[{"paper": {"id": "2607.03691", "title": "Don't Blame the Large Language Model: How Agent Harness Evolution Shapes Coding Agent Quality", "authors": ["Oussama Ben Sghaier", "Hao Li", "Bram Adams", "Ahmed E. Hassan"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.03691", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A controlled longitudinal study isolating how evolving agent harnesses (Codex, Qwen Code, Gemini, OpenCode, OpenHands) affect coding agent quality, holding the LLM constant across 35 sequential Qwen Code releases evaluated on SWE-bench Verified.", "reason": "The paper's object is the coding agents/harnesses themselves rather than proposing a task-performing agent, matching the studies/empirical definition.", "source": "crawl"}, {"paper": {"id": "2607.18462", "title": "Beyond Resolved Rate: A Non-Functional Quality Study", "authors": ["Xin Sun", "Daniel Ståhl", "Kristian Sandahl", "Christoph Kessler"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18462", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "This study empirically compares non-functional quality (static-analysis warnings, CPU time, memory) of patches generated by successive Claude and DeepSeek models via SWE-agent on SWE-bench Lite. It finds resolved rates improve across generations but non-functional patch quality does not consistently improve.", "reason": "An empirical study whose object is analyzing coding-agent output quality trends across model generations, not a new agent or tool.", "source": "crawl"}, {"paper": {"id": "2607.18970", "title": "Skillware: A Software Ontology and Engineering Lifecycle for Persistent Behavioral Artifacts", "authors": ["Haodi Fan", "Zucong Lan"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18970", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "Skillware proposes a software ontology treating 'Agent Skills' as persistent behavioral artifacts with independent identity and lifecycle, grounded in analysis of over 138,000 SKILL.md records and category-boundary case studies.", "reason": "Object of study is the AI-agent-skill ecosystem itself (ontology/taxonomy), not a proposed agent performing a task, matching the studies/position leaf.", "source": "crawl"}, {"paper": {"id": "2604.21744", "title": "Agentic AI-assisted coding offers a unique opportunity to instill epistemic grounding during software development", "authors": ["Magnus Palmblad", "Jared M. Ragland", "Benjamin A. Neely"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2604.21744", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "Proposes GROUNDING.md, a community-governed, field-scoped document encoding hard constraints and conventions to keep domain expertise embedded in agentic AI-assisted software development, illustrated with a proteomics example. It is a position piece advocating a governance practice for agentic coding rather than a working agent or benchmark.", "reason": "Position/vision paper about the practice of agentic software development, not a proposed agent or task-performing method, so it fits the studies off-axis branch.", "source": "crawl"}, {"paper": {"id": "2607.19336", "title": "Agents in the Wild: Where Research Meets Deployment", "authors": ["Grace Hui Yang", "Pranav N. Venkit", "Hooman Sedghamiz", "Enrico Santus", "Victor Dibia", "Ioana Baldini"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19336", "github": "", "website": ""}}, "category": "studies", "tags": ["survey"], "summary": "A tutorial surveying agentic LLM systems moving from research to production deployment across software engineering, scientific discovery, and finance, distilling design patterns, evaluation checklists, and mitigation strategies for reliability and safety.", "reason": "Field-wide tutorial whose object is agentic systems and deployment practice broadly, not a task-specific agent proposal, matching the studies leaf.", "source": "crawl"}, {"paper": {"id": "2607.18859", "title": "PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents", "authors": ["Tianyue Jiang", "Yanlin Wang", "Xin He", "Daya Guo", "Jiachi Chen", "Ming Wen", "Ensheng Shi", "Xilin Liu", "Yuchi Ma", "Guanbin Li"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18859", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "PhoenixRepair is a multi-agent issue-resolution framework that expands repair-strategy exploration via multi-location sampling and iterative reflection/refinement on patches. It achieves state-of-the-art resolved rates on SWE-bench-Verified across multiple backbone LLMs.", "reason": "An agent framework for reproducing, localizing, and repairing reported software issues is squarely software_debugging (SWE-bench-style issue resolution).", "source": "crawl"}, {"paper": {"id": "2506.19045", "title": "Efficient Black-Box Fault Localization for System-Level Test Code Using Large Language Models", "authors": ["Ahmadreza Saboor Yaraghi", "Golnaz Gharachorlu", "Sakina Fatima", "Lionel C. Briand", "Ruiyuan Wan", "Ruifeng Gao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2506.19045", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "A static, LLM-driven fault-localization method for system-level test code that estimates execution traces from a single failure log and prompts an LLM to rank likely faulty locations without executing the test.", "reason": "Fault localization and root-cause analysis of code is explicitly included under software_debugging.", "source": "crawl"}, {"paper": {"id": "2607.18260", "title": "FindStatBench: Evaluating Large Language Models on Combinatorial Code Synthesis", "authors": ["Soham Dan"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-16", "links": {"paper": "https://arxiv.org/abs/2607.18260", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark"], "summary": "FindStatBench is an execution benchmark of 2,329 combinatorial code synthesis tasks requiring LLMs to emit a single Python solve function with no tool use or feedback, scored by sandboxed execution on hidden instances. It finds statistic synthesis easier than map synthesis and reveals brittleness in symbolic rule induction.", "reason": "A benchmark for producing code from specification (competitive-programming style), matching software_code_generation.", "source": "crawl"}, {"paper": {"id": "2607.18357", "title": "Decode-Time Grammars: Constrained LLM Generation over a Refinement Order of Grammar Fragments", "authors": ["Shuoming Zhang", "Ruiyuan Xu", "Haofeng Li", "Qiuchu Yu", "Yangyu Zhang", "Chunwei Xia", "Xiaobing Feng", "Chenxi Wang", "Huimin Cui", "Jiacheng Zhao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18357", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "Proposes decode-time grammars that instantiate environment-typed grammar fragments during LLM decoding to guarantee semantically valid code for low-resource DSLs, APIs, and CLIs, tested on TileLang, SQL, and P4.", "reason": "The technique's purpose is producing semantically correct code, serving the code generation task despite being a decoding method.", "source": "crawl"}, {"paper": {"id": "2607.18742", "title": "Beyond Text Editing: Algebraic Manipulation of Source Code", "authors": ["Kevin Pulo"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18742", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "Proposes source code algebra, a set of composable algebraic operations for modifying codebases as an alternative to text-level edits, aimed at reducing friction for LLM coding agents making non-local changes. A proof-of-concept (SCAS) shows higher success rates and far fewer tokens on cross-file code changes versus text-editing baselines.", "reason": "An agentic method for producing/completing cross-file code edits serves code generation and editing, the task's ultimate purpose is producing code.", "source": "crawl"}, {"paper": {"id": "2607.18886", "title": "TraceDev: A Traceability-Driven Multi-agent Framework for Requirement-to-Code Development", "authors": ["Mingyu Chen", "Yakun Zhang", "Zihao Xie", "Yixing Luo", "Jinrui Xu", "Cuiyun Gao", "Kaiqi Zhao", "Yunming Ye"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18886", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "TraceDev is a multi-agent framework with five role-specific agents that transforms complex natural-language requirements (use cases) into repository-level code, using a traceability graph linking requirements, design, and code artifacts. It outperforms baselines on ETOUR and SMOS datasets.", "reason": "Multi-agent system whose ultimate purpose is producing repository-level code from requirements, matching software_code_generation's whole-repo/app generation.", "source": "crawl"}, {"paper": {"id": "2607.19104", "title": "SciCodePile: A 128GB Corpus and Executable Benchmark for Challenging Scientific Code Generation", "authors": ["Weifeng Sun", "Ye Fan", "Yuchen Chen", "Gou Tan", "Jieke Shi", "Yuan Yidi", "Swee Liang Wong", "Jonathan Pan", "David Lo"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19104", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark", "training-data"], "summary": "SciCodePile is a 128GB corpus of scientific code from 37,737 repositories plus a 200-task executable benchmark for evaluating and improving LLM scientific code generation. Experiments show current models perform poorly on executable scientific code tasks, while pretraining/fine-tuning on the corpus yields substantial gains.", "reason": "A dataset and benchmark serving the downstream task of generating scientific code, per the code-generation leaf.", "source": "crawl"}, {"paper": {"id": "2607.19338", "title": "CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents", "authors": ["Qijia He", "Jiayi Cheng", "Chenqian Le", "Rui Wang", "Xunmei Liu", "Yixian Chen", "Jie Mei", "Zhihao Wang", "Xupeng Chen", "Yuhuan Chen", "Tao Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19338", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "CodeRescue trains a supervised router with conformal risk control to decide, after a coding agent's failed attempt, whether to retry with cheap compute or escalate to a stronger model, improving cost-efficiency across five coding benchmarks.", "reason": "The method targets coding agents' post-failure recovery during code production tasks, so it serves general code generation rather than a specific lifecycle activity like debugging.", "source": "crawl"}, {"paper": {"id": "2607.18575", "title": "RECEIPT: Deterministic, Reward-Hacking-Resistant Verification for White-Box Agentic XSS Discovery", "authors": ["Muxi Lyu", "Karen Shieh", "Yiwei Hou", "Hao Wang", "Koushik Sen", "David Wagner"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18575", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "RECEIPT is a deterministic verification framework for LLM coding agents performing white-box, interactive XSS vulnerability discovery, preventing reward-hacked false claims via isolation, PoC constraints, and role separation. Evaluated on 95 real apps, it found 24 new XSS bugs and recovered labeled CVEs.", "reason": "Agent reasons over source code and executes tests to find and confirm real vulnerabilities, matching software_security.", "source": "crawl"}, {"paper": {"id": "2607.19267", "title": "They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface", "authors": ["Yohann Sidot"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19267", "github": "", "website": ""}}, "category": "software_security", "tags": ["empirical"], "summary": "An empirical red-team study of a five-agent LLM CI/CD pipeline shows that authority-framed prompt injection gets malicious data-exfiltration code laundered as a feature past security-scan and review agents at high rates, while content-based scanners miss the intent entirely.", "reason": "The pipeline's agents generate and security-review code; the study evaluates failure of vulnerability detection/auditing, fitting software_security.", "source": "crawl"}, {"paper": {"id": "2607.18356", "title": "CODENS: Transforming Code Changes into Living, Accessible, and Queryable Documentation", "authors": ["Abdelhak Kelious", "Chyrine Tahri", "Eliot Bardet"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18356", "github": "", "website": ""}}, "category": "software_comprehension", "tags": [], "summary": "CODENS builds a typed software knowledge graph from pull requests to generate living, queryable documentation, using agent-guided graph traversal for repository-level question answering.", "reason": "The served task is understanding and documenting an existing codebase, matching software_comprehension.", "source": "crawl"}, {"paper": {"id": "2607.18642", "title": "Spaghetti Architect: A Contamination-Resistant, By-Construction-Labelled, Multi-Language Code Dataset Generator", "authors": ["Yuxiang Ji"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18642", "github": "", "website": ""}}, "category": "software_comprehension", "tags": ["benchmark"], "summary": "Spaghetti Architect is a contamination-resistant multi-language code dataset generator that produces correct-by-construction programs with controllable messiness and difficulty axes, used to benchmark model comprehension, refactoring, and output-prediction abilities. It also exposes weaknesses like inflated scores from self-annotation and metric collapse on intrinsic complexity.", "reason": "A benchmark/dataset resource whose downstream tasks are code comprehension and refactoring assessment.", "source": "crawl"}, {"paper": {"id": "2607.18555", "title": "LM2Alloy: Investigating LLM-Generated Formal Specifications for Automated Test Derivation in Production Software", "authors": ["Tasmim Rashid", "Muhammad Zubair Malik"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18555", "github": "", "website": ""}}, "category": "software_testing", "tags": ["empirical"], "summary": "LM2Alloy explores using LLMs to generate Alloy formal specifications from requirements and source code, then derive executable test cases, uncovering a real bug missed by existing test suites in two Python libraries.", "reason": "The pipeline autonomously produces formal specifications and executable tests, serving proactive correctness assurance under software_testing.", "source": "crawl"}, {"paper": {"id": "2607.18711", "title": "LLM-Based Invariant Testing for Software Functional Bugs", "authors": ["Ruogu Yang", "Yifeng He", "Yundi Xu", "Yuqing Wei", "Hao Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18711", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "LISA is an LLM-based invariant testing framework that iteratively generates API call sequences and program invariants guided by n-gram feedback to proactively uncover functional (non-crashing) bugs in software libraries. It outperforms fuzzing and prior LLM test-generation baselines in bug detection and coverage.", "reason": "Proactive test/invariant generation for unknown functional defects fits software_testing, not a specific reported bug.", "source": "crawl"}, {"paper": {"id": "2607.19056", "title": "Vector-Bench: Can Models Surgically Edit SVG Code?", "authors": ["Yug Aditi Gupta", "Prannay Hebbar"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19056", "github": "", "website": ""}}, "category": "graphics", "tags": ["benchmark"], "summary": "Vector-Bench is a benchmark of 40 SVG repair tasks testing whether models can surgically edit vector graphics code to fix defects while leaving unrelated structure unchanged. It introduces a specification-based reward and evaluates 34 model endpoints, finding large gaps between repair progress and specification-faithful editing.", "reason": "Task delivers edited SVG code, a visual artifact produced from specification, matching the graphics leaf.", "source": "crawl"}, {"paper": {"id": "2607.18254", "title": "Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR", "authors": ["Plawan Kumar Rath"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-14", "links": {"paper": "https://arxiv.org/abs/2607.18254", "github": "", "website": ""}}, "category": "systems", "tags": ["benchmark"], "summary": "This paper studies inference-time, schema-derived constrained decoding for generating MLIR compiler dialect code without per-dialect fine-tuning, releasing four NL-to-MLIR benchmarks. A small model matches or beats much larger code LMs on structurally-dominated dialects using CFG, type-lattice, and SSA-scope constraints.", "reason": "The deliverable is compiler/IR code generation, which the systems leaf explicitly covers via 'compiler and runtime code'.", "source": "crawl"}, {"paper": {"id": "2604.03245", "title": "FVRuleLearner: Operator-Level Reasoning Tree (Op-Tree)-Based Rules Learning for Formal Verification", "authors": ["Lily Jiaxin Wan", "Chia-Tung Ho", "Yunsheng Bai", "Cunxi Yu", "Ghaith Bany Hamad", "Deming Chen", "Haoxing Ren"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2604.03245", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "FVRuleLearner is a rule-learning framework using an Operator Reasoning Tree to improve LLM generation of SystemVerilog Assertions from natural language for hardware formal verification, boosting syntax and functional correctness.", "reason": "Generates hardware verification code (SVA) from natural language, matching the hardware leaf's 'hardware verification via code'.", "source": "crawl"}, {"paper": {"id": "2605.06936", "title": "Bridging the Last Mile of Circuit Design: PostEDA-Bench, a Hierarchical Benchmark for PPA Convergence and DRC Fixing", "authors": ["Pengju Liu", "Nuo Xu", "Jinwei Tang", "Yu Cao", "Caiwen Ding"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2605.06936", "github": "", "website": ""}}, "category": "hardware", "tags": ["benchmark"], "summary": "PostEDA-Bench is a hierarchical benchmark evaluating LLM-based agents on 'last mile' EDA tasks: fixing sign-off design rule violations and converging power-performance-area targets across multiple toolchains.", "reason": "LLM agents perform hardware design/verification fixes, matching the hardware leaf's chip-design and verification tasks.", "source": "crawl"}, {"paper": {"id": "2607.18519", "title": "A Progressive Approach to Synthesizable RTL Design Generation Using LLMs", "authors": ["Xiangfei Kong", "Tasnim Tabassum", "Marwan Abdelwahab", "Hao Zheng"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18519", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "VeriRefine progressively refines natural-language hardware specifications into schema-constrained per-signal transition functions before generating Verilog, auditing interpretation errors separately from coding errors, achieving high functional correctness on RTLLM and VerilogEval.", "reason": "The task produces synthesizable RTL/Verilog code, matching artifact/hardware.", "source": "crawl"}, {"paper": {"id": "2607.18727", "title": "Formal Verification of an Out-of-Order Multiprocessor against an In-Order Weak-Memory ISA", "authors": ["Janggun Lee", "Jeehoon Kang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18727", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "Presents the first unbounded formal verification of an out-of-order multiprocessor against an in-order weak-memory ISA, decomposing the proof into core refinement and system inclusion. Proofs are mechanized in Rocq using LLM agents to automatically write proof scripts.", "reason": "Hardware microarchitecture verification via code (proof scripts) fits the hardware leaf's 'hardware verification via code' inclusion, not software formal verification.", "source": "crawl"}]
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 112 out-of-scope; 0 failed (retried next run).
1. Don't Blame the Large Language Model: How Agent Harness Evolution Shapes Coding Agent Quality
Oussama Ben Sghaier, Hao Li, Bram Adams, et al. · arXiv 2026/07 · paper
proposed:
studies· tags:empirical2. Beyond Resolved Rate: A Non-Functional Quality Study
Xin Sun, Daniel Ståhl, Kristian Sandahl, et al. · arXiv 2026/07 · paper
proposed:
studies· tags:empirical3. Skillware: A Software Ontology and Engineering Lifecycle for Persistent Behavioral Artifacts
Haodi Fan, Zucong Lan · arXiv 2026/07 · paper
proposed:
studies· tags:position4. Agentic AI-assisted coding offers a unique opportunity to instill epistemic grounding during software development
Magnus Palmblad, Jared M. Ragland, Benjamin A. Neely · arXiv 2026/07 · paper
proposed:
studies· tags:position5. Agents in the Wild: Where Research Meets Deployment
Grace Hui Yang, Pranav N. Venkit, Hooman Sedghamiz, et al. · arXiv 2026/07 · paper
proposed:
studies· tags:survey6. PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents
Tianyue Jiang, Yanlin Wang, Xin He, et al. · arXiv 2026/07 · paper
proposed:
software_debugging· tags: none7. Efficient Black-Box Fault Localization for System-Level Test Code Using Large Language Models
Ahmadreza Saboor Yaraghi, Golnaz Gharachorlu, Sakina Fatima, et al. · arXiv 2026/07 · paper
proposed:
software_debugging· tags: none8. FindStatBench: Evaluating Large Language Models on Combinatorial Code Synthesis
Soham Dan · arXiv 2026/05 · paper
proposed:
software_code_generation· tags:benchmark9. Decode-Time Grammars: Constrained LLM Generation over a Refinement Order of Grammar Fragments
Shuoming Zhang, Ruiyuan Xu, Haofeng Li, et al. · arXiv 2026/07 · paper
proposed:
software_code_generation· tags: none10. Beyond Text Editing: Algebraic Manipulation of Source Code
Kevin Pulo · arXiv 2026/07 · paper
proposed:
software_code_generation· tags: none11. TraceDev: A Traceability-Driven Multi-agent Framework for Requirement-to-Code Development
Mingyu Chen, Yakun Zhang, Zihao Xie, et al. · arXiv 2026/07 · paper
proposed:
software_code_generation· tags: none12. SciCodePile: A 128GB Corpus and Executable Benchmark for Challenging Scientific Code Generation
Weifeng Sun, Ye Fan, Yuchen Chen, et al. · arXiv 2026/07 · paper
proposed:
software_code_generation· tags:benchmarktraining-data13. CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents
Qijia He, Jiayi Cheng, Chenqian Le, et al. · arXiv 2026/07 · paper
proposed:
software_code_generation· tags: none14. RECEIPT: Deterministic, Reward-Hacking-Resistant Verification for White-Box Agentic XSS Discovery
Muxi Lyu, Karen Shieh, Yiwei Hou, et al. · arXiv 2026/07 · paper
proposed:
software_security· tags: none15. They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface
Yohann Sidot · arXiv 2026/07 · paper
proposed:
software_security· tags:empirical16. CODENS: Transforming Code Changes into Living, Accessible, and Queryable Documentation
Abdelhak Kelious, Chyrine Tahri, Eliot Bardet · arXiv 2026/07 · paper
proposed:
software_comprehension· tags: none17. Spaghetti Architect: A Contamination-Resistant, By-Construction-Labelled, Multi-Language Code Dataset Generator
Yuxiang Ji · arXiv 2026/07 · paper
proposed:
software_comprehension· tags:benchmark18. LM2Alloy: Investigating LLM-Generated Formal Specifications for Automated Test Derivation in Production Software
Tasmim Rashid, Muhammad Zubair Malik · arXiv 2026/07 · paper
proposed:
software_testing· tags:empirical19. LLM-Based Invariant Testing for Software Functional Bugs
Ruogu Yang, Yifeng He, Yundi Xu, et al. · arXiv 2026/07 · paper
proposed:
software_testing· tags: none20. Vector-Bench: Can Models Surgically Edit SVG Code?
Yug Aditi Gupta, Prannay Hebbar · arXiv 2026/07 · paper
proposed:
graphics· tags:benchmark21. Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR
Plawan Kumar Rath · arXiv 2026/05 · paper
proposed:
systems· tags:benchmark22. FVRuleLearner: Operator-Level Reasoning Tree (Op-Tree)-Based Rules Learning for Formal Verification
Lily Jiaxin Wan, Chia-Tung Ho, Yunsheng Bai, et al. · arXiv 2026/07 · paper
proposed:
hardware· tags: none23. Bridging the Last Mile of Circuit Design: PostEDA-Bench, a Hierarchical Benchmark for PPA Convergence and DRC Fixing
Pengju Liu, Nuo Xu, Jinwei Tang, et al. · arXiv 2026/07 · paper
proposed:
hardware· tags:benchmark24. A Progressive Approach to Synthesizable RTL Design Generation Using LLMs
Xiangfei Kong, Tasnim Tabassum, Marwan Abdelwahab, et al. · arXiv 2026/07 · paper
proposed:
hardware· tags: none25. Formal Verification of an Out-of-Order Multiprocessor against an In-Order Weak-Memory ISA
Janggun Lee, Jeehoon Kang · arXiv 2026/07 · paper
proposed:
hardware· tags: nonemachine payload (do not edit)
[{"paper": {"id": "2607.03691", "title": "Don't Blame the Large Language Model: How Agent Harness Evolution Shapes Coding Agent Quality", "authors": ["Oussama Ben Sghaier", "Hao Li", "Bram Adams", "Ahmed E. Hassan"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.03691", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A controlled longitudinal study isolating how evolving agent harnesses (Codex, Qwen Code, Gemini, OpenCode, OpenHands) affect coding agent quality, holding the LLM constant across 35 sequential Qwen Code releases evaluated on SWE-bench Verified.", "reason": "The paper's object is the coding agents/harnesses themselves rather than proposing a task-performing agent, matching the studies/empirical definition.", "source": "crawl"}, {"paper": {"id": "2607.18462", "title": "Beyond Resolved Rate: A Non-Functional Quality Study", "authors": ["Xin Sun", "Daniel Ståhl", "Kristian Sandahl", "Christoph Kessler"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18462", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "This study empirically compares non-functional quality (static-analysis warnings, CPU time, memory) of patches generated by successive Claude and DeepSeek models via SWE-agent on SWE-bench Lite. It finds resolved rates improve across generations but non-functional patch quality does not consistently improve.", "reason": "An empirical study whose object is analyzing coding-agent output quality trends across model generations, not a new agent or tool.", "source": "crawl"}, {"paper": {"id": "2607.18970", "title": "Skillware: A Software Ontology and Engineering Lifecycle for Persistent Behavioral Artifacts", "authors": ["Haodi Fan", "Zucong Lan"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18970", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "Skillware proposes a software ontology treating 'Agent Skills' as persistent behavioral artifacts with independent identity and lifecycle, grounded in analysis of over 138,000 SKILL.md records and category-boundary case studies.", "reason": "Object of study is the AI-agent-skill ecosystem itself (ontology/taxonomy), not a proposed agent performing a task, matching the studies/position leaf.", "source": "crawl"}, {"paper": {"id": "2604.21744", "title": "Agentic AI-assisted coding offers a unique opportunity to instill epistemic grounding during software development", "authors": ["Magnus Palmblad", "Jared M. Ragland", "Benjamin A. Neely"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2604.21744", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "Proposes GROUNDING.md, a community-governed, field-scoped document encoding hard constraints and conventions to keep domain expertise embedded in agentic AI-assisted software development, illustrated with a proteomics example. It is a position piece advocating a governance practice for agentic coding rather than a working agent or benchmark.", "reason": "Position/vision paper about the practice of agentic software development, not a proposed agent or task-performing method, so it fits the studies off-axis branch.", "source": "crawl"}, {"paper": {"id": "2607.19336", "title": "Agents in the Wild: Where Research Meets Deployment", "authors": ["Grace Hui Yang", "Pranav N. Venkit", "Hooman Sedghamiz", "Enrico Santus", "Victor Dibia", "Ioana Baldini"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19336", "github": "", "website": ""}}, "category": "studies", "tags": ["survey"], "summary": "A tutorial surveying agentic LLM systems moving from research to production deployment across software engineering, scientific discovery, and finance, distilling design patterns, evaluation checklists, and mitigation strategies for reliability and safety.", "reason": "Field-wide tutorial whose object is agentic systems and deployment practice broadly, not a task-specific agent proposal, matching the studies leaf.", "source": "crawl"}, {"paper": {"id": "2607.18859", "title": "PhoenixRepair: Rethinking Repair Strategy Exploration in Software Agents", "authors": ["Tianyue Jiang", "Yanlin Wang", "Xin He", "Daya Guo", "Jiachi Chen", "Ming Wen", "Ensheng Shi", "Xilin Liu", "Yuchi Ma", "Guanbin Li"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18859", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "PhoenixRepair is a multi-agent issue-resolution framework that expands repair-strategy exploration via multi-location sampling and iterative reflection/refinement on patches. It achieves state-of-the-art resolved rates on SWE-bench-Verified across multiple backbone LLMs.", "reason": "An agent framework for reproducing, localizing, and repairing reported software issues is squarely software_debugging (SWE-bench-style issue resolution).", "source": "crawl"}, {"paper": {"id": "2506.19045", "title": "Efficient Black-Box Fault Localization for System-Level Test Code Using Large Language Models", "authors": ["Ahmadreza Saboor Yaraghi", "Golnaz Gharachorlu", "Sakina Fatima", "Lionel C. Briand", "Ruiyuan Wan", "Ruifeng Gao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2506.19045", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "A static, LLM-driven fault-localization method for system-level test code that estimates execution traces from a single failure log and prompts an LLM to rank likely faulty locations without executing the test.", "reason": "Fault localization and root-cause analysis of code is explicitly included under software_debugging.", "source": "crawl"}, {"paper": {"id": "2607.18260", "title": "FindStatBench: Evaluating Large Language Models on Combinatorial Code Synthesis", "authors": ["Soham Dan"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-16", "links": {"paper": "https://arxiv.org/abs/2607.18260", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark"], "summary": "FindStatBench is an execution benchmark of 2,329 combinatorial code synthesis tasks requiring LLMs to emit a single Python solve function with no tool use or feedback, scored by sandboxed execution on hidden instances. It finds statistic synthesis easier than map synthesis and reveals brittleness in symbolic rule induction.", "reason": "A benchmark for producing code from specification (competitive-programming style), matching software_code_generation.", "source": "crawl"}, {"paper": {"id": "2607.18357", "title": "Decode-Time Grammars: Constrained LLM Generation over a Refinement Order of Grammar Fragments", "authors": ["Shuoming Zhang", "Ruiyuan Xu", "Haofeng Li", "Qiuchu Yu", "Yangyu Zhang", "Chunwei Xia", "Xiaobing Feng", "Chenxi Wang", "Huimin Cui", "Jiacheng Zhao"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18357", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "Proposes decode-time grammars that instantiate environment-typed grammar fragments during LLM decoding to guarantee semantically valid code for low-resource DSLs, APIs, and CLIs, tested on TileLang, SQL, and P4.", "reason": "The technique's purpose is producing semantically correct code, serving the code generation task despite being a decoding method.", "source": "crawl"}, {"paper": {"id": "2607.18742", "title": "Beyond Text Editing: Algebraic Manipulation of Source Code", "authors": ["Kevin Pulo"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18742", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "Proposes source code algebra, a set of composable algebraic operations for modifying codebases as an alternative to text-level edits, aimed at reducing friction for LLM coding agents making non-local changes. A proof-of-concept (SCAS) shows higher success rates and far fewer tokens on cross-file code changes versus text-editing baselines.", "reason": "An agentic method for producing/completing cross-file code edits serves code generation and editing, the task's ultimate purpose is producing code.", "source": "crawl"}, {"paper": {"id": "2607.18886", "title": "TraceDev: A Traceability-Driven Multi-agent Framework for Requirement-to-Code Development", "authors": ["Mingyu Chen", "Yakun Zhang", "Zihao Xie", "Yixing Luo", "Jinrui Xu", "Cuiyun Gao", "Kaiqi Zhao", "Yunming Ye"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18886", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "TraceDev is a multi-agent framework with five role-specific agents that transforms complex natural-language requirements (use cases) into repository-level code, using a traceability graph linking requirements, design, and code artifacts. It outperforms baselines on ETOUR and SMOS datasets.", "reason": "Multi-agent system whose ultimate purpose is producing repository-level code from requirements, matching software_code_generation's whole-repo/app generation.", "source": "crawl"}, {"paper": {"id": "2607.19104", "title": "SciCodePile: A 128GB Corpus and Executable Benchmark for Challenging Scientific Code Generation", "authors": ["Weifeng Sun", "Ye Fan", "Yuchen Chen", "Gou Tan", "Jieke Shi", "Yuan Yidi", "Swee Liang Wong", "Jonathan Pan", "David Lo"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19104", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark", "training-data"], "summary": "SciCodePile is a 128GB corpus of scientific code from 37,737 repositories plus a 200-task executable benchmark for evaluating and improving LLM scientific code generation. Experiments show current models perform poorly on executable scientific code tasks, while pretraining/fine-tuning on the corpus yields substantial gains.", "reason": "A dataset and benchmark serving the downstream task of generating scientific code, per the code-generation leaf.", "source": "crawl"}, {"paper": {"id": "2607.19338", "title": "CodeRescue: Budget-Calibrated Recovery Routing for Coding Agents", "authors": ["Qijia He", "Jiayi Cheng", "Chenqian Le", "Rui Wang", "Xunmei Liu", "Yixian Chen", "Jie Mei", "Zhihao Wang", "Xupeng Chen", "Yuhuan Chen", "Tao Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19338", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "CodeRescue trains a supervised router with conformal risk control to decide, after a coding agent's failed attempt, whether to retry with cheap compute or escalate to a stronger model, improving cost-efficiency across five coding benchmarks.", "reason": "The method targets coding agents' post-failure recovery during code production tasks, so it serves general code generation rather than a specific lifecycle activity like debugging.", "source": "crawl"}, {"paper": {"id": "2607.18575", "title": "RECEIPT: Deterministic, Reward-Hacking-Resistant Verification for White-Box Agentic XSS Discovery", "authors": ["Muxi Lyu", "Karen Shieh", "Yiwei Hou", "Hao Wang", "Koushik Sen", "David Wagner"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18575", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "RECEIPT is a deterministic verification framework for LLM coding agents performing white-box, interactive XSS vulnerability discovery, preventing reward-hacked false claims via isolation, PoC constraints, and role separation. Evaluated on 95 real apps, it found 24 new XSS bugs and recovered labeled CVEs.", "reason": "Agent reasons over source code and executes tests to find and confirm real vulnerabilities, matching software_security.", "source": "crawl"}, {"paper": {"id": "2607.19267", "title": "They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface", "authors": ["Yohann Sidot"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19267", "github": "", "website": ""}}, "category": "software_security", "tags": ["empirical"], "summary": "An empirical red-team study of a five-agent LLM CI/CD pipeline shows that authority-framed prompt injection gets malicious data-exfiltration code laundered as a feature past security-scan and review agents at high rates, while content-based scanners miss the intent entirely.", "reason": "The pipeline's agents generate and security-review code; the study evaluates failure of vulnerability detection/auditing, fitting software_security.", "source": "crawl"}, {"paper": {"id": "2607.18356", "title": "CODENS: Transforming Code Changes into Living, Accessible, and Queryable Documentation", "authors": ["Abdelhak Kelious", "Chyrine Tahri", "Eliot Bardet"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18356", "github": "", "website": ""}}, "category": "software_comprehension", "tags": [], "summary": "CODENS builds a typed software knowledge graph from pull requests to generate living, queryable documentation, using agent-guided graph traversal for repository-level question answering.", "reason": "The served task is understanding and documenting an existing codebase, matching software_comprehension.", "source": "crawl"}, {"paper": {"id": "2607.18642", "title": "Spaghetti Architect: A Contamination-Resistant, By-Construction-Labelled, Multi-Language Code Dataset Generator", "authors": ["Yuxiang Ji"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18642", "github": "", "website": ""}}, "category": "software_comprehension", "tags": ["benchmark"], "summary": "Spaghetti Architect is a contamination-resistant multi-language code dataset generator that produces correct-by-construction programs with controllable messiness and difficulty axes, used to benchmark model comprehension, refactoring, and output-prediction abilities. It also exposes weaknesses like inflated scores from self-annotation and metric collapse on intrinsic complexity.", "reason": "A benchmark/dataset resource whose downstream tasks are code comprehension and refactoring assessment.", "source": "crawl"}, {"paper": {"id": "2607.18555", "title": "LM2Alloy: Investigating LLM-Generated Formal Specifications for Automated Test Derivation in Production Software", "authors": ["Tasmim Rashid", "Muhammad Zubair Malik"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18555", "github": "", "website": ""}}, "category": "software_testing", "tags": ["empirical"], "summary": "LM2Alloy explores using LLMs to generate Alloy formal specifications from requirements and source code, then derive executable test cases, uncovering a real bug missed by existing test suites in two Python libraries.", "reason": "The pipeline autonomously produces formal specifications and executable tests, serving proactive correctness assurance under software_testing.", "source": "crawl"}, {"paper": {"id": "2607.18711", "title": "LLM-Based Invariant Testing for Software Functional Bugs", "authors": ["Ruogu Yang", "Yifeng He", "Yundi Xu", "Yuqing Wei", "Hao Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18711", "github": "", "website": ""}}, "category": "software_testing", "tags": [], "summary": "LISA is an LLM-based invariant testing framework that iteratively generates API call sequences and program invariants guided by n-gram feedback to proactively uncover functional (non-crashing) bugs in software libraries. It outperforms fuzzing and prior LLM test-generation baselines in bug detection and coverage.", "reason": "Proactive test/invariant generation for unknown functional defects fits software_testing, not a specific reported bug.", "source": "crawl"}, {"paper": {"id": "2607.19056", "title": "Vector-Bench: Can Models Surgically Edit SVG Code?", "authors": ["Yug Aditi Gupta", "Prannay Hebbar"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.19056", "github": "", "website": ""}}, "category": "graphics", "tags": ["benchmark"], "summary": "Vector-Bench is a benchmark of 40 SVG repair tasks testing whether models can surgically edit vector graphics code to fix defects while leaving unrelated structure unchanged. It introduces a specification-based reward and evaluates 34 model endpoints, finding large gaps between repair progress and specification-faithful editing.", "reason": "Task delivers edited SVG code, a visual artifact produced from specification, matching the graphics leaf.", "source": "crawl"}, {"paper": {"id": "2607.18254", "title": "Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR", "authors": ["Plawan Kumar Rath"], "venue": "arXiv 2026/05", "category": "", "published": "2026-05-14", "links": {"paper": "https://arxiv.org/abs/2607.18254", "github": "", "website": ""}}, "category": "systems", "tags": ["benchmark"], "summary": "This paper studies inference-time, schema-derived constrained decoding for generating MLIR compiler dialect code without per-dialect fine-tuning, releasing four NL-to-MLIR benchmarks. A small model matches or beats much larger code LMs on structurally-dominated dialects using CFG, type-lattice, and SSA-scope constraints.", "reason": "The deliverable is compiler/IR code generation, which the systems leaf explicitly covers via 'compiler and runtime code'.", "source": "crawl"}, {"paper": {"id": "2604.03245", "title": "FVRuleLearner: Operator-Level Reasoning Tree (Op-Tree)-Based Rules Learning for Formal Verification", "authors": ["Lily Jiaxin Wan", "Chia-Tung Ho", "Yunsheng Bai", "Cunxi Yu", "Ghaith Bany Hamad", "Deming Chen", "Haoxing Ren"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2604.03245", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "FVRuleLearner is a rule-learning framework using an Operator Reasoning Tree to improve LLM generation of SystemVerilog Assertions from natural language for hardware formal verification, boosting syntax and functional correctness.", "reason": "Generates hardware verification code (SVA) from natural language, matching the hardware leaf's 'hardware verification via code'.", "source": "crawl"}, {"paper": {"id": "2605.06936", "title": "Bridging the Last Mile of Circuit Design: PostEDA-Bench, a Hierarchical Benchmark for PPA Convergence and DRC Fixing", "authors": ["Pengju Liu", "Nuo Xu", "Jinwei Tang", "Yu Cao", "Caiwen Ding"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2605.06936", "github": "", "website": ""}}, "category": "hardware", "tags": ["benchmark"], "summary": "PostEDA-Bench is a hierarchical benchmark evaluating LLM-based agents on 'last mile' EDA tasks: fixing sign-off design rule violations and converging power-performance-area targets across multiple toolchains.", "reason": "LLM agents perform hardware design/verification fixes, matching the hardware leaf's chip-design and verification tasks.", "source": "crawl"}, {"paper": {"id": "2607.18519", "title": "A Progressive Approach to Synthesizable RTL Design Generation Using LLMs", "authors": ["Xiangfei Kong", "Tasnim Tabassum", "Marwan Abdelwahab", "Hao Zheng"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.18519", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "VeriRefine progressively refines natural-language hardware specifications into schema-constrained per-signal transition functions before generating Verilog, auditing interpretation errors separately from coding errors, achieving high functional correctness on RTLLM and VerilogEval.", "reason": "The task produces synthesizable RTL/Verilog code, matching artifact/hardware.", "source": "crawl"}, {"paper": {"id": "2607.18727", "title": "Formal Verification of an Out-of-Order Multiprocessor against an In-Order Weak-Memory ISA", "authors": ["Janggun Lee", "Jeehoon Kang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-21", "links": {"paper": "https://arxiv.org/abs/2607.18727", "github": "", "website": ""}}, "category": "hardware", "tags": [], "summary": "Presents the first unbounded formal verification of an out-of-order multiprocessor against an in-order weak-memory ISA, decomposing the proof into core refinement and system inclusion. Proofs are mechanized in Rocq using LLM agents to automatically write proof scripts.", "reason": "Hardware microarchitecture verification via code (proof scripts) fits the hardware leaf's 'hardware verification via code' inclusion, not software formal verification.", "source": "crawl"}]