors": ["Hong Yi Lin", "Mingzhao Liang", "Patanamon Thongtanunam", "Kla Tantithamthavorn"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.03316", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical study mining over 31,000 CodeRabbit agentic code review comments and developer responses across 10,191 GitHub pull requests, analyzing acceptance/rejection patterns and predictors of rejection. It measures real-world developer reception of an existing agentic code review tool rather than proposing a new reviewing method.", "reason": "Behavioral/empirical study of real-world usage of an existing code-review agent, not a proposed agent or method, fitting the studies/empirical precedent (SWE-chat).", "source": "crawl"}, {"paper": {"id": "2607.21079", "title": "Maintenance Signals in AI-Assisted GitHub Repositories: Evidence from GenAI Adopters", "authors": ["Rikuto Tsuchida", "Youmei Fan", "Kazumasa Shimari", "Raula Gaikovina Kula", "Gema Rodríguez-Pérez", "Kenichi Matsumoto"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.21079", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical study analyzing GitHub repositories and issues from GenAI adopters, finding that AI-assisted repos show distinct documentation patterns and maintenance issues tied to external AI-provider dependencies. The study suggests GenAI shifts developer effort toward validation and maintenance rather than eliminating it.", "reason": "A behavioral/empirical study of how GenAI adoption affects real-world software maintenance practices, matching the studies leaf's empirical-study definition.", "source": "crawl"}, {"paper": {"id": "2607.19843", "title": "Beyond Fail-to-Pass: Iterative Hardening of Co-Generated Bug Reproduction Tests and Fixes", "authors": ["Yuhao Tan", "Zhibang Yang", "Fangkai Yang", "Yuan Yao", "Yu Kang", "Lu Wang", "Pu Zhao", "Xin Zhang", "Xiaoxing Ma", "Qingwei Lin", "Saravan Rajmohan", "Dongmei Zhang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.19843", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "CoHarden is a co-generation framework that iteratively hardens bug reproduction tests and patches together, using a 'Lax' mutation-survival signal beyond fail-to-pass to ensure tests actually validate correct fixes for automated program repair.", "reason": "Targets bug reproduction and repair (APR) evaluated on SWE-bench Verified, the core of software_debugging.", "source": "crawl"}, {"paper": {"id": "2607.20709", "title": "NVIDIA-labs OO Agents: Native Python Object-Oriented Agents", "authors": ["Paul Furgale", "Severin Klingler", "James Nolan", "Matt Staats", "Gaia Di Lorenzo", "Elisa Martinez Abad", "Christian Schüller", "Razvan Dinu", "Alessio Devoto", "Pascal Berard", "Gal Kaplun", "Elad Sarafian", "Riccardo Roveri", "Leon Derczynski", "Ricardo Silveira Cabral"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20709", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "NOOA is a model-agnostic Python framework where agents are objects whose methods (bodies of '...') are completed by an LLM agent loop, unifying prompts, tools, and state into standard Python. It is evaluated on agentic benchmarks including SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3.", "reason": "A generalist agent-programming SDK that performs tasks must route to its dominant evaluated activity rather than studies; SWE-bench Verified is its lead benchmark.", "source": "crawl"}, {"paper": {"id": "2607.20911", "title": "Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction", "authors": ["Tencent WorkBuddy Bench Team", "Siqi Cai", "Shaopeng Chen", "Xiang Fei", "Yong Mao", "Zihan Xu", "Zhiheng Lyu", "Zhijian Shao", "Yuchen Shi", "Shuwen Zhang", "Chaofan Qiu", "Linjie Che", "Xiaoxi Zhao", "Feng Wu", "Kai Zhang", "Chaofan Zhu", "Yubin Qi", "Xiaoyun Liang", "Peijie Dong", "Yunhao Zhang", "Yuanjie Zhu", "Ling Jiang", "Xianjun Zhang", "Zhehang Chu", "Anyuan Sang", "Zhen Feng", "Sen Nie", "Shi Wu", "Yuanzhen Xu", "Xin Li", "Ning Yang", "Zhiqiang Dong", "Hande Dong", "Qiang Lin", "Yi Liu", "Yunsheng Wu", "Ke Li", "Xing Sun"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.20911", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["benchmark"], "summary": "Tencent WorkBuddy Bench is a contamination-resistant, multi-domain coding-agent benchmark spanning Code, Web, Office, and Security tasks, each reverse-engineered from real commits, PRs, or business scenarios and rewritten as role-played requests. It reports a cross-model leaderboard using two agent harnesses under a uniform, reproducible task-directory protocol.", "reason": "A multi-domain coding-agent evaluation suite led by repository-level engineering tasks, routed like SWE-Compass to the debugging/issue-resolution activity despite broader framing.", "source": "crawl"}, {"paper": {"id": "2607.19354", "title": "FormulaSPIN: Self-Play Fine-Tuning for Natural Language to Spreadsheet Formula Generation", "authors": ["Cy Xie"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-21", "links": {"paper": "https://arxiv.org/abs/2607.19354", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "FormulaSPIN is a self-play fine-tuning framework that uses execution feedback to train models for natural-language-to-spreadsheet-formula generation, resolving contradictory gradients from execution-equivalent formulas. It achieves state-of-the-art exact-match and execution accuracy on NL2FORMULA benchmarks.", "reason": "A training method whose downstream task is generating executable code (formulas) places it in artifact/software_code_generation.", "source": "crawl"}, {"paper": {"id": "2607.19703", "title": "Bridging Behavior and Implementation: Automated Java Glue Code Generation for Behavior-Driven Development", "authors": ["Xinyu Shi", "Zhou Yang", "An Ran Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.19703", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "AutoGlue is a hierarchical multi-agent framework that generates Java glue code linking BDD natural-language scenarios to project implementations, using behavior interpretation and context retrieval agents. It substantially improves API F1 and CodeBLEU over few-shot prompting on eight open-source projects.", "reason": "A multi-agent system that produces executable code from specifications and project context, serving code generation.", "source": "crawl"}, {"paper": {"id": "2605.25572", "title": "PennySynth: RAG-Driven Data Synthesis for Automated Quantum Code Generation", "authors": ["Minghao Shao", "Nouhaila Innan", "Hariharan Janardhanan", "Muhammad Kashif", "Alberto Marchisio", "Muhammad Shafique"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2605.25572", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["training-data"], "summary": "PennySynth is a retrieval-augmented generation framework that conditions LLM code generation on a curated corpus of PennyLane instruction-code pairs to improve quantum circuit code correctness. It boosts pass@5 on QHack coding challenges substantially over non-retrieval baselines.", "reason": "The task's ultimate purpose is producing correct quantum program code from specifications, matching software_code_generation.", "source": "crawl"}, {"paper": {"id": "2607.20456", "title": "Learn2Zinc: Fine-tuning Small Language Models for Text-to-Model Translation in MiniZinc", "authors": ["Serdar Kadioglu", "Karthik Uppuluri"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-14", "links": {"paper": "https://arxiv.org/abs/2607.20456", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["model", "training-data"], "summary": "Fine-tunes small language models to translate natural language into MiniZinc constraint models, using cross-model error bootstrapping to fix syntax failures. Achieves up to 98% execution accuracy though constraint-reasoning solution accuracy remains lower at 35%.", "reason": "Text-to-code generation for a domain-specific language, analogous to text-to-SQL, serving the code-generation task with released models/data.", "source": "crawl"}, {"paper": {"id": "2607.20499", "title": "ExecuGraph: A Multi-Agent, Execution-Grounded Framework for Reliable Backend Code Synthesis with Large Language Models", "authors": ["Sai Deekshith Lekkala", "Jothi Prabha Appadurai", "Rohith Reddy Bellibatlu", "Manpreet Singh"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-18", "links": {"paper": "https://arxiv.org/abs/2607.20499", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "ExecuGraph is a six-agent workflow (Planner, Generator, Reviewer, Evaluator, Optimizer, Explainer) that grounds backend code synthesis in sandboxed execution-based validation with bounded retries. Evaluated on HumanEval, APPS, and a custom DSA suite, it isolates the marginal contribution of multi-agent decomposition versus single-agent retry baselines.", "reason": "The agent's ultimate purpose is producing correct backend code, matching software_code_generation.", "source": "crawl"}, {"paper": {"id": "2607.20972", "title": "Delivery, Not Storage: Cue-Anchored Working Memory as a Harness Property for Coding Agents", "authors": ["Swapnanil Saha"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.20972", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "Argues that long-running coding agents need a harness-level 'cue-anchored' working memory tier, distinct from deliberately authored documents, to reliably surface situational facts during coding sessions. A controlled evaluation shows deterministic memory injection outperforms voluntary agent memory use and survives compaction where conversational memory does not.", "reason": "Proposes a harness/memory design for coding agents performing general coding tasks, not tied to one SWE lifecycle activity, so it defaults to the general code-production leaf.", "source": "crawl"}, {"paper": {"id": "2607.21217", "title": "ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders", "authors": ["Zhongyuan Peng", "Dan Huang", "Chuyu Zhang", "Caijun Xu", "Changyi Xiao", "Shibo Hong", "David Lo", "Lin Qiu", "Xuezhi Cao", "Jiyuan He", "Yixin Cao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2607.21217", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark"], "summary": "ICAE-Bench evaluates coding agents that must transform fuzzy product requirements into working software via planning, requirement clarification with a simulated user agent, and repository-level construction. Tasks derive from real open-source repositories with multi-dimensional correctness diagnostics.", "reason": "The served task is building software from an underspecified intent, matching whole-project/repository code generation.", "source": "crawl"}, {"paper": {"id": "2602.09345", "title": "AgentCgroup: Understanding and Controlling OS Resources of AI Agents", "authors": ["Yusheng Zheng", "Jiakun Fan", "Quanzhi Fu", "Yiwei Yang", "Wei Zhang", "Andi Quinn"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2602.09345", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["empirical"], "summary": "AgentCgroup characterizes OS-level resource dynamics (CPU, memory, latency) of sandboxed AI coding agents performing SWE tasks and proposes an eBPF-based, tool-call-aware cgroup controller to improve multi-tenant isolation and reduce resource waste.", "reason": "Enabling infrastructure work (sandbox/container resource management) that supports running code agents, matching software_infrastructure's environment-setup scope.", "source": "crawl"}, {"paper": {"id": "2607.20478", "title": "Verifier-First Evaluation of Agentic LLMs for Infrastructure-as-Code Generation", "authors": ["Mohamed Jouini"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-29", "links": {"paper": "https://arxiv.org/abs/2607.20478", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["empirical", "benchmark"], "summary": "An empirical, verifier-first study of seven agentic strategies (RAG, iterative refinement, prompt optimization) for generating Terraform Infrastructure-as-Code, evaluated on the IaC-Eval v2 benchmark with staged validate/plan/policy verifiers. Finds active retrieval and verifier-guided refinement substantially raise pass@1 rates.", "reason": "Agentic generation of infrastructure/configuration code serves the environment-setup and provisioning activity, matching software_infrastructure.", "source": "crawl"}, {"paper": {"id": "2504.15284", "title": "EditLord: Learning Code Transformation Rules for Code Editing", "authors": ["Weichen Li", "Albert Jan", "Baishakhi Ray", "Junfeng Yang", "Chengzhi Mao", "Kexin Pei"], "venue": "arXiv 2025/07", "category": "", "published": "2025-07-09", "links": {"paper": "https://arxiv.org/abs/2504.15284", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "EditLord extracts explicit code transformation rules via an LM to guide behavior-preserving code editing, improving performance, robustness, and functional correctness across software engineering and security editing tasks.", "reason": "Task is behavior-preserving code transformation (property changes without altering functionality), matching software_maintenance.", "source": "crawl"}, {"paper": {"id": "2607.19653", "title": "PerfAgent: Profiler-Guided Iterative Refinement for Repository-Level Code Optimization", "authors": ["Ryan Deng", "Yuanzhe Liu", "Bastian Lipka", "Yao Ma", "Xuhao Chen", "Tim Kaler", "Jatin Ganhotra"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.19653", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "PerfAgent is a profiler-guided, verifier-in-the-loop workflow that augments coding agents to perform repository-level performance optimization, using profiler evidence to find real bottlenecks while preserving behavior. It roughly doubles the rate of expert-matching optimization patches on GSO and SWE-fficiency-Lite benchmarks.", "reason": "Task is behavior-preserving runtime performance optimization of existing repositories, matching the software_maintenance leaf.", "source": "crawl"}, {"paper": {"id": "2607.20353", "title": "Multi-Source and Cross-Scenario Strategy-Guided Code Optimization", "authors": ["Yuwei Zhao", "Qianyu Xiao", "Ye Cui", "Yijun Yu", "Yingfei Xiong"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20353", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "MoST is an LLM-based code optimization framework that unifies strategies from multiple knowledge sources and transfers them across programming-language scenarios to generate performance-optimization patches. It outperforms prior strategy-guided baselines on C/C++, Python, and Rust optimization benchmarks and real-world projects.", "reason": "The task is performance optimization of existing code while preserving behavior, matching software_maintenance.", "source": "crawl"}, {"paper": {"id": "2607.20690", "title": "Learning to Detect UI Principle Violations via Reinforcement Learning", "authors": ["Nishi Mehta", "Swathi Alse", "Himani Kumavat", "Yue Yu", "Pratik Jayarao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20690", "github": "", "website": ""}}, "category": "software_review", "tags": ["model", "training-data"], "summary": "This work trains a lightweight vision-language critic via reinforcement learning to detect UI principle violations (accessibility, dark patterns, poor hierarchy) in LLM-generated web front-end code, using a synthetic dataset of ~10,000 injected-violation web pages. The critic can audit generated interfaces and provide reward signals for design-aware code generation.", "reason": "The task judges the quality of generated code artifacts, matching the code-review lifecycle activity rather than code production itself.", "source": "crawl"}, {"paper": {"id": "2606.02006", "title": "An Agentic Approach Towards Replication Package Quality Evaluation", "authors": ["Maximilian Alexander Amougou Mbida", "Florian Angermeir"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2606.02006", "github": "", "website": ""}}, "category": "software_review", "tags": [], "summary": "Proposes a multi-agent prototype that operationalizes open-science reproducibility guidelines into machine-verifiable criteria to automatically inspect and score replication package quality, producing evidence-grounded improvement reports. Preliminary evaluation shows high consistency and researcher-perceived usefulness.", "reason": "The agent judges the quality of research code/artifacts and drafts improvement reports, analogous to automated review and change-quality judgment.", "source": "crawl"}, {"paper": {"id": "2602.00667", "title": "zkCraft: Prompt-Guided LLM as a Zero-Shot Mutation Pattern Oracle for TCCT-Powered ZK Fuzzing", "authors": ["Rong Fu", "Jia Yee Tan", "Ziyu Kong", "Shuning Zhang", "Xianda Li", "Kun Liu", "Youjin Wang", "Simon Fong"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2602.00667", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "zkCraft combines LLM-guided mutation templates with proof-bearing algebraic verification to detect under- and over-constrained faults in zero-knowledge circuits. It reduces solver interaction while achieving low false positives on real Circom code.", "reason": "Detects and localizes security-critical constraint bugs in ZK circuits, matching software_security's vulnerability detection/repair scope.", "source": "crawl"}, {"paper": {"id": "2605.02187", "title": "Rewriting the Response Path: Silent Tampering and Provider-Signed Defense in BYOK LLM Agents", "authors": ["Mingyu Luo", "Zihan Zhang", "Zesen Liu", "Yuchong Xie", "Zhixiang Zhang", "Dung Hiu Hilton Yeung", "Wai Ip Lai", "Ping Chen", "Ming Wen", "Dongdong She"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2605.02187", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "This paper identifies a response-path integrity gap in Bring-Your-Own-Key LLM agent setups where a relay can silently tamper with execution-bearing fields of responses, causing malicious code changes to pass tests while defeating security checks; it proposes a provider-signed defense (sign-c) that rejects tampered responses with negligible overhead.", "reason": "Centers on protecting the integrity of agent-produced code changes against tampering that introduces vulnerabilities, matching the security-assurance activity on code.", "source": "crawl"}, {"paper": {"id": "2601.22952", "title": "Sifting the Noise: A Comparative Study of LLM Agents in Vulnerability False Positive Filtering", "authors": ["Yunpeng Xiong", "Ting Zhang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2601.22952", "github": "", "website": ""}}, "category": "software_security", "tags": ["empirical"], "summary": "A comparative study evaluating three LLM-based agent frameworks (Aider, OpenHands, SWE-agent) for filtering false-positive static-analysis vulnerability alerts across OWASP, real-world Java, and post-cutoff C/C++ datasets.", "reason": "An empirical study focused on a single security activity (vulnerability FP filtering) routes to software_security plus empirical tag, not to the field-wide studies leaf.", "source": "crawl"}, {"paper": {"id": "2607.20698", "title": "Buzz to Boom: Detecting Message Progression Vulnerabilities in Electron Applications via Segmented Directed Fuzzing", "authors": ["Jianjia Yu", "Zhengyu Liu", "Ziyang Li", "Yu Sun", "Yinzhi Cao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20698", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "Proton is a segmented directed fuzzing framework that decomposes cross-process message-passing in Electron apps to detect Message Progression Vulnerabilities, synthesizing end-to-end exploit inputs. Evaluated on 589 real-world apps, it found 23 zero-day vulnerabilities leading to CVEs and bug bounties.", "reason": "Vulnerability-targeted fuzzing to find and validate security flaws in code matches the software_security leaf.", "source": "crawl"}, {"paper": {"id": "2607.20759", "title": "IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests", "authors": ["Ankur Singh", "Jinqiu Yang", "Tse-Hsun Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-22", "links": {"paper": "https://arxiv.org/abs/2607.20759", "github": "", "website": ""}}, "category": "software_security", "tags": ["benchmark"], "summary": "IssueTrojanBench benchmarks state-of-the-art coding agents (Cursor, Claude Code, Codex Desktop) against malicious issue requests using four attack categories and six delivery vectors, finding 66.5% penetrate all guardrails. It shows current agent- and LLM-level defenses are insufficient against prompt-injected malicious instructions embedded in development artifacts.", "reason": "Evaluates vulnerability/attack surface of code-writing agents themselves, fitting software_security's auditing and vulnerability-benchmark scope.", "source": "crawl"}, {"paper": {"id": "2605.09504", "title": "AI Security Policy Should Assess Systems, Not Only Models", "authors": ["Michael A. Riegler", "Inga Strümke"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-23", "links": {"paper": "https://arxiv.org/abs/2605.09504", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "Presents swarm-attack, a multi-LLM-agent adversarial testing framework combining jailbreak attacks and automated vulnerability discovery via source-code analysis and binary fuzzing. It recovers all 9 planted CWEs in a vulnerable C application within minutes, arguing security policy must assess whole systems.", "reason": "Fuzzing and source-code analysis aimed at discovering CVEs/CWEs is vulnerability detection, per the software_testing/software_security boundary.", "source": "crawl"}]
```

</details>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Paper review 2026-07-24: 25 candidates #337

1. Is Agentic Code Review Helpful? Mining Developers' Feedback to CodeRabbit Reviews in the Wild

2. Maintenance Signals in AI-Assisted GitHub Repositories: Evidence from GenAI Adopters

3. Beyond Fail-to-Pass: Iterative Hardening of Co-Generated Bug Reproduction Tests and Fixes

4. NVIDIA-labs OO Agents: Native Python Object-Oriented Agents

5. Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction

6. FormulaSPIN: Self-Play Fine-Tuning for Natural Language to Spreadsheet Formula Generation

7. Bridging Behavior and Implementation: Automated Java Glue Code Generation for Behavior-Driven Development

8. PennySynth: RAG-Driven Data Synthesis for Automated Quantum Code Generation

9. Learn2Zinc: Fine-tuning Small Language Models for Text-to-Model Translation in MiniZinc

10. ExecuGraph: A Multi-Agent, Execution-Grounded Framework for Reliable Backend Code Synthesis with Large Language Models

11. Delivery, Not Storage: Cue-Anchored Working Memory as a Harness Property for Coding Agents

12. ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders

13. AgentCgroup: Understanding and Controlling OS Resources of AI Agents

14. Verifier-First Evaluation of Agentic LLMs for Infrastructure-as-Code Generation

15. EditLord: Learning Code Transformation Rules for Code Editing

16. PerfAgent: Profiler-Guided Iterative Refinement for Repository-Level Code Optimization

17. Multi-Source and Cross-Scenario Strategy-Guided Code Optimization

18. Learning to Detect UI Principle Violations via Reinforcement Learning

19. An Agentic Approach Towards Replication Package Quality Evaluation

20. zkCraft: Prompt-Guided LLM as a Zero-Shot Mutation Pattern Oracle for TCCT-Powered ZK Fuzzing

21. Rewriting the Response Path: Silent Tampering and Provider-Signed Defense in BYOK LLM Agents

22. Sifting the Noise: A Comparative Study of LLM Agents in Vulnerability False Positive Filtering

23. Buzz to Boom: Detecting Message Progression Vulnerabilities in Electron Applications via Segmented Directed Fuzzing

24. IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests

25. AI Security Policy Should Assess Systems, Not Only Models

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Uh oh!

Paper review 2026-07-24: 25 candidates #337

Description

1. Is Agentic Code Review Helpful? Mining Developers' Feedback to CodeRabbit Reviews in the Wild

2. Maintenance Signals in AI-Assisted GitHub Repositories: Evidence from GenAI Adopters

3. Beyond Fail-to-Pass: Iterative Hardening of Co-Generated Bug Reproduction Tests and Fixes

4. NVIDIA-labs OO Agents: Native Python Object-Oriented Agents

5. Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction

6. FormulaSPIN: Self-Play Fine-Tuning for Natural Language to Spreadsheet Formula Generation

7. Bridging Behavior and Implementation: Automated Java Glue Code Generation for Behavior-Driven Development

8. PennySynth: RAG-Driven Data Synthesis for Automated Quantum Code Generation

9. Learn2Zinc: Fine-tuning Small Language Models for Text-to-Model Translation in MiniZinc

10. ExecuGraph: A Multi-Agent, Execution-Grounded Framework for Reliable Backend Code Synthesis with Large Language Models

11. Delivery, Not Storage: Cue-Anchored Working Memory as a Harness Property for Coding Agents

12. ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders

13. AgentCgroup: Understanding and Controlling OS Resources of AI Agents

14. Verifier-First Evaluation of Agentic LLMs for Infrastructure-as-Code Generation

15. EditLord: Learning Code Transformation Rules for Code Editing

16. PerfAgent: Profiler-Guided Iterative Refinement for Repository-Level Code Optimization

17. Multi-Source and Cross-Scenario Strategy-Guided Code Optimization

18. Learning to Detect UI Principle Violations via Reinforcement Learning

19. An Agentic Approach Towards Replication Package Quality Evaluation

20. zkCraft: Prompt-Guided LLM as a Zero-Shot Mutation Pattern Oracle for TCCT-Powered ZK Fuzzing

21. Rewriting the Response Path: Silent Tampering and Provider-Signed Defense in BYOK LLM Agents

22. Sifting the Noise: A Comparative Study of LLM Agents in Vulnerability False Positive Filtering

23. Buzz to Boom: Detecting Message Progression Vulnerabilities in Electron Applications via Segmented Directed Fuzzing

24. IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests

25. AI Security Policy Should Assess Systems, Not Only Models

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions