[{"paper": {"id": "2604.02544", "title": "Developer Experience with AI Coding Agents: HTTP Behavioral Signatures in Documentation Portals", "authors": ["Oleksii Borysenko"], "venue": "arXiv 2026/04", "category": "", "published": "2026-04-02", "links": {"paper": "https://arxiv.org/abs/2604.02544", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical study analyzing HTTP request fingerprints from nine AI coding agents and six AI assistant services accessing developer documentation portals, revealing distinct behavioral signatures. It discusses implications for documentation design and analytics given AI-driven access patterns.", "reason": "Object of study is behavior of AI coding agents in the wild, not a proposed method for a task, so it fits the empirical studies leaf.", "source": "crawl"}, {"paper": {"id": "2607.22585", "title": "The Scaffold Effect in Coding Agents: Harness Choice as a Hidden Variable in Coding-Agent Evaluation", "authors": ["Naman Vats", "Oleg Golev"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-08", "links": {"paper": "https://arxiv.org/abs/2607.22585", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical study showing that the choice of agent harness (Goose, OpenCode, OpenHands-SDK) causes up to 40x differences in token cost for coding agents on Terminal-Bench Pro, while pass rates stay similar across harnesses. It argues model name alone is an incomplete unit for coding-agent evaluation and harness-level failure biases must be reported.", "reason": "Object of study is coding-agent evaluation methodology itself (harness vs model effects), an empirical study rather than a proposed agent/method for a task.", "source": "crawl"}, {"paper": {"id": "2607.22807", "title": "The Best Programming Language for Tokenmaxxing: An Investigation of Coding Agent Behavior Across Programming Languages", "authors": ["Zixuan Wu", "Carolyn Jane Anderson", "Arjun Guha"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22807", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical investigation of coding agent token consumption across Python, Java, Rust, and OCaml, analyzing trajectory structure and text to explain why costs vary by language. Finds agents produce more noncompiling code and distrust tests in unfamiliar languages.", "reason": "This is a behavioral/empirical study of existing coding agents' behavior rather than a new agent, method, or benchmark performing a task.", "source": "crawl"}, {"paper": {"id": "2607.23332", "title": "AlloBench: Measuring Online Tool Allocation Capability in LLM Agents", "authors": ["Daniel Wang", "Andrew Xu"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-25", "links": {"paper": "https://arxiv.org/abs/2607.23332", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical", "benchmark"], "summary": "AlloBench is a paired benchmark testing whether LLM agents rationally allocate effort toward building reusable tools versus one-off scripts, comparing abstract and code-construction settings across frontier models.", "reason": "An empirical/capability study of agent behavior (tool-investment decision-making) rather than a task-performing system, so it fits the off-axis studies leaf.", "source": "crawl"}, {"paper": {"id": "2607.23670", "title": "Plans Work in Mysterious Ways: Evaluating a Plan Mode for Spreadsheet Agents", "authors": ["Aayush Kumar", "Avik Dutta", "Sumit Gulwani", "Gustavo Soares", "Advait Sarkar", "Emerson Murphy-Hill"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-26", "links": {"paper": "https://arxiv.org/abs/2607.23670", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A within-subjects user study (N=24) evaluating a prototype Plan Mode for spreadsheet agents against a non-planning baseline, examining effects on refinement behavior and user perception in end-user programming. Finds Plan Mode improves perceived creativity support and collaboration despite similar task outcomes.", "reason": "An empirical human-agent interaction study of an existing agentic feature, proposing no new task-performing method, per the studies/empirical precedent.", "source": "crawl"}, {"paper": {"id": "2607.24601", "title": "Evaluating the Impact of Explainable AI on Trust in AI-Assisted Code Review", "authors": ["Zhenhan Gao", "Marvin Muñoz Barón", "Umm-e Habiba", "Daniel Graziotin", "Stefan Wagner"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24601", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A within-subjects user study of 34 developers examining how explanation level in LLM-based code review systems affects developer trust and agreement with AI recommendations. It studies human-agent interaction around AI code review tools rather than proposing a new review agent.", "reason": "Empirical/behavioral study of human-agent interaction with AI-assisted code review tools fits the studies leaf.", "source": "crawl"}, {"paper": {"id": "2605.17675", "title": "Bridging the Gap on AI-Assisted Scientific Software Development Through Transparency and Traceability", "authors": ["Chaitanya Bhave", "Pierre-Clément A. Simon", "Casey Icenhour", "Lin Yang", "Cody J. Permann", "Daniel Schwen", "Christopher S. Ritter"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2605.17675", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "A position paper proposing governance guidance for AI-assisted development of safety-critical scientific software under strict Software Quality Assurance (e.g. NQA-1) standards, using a fusion-energy simulation code as a demonstration case. It focuses on transparency, traceability, and human accountability rather than proposing a new agent or tool.", "reason": "Vision/position paper about the practice of AI-assisted software development, matching the studies leaf's position-paper scope.", "source": "crawl"}, {"paper": {"id": "2607.22711", "title": "CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents", "authors": ["Mingwei Zheng", "David OBrien", "Siwei Cui", "Pardis Pashakhanloo", "Rajdeep Mukherjee", "Myeongsoo Kim", "Sachit Kuhar"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.22711", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "Proposes CORVUS, a trajectory architecture for LLM coding agents that decouples file-read actions from stale observation snapshots, keeping context synchronized with the live codebase. Evaluated on SWE-PolyBench and SWE-Bench Pro, it cuts input tokens and reasoning cycles while preserving pass rates.", "reason": "Evaluated on SWE-bench-style issue-resolution benchmarks, routing to software_debugging per benchmark-routing rule.", "source": "crawl"}, {"paper": {"id": "2607.24604", "title": "Looping Is Not Reliability: State-Bound Evidence and Typed Revision Contracts for Agentic Code Repair", "authors": ["Xueping Gao", "Jianwei Yang", "Qiang Yang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24604", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["empirical"], "summary": "An empirical study of generate-test-revise loops in agentic code repair on HumanEval, showing repeated revision can degrade correctness, and proposing a typed revision contract binding verifier evidence to code states.", "reason": "The served task is program repair/issue resolution reliability, matching software_debugging.", "source": "crawl"}, {"paper": {"id": "2607.24882", "title": "Agent Retrieval Bench: Evaluating Repository Context Retrieval for Coding Agents", "authors": ["Bowen Qin", "Yi Xie"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24882", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["benchmark"], "summary": "Agent Retrieval Bench evaluates repository context retrieval methods that feed patch generation for coding agents, covering tasks like code2test, comment2context, trace2code, and edit2ripple across 427 samples in 25 repositories.", "reason": "Serves the upstream context-acquisition stage of SWE-bench-style issue resolution/patch generation, so it routes to software_debugging plus benchmark tag.", "source": "crawl"}, {"paper": {"id": "2607.25873", "title": "How Do LLMs Read Bug Reports? An Empirical Study of Attention in LLMs for Automated Program Repair", "authors": ["Ramtin Ehsani", "Irene Manotas", "Saurabh Pujar", "Luca Buratti", "Preetha Chatterjee"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25873", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["empirical"], "summary": "An empirical study analyzing attention patterns of LLM-based automated program repair systems over bug reports, finding that diffused attention across diagnostic sections correlates with repair success while over-localized attention correlates with failure.", "reason": "A single-activity (program repair) empirical/failure-analysis study routes to software_debugging plus the empirical tag, per the studies boundary rule.", "source": "crawl"}, {"paper": {"id": "2507.22080", "title": "CodeEvo: Interaction-Driven Synthesis of Code-centric Data through Hybrid and Iterative Feedback", "authors": ["Qiushi Sun", "Jinyang Gong", "Lei Li", "Qipeng Guo", "Fei Yuan"], "venue": "arXiv 2025/07", "category": "", "published": "2025-07-25", "links": {"paper": "https://arxiv.org/abs/2507.22080", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["training-data"], "summary": "CodeEvo uses a dual-agent Coder/Reviewer architecture with hybrid compiler and semantic feedback to iteratively synthesize high-quality instruction-code pairs. The resulting CodeEvo-100K dataset improves fine-tuned models' performance across code generation benchmarks.", "reason": "Resource paper whose downstream task is code generation, so artifact/software_code_generation plus training-data tag.", "source": "crawl"}, {"paper": {"id": "2607.22898", "title": "AssumptionMiner: Extracting, Tracing, and Revising Implicit Assumptions in LLM Code Generation", "authors": ["Jie \"JW\" Wu"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22898", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark"], "summary": "AssumptionMiner extracts, traces, and revises implicit assumptions LLMs make when generating code from underspecified prompts, producing an explicit assumption layer and enabling AST-guided targeted regeneration. Introduces a 180-task benchmark with 676 annotated assumptions to evaluate extraction, localization, and revision.", "reason": "Serves the code generation task by improving correctness of LLM-generated code from ambiguous specs, with a released benchmark.", "source": "crawl"}, {"paper": {"id": "2607.24884", "title": "Beyond \"What to Retrieve\": Uncertainty in Retrieval-Augmented Code Generation", "authors": ["Chandan Kumar Sah", "Xiaoli Lian", "Li Zhang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24884", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "OpenCoder is a framework that estimates uncertainty in heterogeneous retrieved evidence (similar code, repo context, APIs) to filter and rank it, guiding repository-level code generation, verification, and repair. It improves selected-output correctness on an expanded RepoExec-inline benchmark, though gains are backend-dependent.", "reason": "Task ultimately serves repository-level code generation with an active generate-verify-repair pipeline, not a pure metric.", "source": "crawl"}, {"paper": {"id": "2607.25130", "title": "Learning from 53.6K Real-World Developer Edits of AI-Generated Code", "authors": ["Jenny T. Liang", "Mihika Bairathi", "Wayne Chi", "Ameet Talwalkar", "Nishant Subramani", "Valerie Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.25130", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["training-data", "benchmark"], "summary": "DECODE is a 53.6K-example dataset of real-world in-IDE developer edits to AI-generated code across Python, TypeScript, and JavaScript, used both to analyze editing behavior and to benchmark/finetune LLMs on code-edit prediction. Finetuned small models substantially outperform frontier LLMs at predicting edits.", "reason": "A resource paper whose downstream task is repo-aware code completion/next-edit prediction, matching software_code_generation's resource-paper scope.", "source": "crawl"}, {"paper": {"id": "2607.22917", "title": "Agent Team Work Zone: An Automated, Persistent Workspace for Long-Lived Coding Agent Teams", "authors": ["Shouren Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22917", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": [], "summary": "ATWZ is a filesystem-based operations layer built around Claude Code's Agent Teams that persists agent working state, skills, and inter-agent messages to enable recovery, resumption, and reduced prompt overhead in long-lived coding agent teams.", "reason": "Enabling tooling that preserves and restores agent operational state in service of sustained software development work, matching software_infrastructure's scope of enabling code production.", "source": "crawl"}, {"paper": {"id": "2607.25141", "title": "Specification-Driven DevOps for Multi-Service Environments", "authors": ["Oleg Grynets", "Kyrylo Fursov", "Vasyl Lyashkevych", "Volodymyr Veres"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.25141", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["empirical"], "summary": "This study evaluates whether a frontier LLM can generate functionally correct Dockerfiles and Docker Compose configurations for multi-service repositories without developer-authored deployment artifacts, finding functional success but gaps in deployment-intent fidelity (e.g., missing network segmentation, multi-stage builds). It proposes a minimal deployment specification to close this gap.", "reason": "LLM-generated build/deployment configuration for running code is environment-setup/CI-CD enabling work under software_infrastructure.", "source": "crawl"}, {"paper": {"id": "2606.31767", "title": "JETO-Bench: A Reproducible Benchmark for Execution Time Improvement Patches in Java", "authors": ["Khashayar Etemadi", "Zhendong Su"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-30", "links": {"paper": "https://arxiv.org/abs/2606.31767", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["benchmark"], "summary": "Presents JETO-Mine, a configurable pipeline to mine reproducible execution-time-improvement patch (ETIP) benchmarks from Java repositories, and JETO-Bench, a benchmark of 660 mined and 91 verified ETIPs. Evaluates OpenHands with GPT-5-mini, which fixes 14.3% of issues.", "reason": "Serves performance optimization of existing code while preserving behavior, matching software_maintenance's boundary definition.", "source": "crawl"}, {"paper": {"id": "2607.22588", "title": "ParBench: A Benchmark for Reliable Evaluation of LLM Parallel Code Translation", "authors": ["Samyak Jhaveri", "Erel Kaplan", "Tom Yotam", "Le Chen", "Tomer Bitan", "Niranjan Hasabnis", "Gal Oren"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-09", "links": {"paper": "https://arxiv.org/abs/2607.22588", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["benchmark"], "summary": "ParBench is a kernel-centric benchmark for evaluating LLM-based translation of parallel code kernels across CUDA, OpenMP, OpenCL, and OpenMP target offload, testing behavioral fidelity and robustness to source perturbations. It reveals persistent weaknesses in cross-API translation, multi-file coordination, and robustness.", "reason": "Behavior-preserving code translation across APIs/languages matches software_maintenance's migration boundary.", "source": "crawl"}, {"paper": {"id": "2607.23665", "title": "Multi-level Code Optimization via Mixture of Prompts", "authors": ["Yun Peng", "Jun Wan", "Jiakun Liu", "Shuzheng Gao", "David Lo", "Xiaoxue Ren"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-26", "links": {"paper": "https://arxiv.org/abs/2607.23665", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "Optimo is an LLM-based multi-level code optimization approach using a Mixture-of-Prompts architecture that identifies performance bottlenecks via differential profiling and applies tailored optimization strategies across four abstraction levels. It achieves up to 57.48% opt% and 13.51x speedups on efficiency benchmarks COFFE and Effibench.", "reason": "Performance optimization of existing code while preserving behavior falls under software_maintenance.", "source": "crawl"}, {"paper": {"id": "2602.07195", "title": "Automated Modernization of Machine Learning Engineering Notebooks for Reproducibility", "authors": ["Bihui Jin", "Kaiyuan Wang", "Pengyu Nie"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2602.07195", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "MLEModernizer is an LLM-driven agentic framework that iteratively executes, diagnoses, and repairs Jupyter notebooks to restore reproducibility as ML software and hardware ecosystems evolve. Evaluated on over 8,000 non-reproducible notebooks, it successfully modernizes roughly 40-45 percent of them.", "reason": "Behavior-preserving adaptation of code to new dependency and environment versions is software_maintenance.", "source": "crawl"}, {"paper": {"id": "2607.24998", "title": "VPR-Evolve: Multi-Agent-Driven Algorithm Evolution for FPGA Place and Route", "authors": ["Qihang Wu", "Taizun Jafri", "Aman Arora", "Vidya A. Chhabria"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24998", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "VPR-Evolve uses LLM agents to propose, implement, and evaluate source-code modifications to the VPR FPGA place-and-route engine, guided by shared memory of prior outcomes. It improves critical-path delay, wirelength, and runtime over stock VPR and hyperparameter-tuning baselines on VTR-9 benchmarks.", "reason": "Agents evolve an existing CAD tool's source code to optimize performance metrics while preserving functionality, matching maintenance/performance-optimization.", "source": "crawl"}, {"paper": {"id": "2607.25970", "title": "Reinforcement Learning for Code Optimization", "authors": ["Pierre Chambon", "Kunhao Zheng", "Juliette Decugis", "Benoit Sagot", "Gabriel Synnaeve"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25970", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["benchmark"], "summary": "Trains RL agents to optimize code execution speed while preserving correctness, introducing DMC-Optim with a calibrated timing sandbox and reward design to overcome noise in timed-execution RL. Shows large pass@1 gains in performance-oriented code optimization over standard RLVR.", "reason": "behavior-preserving performance optimization of existing code matches software_maintenance", "source": "crawl"}, {"paper": {"id": "2607.22569", "title": "Execution-Grounded Security Testing for Coding Agents in Software Engineering Pipelines", "authors": ["Yifei Ge", "Weisong Sun", "Jinkun Xiao", "Yuchen Chen", "Yebo Feng", "Peizhuo Lv", "Xia Feng", "Chunrong Fang", "Zhihong Zhao", "Zhenyu Chen", "Yang Liu"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-01", "links": {"paper": "https://arxiv.org/abs/2607.22569", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "Presents an execution-grounded red-team testing framework that embeds unsafe system operations inside routine software-engineering tasks to probe whether coding agents can be induced to perform unauthorized system-level actions. Experiments show high rates of verified unsafe execution across agent frameworks and models.", "reason": "A security-assurance testing framework targeting coding agents' unsafe actions during SE pipeline tasks fits software_security's security-testing scope.", "source": "crawl"}, {"paper": {"id": "2607.23088", "title": "Poster: Rethinking Security in LLM Code Generation through Real-World Risk Scenarios", "authors": ["Lixun Ma", "Ruolong Ma", "Bei Wang", "Feng Wei", "Zhenguang Liu", "Lorenzo Cavallaro", "Wentao Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-25", "links": {"paper": "https://arxiv.org/abs/2607.23088", "github": "", "website": ""}}, "category": "software_security", "tags": ["benchmark"], "summary": "Introduces a 2,700-case benchmark of realistic ambiguous/under-specified prompts to evaluate security vulnerabilities in LLM-generated code, finding over 56% average vulnerability rates across eight LLMs. Shows security-aware prompting can mitigate risks by up to 45%.", "reason": "A benchmark resource serving secure code generation, routed to software_security per code-generation security scope.", "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 344 out-of-scope; 0 failed (retried next run).
1. Developer Experience with AI Coding Agents: HTTP Behavioral Signatures in Documentation Portals
Oleksii Borysenko · arXiv 2026/04 · paper
proposed:
studies· tags:empirical2. The Scaffold Effect in Coding Agents: Harness Choice as a Hidden Variable in Coding-Agent Evaluation
Naman Vats, Oleg Golev · arXiv 2026/06 · paper
proposed:
studies· tags:empirical3. The Best Programming Language for Tokenmaxxing: An Investigation of Coding Agent Behavior Across Programming Languages
Zixuan Wu, Carolyn Jane Anderson, Arjun Guha · arXiv 2026/07 · paper
proposed:
studies· tags:empirical4. AlloBench: Measuring Online Tool Allocation Capability in LLM Agents
Daniel Wang, Andrew Xu · arXiv 2026/07 · paper
proposed:
studies· tags:empiricalbenchmark5. Plans Work in Mysterious Ways: Evaluating a Plan Mode for Spreadsheet Agents
Aayush Kumar, Avik Dutta, Sumit Gulwani, et al. · arXiv 2026/07 · paper
proposed:
studies· tags:empirical6. Evaluating the Impact of Explainable AI on Trust in AI-Assisted Code Review
Zhenhan Gao, Marvin Muñoz Barón, Umm-e Habiba, et al. · arXiv 2026/07 · paper
proposed:
studies· tags:empirical7. Bridging the Gap on AI-Assisted Scientific Software Development Through Transparency and Traceability
Chaitanya Bhave, Pierre-Clément A. Simon, Casey Icenhour, et al. · arXiv 2026/07 · paper
proposed:
studies· tags:position8. CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents
Mingwei Zheng, David OBrien, Siwei Cui, et al. · arXiv 2026/07 · paper
proposed:
software_debugging· tags: none9. Looping Is Not Reliability: State-Bound Evidence and Typed Revision Contracts for Agentic Code Repair
Xueping Gao, Jianwei Yang, Qiang Yang · arXiv 2026/07 · paper
proposed:
software_debugging· tags:empirical10. Agent Retrieval Bench: Evaluating Repository Context Retrieval for Coding Agents
Bowen Qin, Yi Xie · arXiv 2026/07 · paper
proposed:
software_debugging· tags:benchmark11. How Do LLMs Read Bug Reports? An Empirical Study of Attention in LLMs for Automated Program Repair
Ramtin Ehsani, Irene Manotas, Saurabh Pujar, et al. · arXiv 2026/07 · paper
proposed:
software_debugging· tags:empirical12. CodeEvo: Interaction-Driven Synthesis of Code-centric Data through Hybrid and Iterative Feedback
Qiushi Sun, Jinyang Gong, Lei Li, et al. · arXiv 2025/07 · paper
proposed:
software_code_generation· tags:training-data13. AssumptionMiner: Extracting, Tracing, and Revising Implicit Assumptions in LLM Code Generation
Jie "JW" Wu · arXiv 2026/07 · paper
proposed:
software_code_generation· tags:benchmark14. Beyond "What to Retrieve": Uncertainty in Retrieval-Augmented Code Generation
Chandan Kumar Sah, Xiaoli Lian, Li Zhang · arXiv 2026/07 · paper
proposed:
software_code_generation· tags: none15. Learning from 53.6K Real-World Developer Edits of AI-Generated Code
Jenny T. Liang, Mihika Bairathi, Wayne Chi, et al. · arXiv 2026/07 · paper
proposed:
software_code_generation· tags:training-databenchmark16. Agent Team Work Zone: An Automated, Persistent Workspace for Long-Lived Coding Agent Teams
Shouren Wang · arXiv 2026/07 · paper
proposed:
software_infrastructure· tags: none17. Specification-Driven DevOps for Multi-Service Environments
Oleg Grynets, Kyrylo Fursov, Vasyl Lyashkevych, et al. · arXiv 2026/07 · paper
proposed:
software_infrastructure· tags:empirical18. JETO-Bench: A Reproducible Benchmark for Execution Time Improvement Patches in Java
Khashayar Etemadi, Zhendong Su · arXiv 2026/06 · paper
proposed:
software_maintenance· tags:benchmark19. ParBench: A Benchmark for Reliable Evaluation of LLM Parallel Code Translation
Samyak Jhaveri, Erel Kaplan, Tom Yotam, et al. · arXiv 2026/06 · paper
proposed:
software_maintenance· tags:benchmark20. Multi-level Code Optimization via Mixture of Prompts
Yun Peng, Jun Wan, Jiakun Liu, et al. · arXiv 2026/07 · paper
proposed:
software_maintenance· tags: none21. Automated Modernization of Machine Learning Engineering Notebooks for Reproducibility
Bihui Jin, Kaiyuan Wang, Pengyu Nie · arXiv 2026/07 · paper
proposed:
software_maintenance· tags: none22. VPR-Evolve: Multi-Agent-Driven Algorithm Evolution for FPGA Place and Route
Qihang Wu, Taizun Jafri, Aman Arora, et al. · arXiv 2026/07 · paper
proposed:
software_maintenance· tags: none23. Reinforcement Learning for Code Optimization
Pierre Chambon, Kunhao Zheng, Juliette Decugis, et al. · arXiv 2026/07 · paper
proposed:
software_maintenance· tags:benchmark24. Execution-Grounded Security Testing for Coding Agents in Software Engineering Pipelines
Yifei Ge, Weisong Sun, Jinkun Xiao, et al. · arXiv 2026/06 · paper
proposed:
software_security· tags: none25. Poster: Rethinking Security in LLM Code Generation through Real-World Risk Scenarios
Lixun Ma, Ruolong Ma, Bei Wang, et al. · arXiv 2026/07 · paper
proposed:
software_security· tags:benchmarkmachine payload (do not edit)
[{"paper": {"id": "2604.02544", "title": "Developer Experience with AI Coding Agents: HTTP Behavioral Signatures in Documentation Portals", "authors": ["Oleksii Borysenko"], "venue": "arXiv 2026/04", "category": "", "published": "2026-04-02", "links": {"paper": "https://arxiv.org/abs/2604.02544", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical study analyzing HTTP request fingerprints from nine AI coding agents and six AI assistant services accessing developer documentation portals, revealing distinct behavioral signatures. It discusses implications for documentation design and analytics given AI-driven access patterns.", "reason": "Object of study is behavior of AI coding agents in the wild, not a proposed method for a task, so it fits the empirical studies leaf.", "source": "crawl"}, {"paper": {"id": "2607.22585", "title": "The Scaffold Effect in Coding Agents: Harness Choice as a Hidden Variable in Coding-Agent Evaluation", "authors": ["Naman Vats", "Oleg Golev"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-08", "links": {"paper": "https://arxiv.org/abs/2607.22585", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical study showing that the choice of agent harness (Goose, OpenCode, OpenHands-SDK) causes up to 40x differences in token cost for coding agents on Terminal-Bench Pro, while pass rates stay similar across harnesses. It argues model name alone is an incomplete unit for coding-agent evaluation and harness-level failure biases must be reported.", "reason": "Object of study is coding-agent evaluation methodology itself (harness vs model effects), an empirical study rather than a proposed agent/method for a task.", "source": "crawl"}, {"paper": {"id": "2607.22807", "title": "The Best Programming Language for Tokenmaxxing: An Investigation of Coding Agent Behavior Across Programming Languages", "authors": ["Zixuan Wu", "Carolyn Jane Anderson", "Arjun Guha"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22807", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "An empirical investigation of coding agent token consumption across Python, Java, Rust, and OCaml, analyzing trajectory structure and text to explain why costs vary by language. Finds agents produce more noncompiling code and distrust tests in unfamiliar languages.", "reason": "This is a behavioral/empirical study of existing coding agents' behavior rather than a new agent, method, or benchmark performing a task.", "source": "crawl"}, {"paper": {"id": "2607.23332", "title": "AlloBench: Measuring Online Tool Allocation Capability in LLM Agents", "authors": ["Daniel Wang", "Andrew Xu"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-25", "links": {"paper": "https://arxiv.org/abs/2607.23332", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical", "benchmark"], "summary": "AlloBench is a paired benchmark testing whether LLM agents rationally allocate effort toward building reusable tools versus one-off scripts, comparing abstract and code-construction settings across frontier models.", "reason": "An empirical/capability study of agent behavior (tool-investment decision-making) rather than a task-performing system, so it fits the off-axis studies leaf.", "source": "crawl"}, {"paper": {"id": "2607.23670", "title": "Plans Work in Mysterious Ways: Evaluating a Plan Mode for Spreadsheet Agents", "authors": ["Aayush Kumar", "Avik Dutta", "Sumit Gulwani", "Gustavo Soares", "Advait Sarkar", "Emerson Murphy-Hill"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-26", "links": {"paper": "https://arxiv.org/abs/2607.23670", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A within-subjects user study (N=24) evaluating a prototype Plan Mode for spreadsheet agents against a non-planning baseline, examining effects on refinement behavior and user perception in end-user programming. Finds Plan Mode improves perceived creativity support and collaboration despite similar task outcomes.", "reason": "An empirical human-agent interaction study of an existing agentic feature, proposing no new task-performing method, per the studies/empirical precedent.", "source": "crawl"}, {"paper": {"id": "2607.24601", "title": "Evaluating the Impact of Explainable AI on Trust in AI-Assisted Code Review", "authors": ["Zhenhan Gao", "Marvin Muñoz Barón", "Umm-e Habiba", "Daniel Graziotin", "Stefan Wagner"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24601", "github": "", "website": ""}}, "category": "studies", "tags": ["empirical"], "summary": "A within-subjects user study of 34 developers examining how explanation level in LLM-based code review systems affects developer trust and agreement with AI recommendations. It studies human-agent interaction around AI code review tools rather than proposing a new review agent.", "reason": "Empirical/behavioral study of human-agent interaction with AI-assisted code review tools fits the studies leaf.", "source": "crawl"}, {"paper": {"id": "2605.17675", "title": "Bridging the Gap on AI-Assisted Scientific Software Development Through Transparency and Traceability", "authors": ["Chaitanya Bhave", "Pierre-Clément A. Simon", "Casey Icenhour", "Lin Yang", "Cody J. Permann", "Daniel Schwen", "Christopher S. Ritter"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2605.17675", "github": "", "website": ""}}, "category": "studies", "tags": ["position"], "summary": "A position paper proposing governance guidance for AI-assisted development of safety-critical scientific software under strict Software Quality Assurance (e.g. NQA-1) standards, using a fusion-energy simulation code as a demonstration case. It focuses on transparency, traceability, and human accountability rather than proposing a new agent or tool.", "reason": "Vision/position paper about the practice of AI-assisted software development, matching the studies leaf's position-paper scope.", "source": "crawl"}, {"paper": {"id": "2607.22711", "title": "CORVUS: Context Optimization and Reduction Via Underlying Synchronization for LLM Coding Agents", "authors": ["Mingwei Zheng", "David OBrien", "Siwei Cui", "Pardis Pashakhanloo", "Rajdeep Mukherjee", "Myeongsoo Kim", "Sachit Kuhar"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-20", "links": {"paper": "https://arxiv.org/abs/2607.22711", "github": "", "website": ""}}, "category": "software_debugging", "tags": [], "summary": "Proposes CORVUS, a trajectory architecture for LLM coding agents that decouples file-read actions from stale observation snapshots, keeping context synchronized with the live codebase. Evaluated on SWE-PolyBench and SWE-Bench Pro, it cuts input tokens and reasoning cycles while preserving pass rates.", "reason": "Evaluated on SWE-bench-style issue-resolution benchmarks, routing to software_debugging per benchmark-routing rule.", "source": "crawl"}, {"paper": {"id": "2607.24604", "title": "Looping Is Not Reliability: State-Bound Evidence and Typed Revision Contracts for Agentic Code Repair", "authors": ["Xueping Gao", "Jianwei Yang", "Qiang Yang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24604", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["empirical"], "summary": "An empirical study of generate-test-revise loops in agentic code repair on HumanEval, showing repeated revision can degrade correctness, and proposing a typed revision contract binding verifier evidence to code states.", "reason": "The served task is program repair/issue resolution reliability, matching software_debugging.", "source": "crawl"}, {"paper": {"id": "2607.24882", "title": "Agent Retrieval Bench: Evaluating Repository Context Retrieval for Coding Agents", "authors": ["Bowen Qin", "Yi Xie"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24882", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["benchmark"], "summary": "Agent Retrieval Bench evaluates repository context retrieval methods that feed patch generation for coding agents, covering tasks like code2test, comment2context, trace2code, and edit2ripple across 427 samples in 25 repositories.", "reason": "Serves the upstream context-acquisition stage of SWE-bench-style issue resolution/patch generation, so it routes to software_debugging plus benchmark tag.", "source": "crawl"}, {"paper": {"id": "2607.25873", "title": "How Do LLMs Read Bug Reports? An Empirical Study of Attention in LLMs for Automated Program Repair", "authors": ["Ramtin Ehsani", "Irene Manotas", "Saurabh Pujar", "Luca Buratti", "Preetha Chatterjee"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25873", "github": "", "website": ""}}, "category": "software_debugging", "tags": ["empirical"], "summary": "An empirical study analyzing attention patterns of LLM-based automated program repair systems over bug reports, finding that diffused attention across diagnostic sections correlates with repair success while over-localized attention correlates with failure.", "reason": "A single-activity (program repair) empirical/failure-analysis study routes to software_debugging plus the empirical tag, per the studies boundary rule.", "source": "crawl"}, {"paper": {"id": "2507.22080", "title": "CodeEvo: Interaction-Driven Synthesis of Code-centric Data through Hybrid and Iterative Feedback", "authors": ["Qiushi Sun", "Jinyang Gong", "Lei Li", "Qipeng Guo", "Fei Yuan"], "venue": "arXiv 2025/07", "category": "", "published": "2025-07-25", "links": {"paper": "https://arxiv.org/abs/2507.22080", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["training-data"], "summary": "CodeEvo uses a dual-agent Coder/Reviewer architecture with hybrid compiler and semantic feedback to iteratively synthesize high-quality instruction-code pairs. The resulting CodeEvo-100K dataset improves fine-tuned models' performance across code generation benchmarks.", "reason": "Resource paper whose downstream task is code generation, so artifact/software_code_generation plus training-data tag.", "source": "crawl"}, {"paper": {"id": "2607.22898", "title": "AssumptionMiner: Extracting, Tracing, and Revising Implicit Assumptions in LLM Code Generation", "authors": ["Jie \"JW\" Wu"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22898", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["benchmark"], "summary": "AssumptionMiner extracts, traces, and revises implicit assumptions LLMs make when generating code from underspecified prompts, producing an explicit assumption layer and enabling AST-guided targeted regeneration. Introduces a 180-task benchmark with 676 annotated assumptions to evaluate extraction, localization, and revision.", "reason": "Serves the code generation task by improving correctness of LLM-generated code from ambiguous specs, with a released benchmark.", "source": "crawl"}, {"paper": {"id": "2607.24884", "title": "Beyond \"What to Retrieve\": Uncertainty in Retrieval-Augmented Code Generation", "authors": ["Chandan Kumar Sah", "Xiaoli Lian", "Li Zhang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24884", "github": "", "website": ""}}, "category": "software_code_generation", "tags": [], "summary": "OpenCoder is a framework that estimates uncertainty in heterogeneous retrieved evidence (similar code, repo context, APIs) to filter and rank it, guiding repository-level code generation, verification, and repair. It improves selected-output correctness on an expanded RepoExec-inline benchmark, though gains are backend-dependent.", "reason": "Task ultimately serves repository-level code generation with an active generate-verify-repair pipeline, not a pure metric.", "source": "crawl"}, {"paper": {"id": "2607.25130", "title": "Learning from 53.6K Real-World Developer Edits of AI-Generated Code", "authors": ["Jenny T. Liang", "Mihika Bairathi", "Wayne Chi", "Ameet Talwalkar", "Nishant Subramani", "Valerie Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.25130", "github": "", "website": ""}}, "category": "software_code_generation", "tags": ["training-data", "benchmark"], "summary": "DECODE is a 53.6K-example dataset of real-world in-IDE developer edits to AI-generated code across Python, TypeScript, and JavaScript, used both to analyze editing behavior and to benchmark/finetune LLMs on code-edit prediction. Finetuned small models substantially outperform frontier LLMs at predicting edits.", "reason": "A resource paper whose downstream task is repo-aware code completion/next-edit prediction, matching software_code_generation's resource-paper scope.", "source": "crawl"}, {"paper": {"id": "2607.22917", "title": "Agent Team Work Zone: An Automated, Persistent Workspace for Long-Lived Coding Agent Teams", "authors": ["Shouren Wang"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-24", "links": {"paper": "https://arxiv.org/abs/2607.22917", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": [], "summary": "ATWZ is a filesystem-based operations layer built around Claude Code's Agent Teams that persists agent working state, skills, and inter-agent messages to enable recovery, resumption, and reduced prompt overhead in long-lived coding agent teams.", "reason": "Enabling tooling that preserves and restores agent operational state in service of sustained software development work, matching software_infrastructure's scope of enabling code production.", "source": "crawl"}, {"paper": {"id": "2607.25141", "title": "Specification-Driven DevOps for Multi-Service Environments", "authors": ["Oleg Grynets", "Kyrylo Fursov", "Vasyl Lyashkevych", "Volodymyr Veres"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.25141", "github": "", "website": ""}}, "category": "software_infrastructure", "tags": ["empirical"], "summary": "This study evaluates whether a frontier LLM can generate functionally correct Dockerfiles and Docker Compose configurations for multi-service repositories without developer-authored deployment artifacts, finding functional success but gaps in deployment-intent fidelity (e.g., missing network segmentation, multi-stage builds). It proposes a minimal deployment specification to close this gap.", "reason": "LLM-generated build/deployment configuration for running code is environment-setup/CI-CD enabling work under software_infrastructure.", "source": "crawl"}, {"paper": {"id": "2606.31767", "title": "JETO-Bench: A Reproducible Benchmark for Execution Time Improvement Patches in Java", "authors": ["Khashayar Etemadi", "Zhendong Su"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-30", "links": {"paper": "https://arxiv.org/abs/2606.31767", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["benchmark"], "summary": "Presents JETO-Mine, a configurable pipeline to mine reproducible execution-time-improvement patch (ETIP) benchmarks from Java repositories, and JETO-Bench, a benchmark of 660 mined and 91 verified ETIPs. Evaluates OpenHands with GPT-5-mini, which fixes 14.3% of issues.", "reason": "Serves performance optimization of existing code while preserving behavior, matching software_maintenance's boundary definition.", "source": "crawl"}, {"paper": {"id": "2607.22588", "title": "ParBench: A Benchmark for Reliable Evaluation of LLM Parallel Code Translation", "authors": ["Samyak Jhaveri", "Erel Kaplan", "Tom Yotam", "Le Chen", "Tomer Bitan", "Niranjan Hasabnis", "Gal Oren"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-09", "links": {"paper": "https://arxiv.org/abs/2607.22588", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["benchmark"], "summary": "ParBench is a kernel-centric benchmark for evaluating LLM-based translation of parallel code kernels across CUDA, OpenMP, OpenCL, and OpenMP target offload, testing behavioral fidelity and robustness to source perturbations. It reveals persistent weaknesses in cross-API translation, multi-file coordination, and robustness.", "reason": "Behavior-preserving code translation across APIs/languages matches software_maintenance's migration boundary.", "source": "crawl"}, {"paper": {"id": "2607.23665", "title": "Multi-level Code Optimization via Mixture of Prompts", "authors": ["Yun Peng", "Jun Wan", "Jiakun Liu", "Shuzheng Gao", "David Lo", "Xiaoxue Ren"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-26", "links": {"paper": "https://arxiv.org/abs/2607.23665", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "Optimo is an LLM-based multi-level code optimization approach using a Mixture-of-Prompts architecture that identifies performance bottlenecks via differential profiling and applies tailored optimization strategies across four abstraction levels. It achieves up to 57.48% opt% and 13.51x speedups on efficiency benchmarks COFFE and Effibench.", "reason": "Performance optimization of existing code while preserving behavior falls under software_maintenance.", "source": "crawl"}, {"paper": {"id": "2602.07195", "title": "Automated Modernization of Machine Learning Engineering Notebooks for Reproducibility", "authors": ["Bihui Jin", "Kaiyuan Wang", "Pengyu Nie"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2602.07195", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "MLEModernizer is an LLM-driven agentic framework that iteratively executes, diagnoses, and repairs Jupyter notebooks to restore reproducibility as ML software and hardware ecosystems evolve. Evaluated on over 8,000 non-reproducible notebooks, it successfully modernizes roughly 40-45 percent of them.", "reason": "Behavior-preserving adaptation of code to new dependency and environment versions is software_maintenance.", "source": "crawl"}, {"paper": {"id": "2607.24998", "title": "VPR-Evolve: Multi-Agent-Driven Algorithm Evolution for FPGA Place and Route", "authors": ["Qihang Wu", "Taizun Jafri", "Aman Arora", "Vidya A. Chhabria"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-27", "links": {"paper": "https://arxiv.org/abs/2607.24998", "github": "", "website": ""}}, "category": "software_maintenance", "tags": [], "summary": "VPR-Evolve uses LLM agents to propose, implement, and evaluate source-code modifications to the VPR FPGA place-and-route engine, guided by shared memory of prior outcomes. It improves critical-path delay, wirelength, and runtime over stock VPR and hyperparameter-tuning baselines on VTR-9 benchmarks.", "reason": "Agents evolve an existing CAD tool's source code to optimize performance metrics while preserving functionality, matching maintenance/performance-optimization.", "source": "crawl"}, {"paper": {"id": "2607.25970", "title": "Reinforcement Learning for Code Optimization", "authors": ["Pierre Chambon", "Kunhao Zheng", "Juliette Decugis", "Benoit Sagot", "Gabriel Synnaeve"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-28", "links": {"paper": "https://arxiv.org/abs/2607.25970", "github": "", "website": ""}}, "category": "software_maintenance", "tags": ["benchmark"], "summary": "Trains RL agents to optimize code execution speed while preserving correctness, introducing DMC-Optim with a calibrated timing sandbox and reward design to overcome noise in timed-execution RL. Shows large pass@1 gains in performance-oriented code optimization over standard RLVR.", "reason": "behavior-preserving performance optimization of existing code matches software_maintenance", "source": "crawl"}, {"paper": {"id": "2607.22569", "title": "Execution-Grounded Security Testing for Coding Agents in Software Engineering Pipelines", "authors": ["Yifei Ge", "Weisong Sun", "Jinkun Xiao", "Yuchen Chen", "Yebo Feng", "Peizhuo Lv", "Xia Feng", "Chunrong Fang", "Zhihong Zhao", "Zhenyu Chen", "Yang Liu"], "venue": "arXiv 2026/06", "category": "", "published": "2026-06-01", "links": {"paper": "https://arxiv.org/abs/2607.22569", "github": "", "website": ""}}, "category": "software_security", "tags": [], "summary": "Presents an execution-grounded red-team testing framework that embeds unsafe system operations inside routine software-engineering tasks to probe whether coding agents can be induced to perform unauthorized system-level actions. Experiments show high rates of verified unsafe execution across agent frameworks and models.", "reason": "A security-assurance testing framework targeting coding agents' unsafe actions during SE pipeline tasks fits software_security's security-testing scope.", "source": "crawl"}, {"paper": {"id": "2607.23088", "title": "Poster: Rethinking Security in LLM Code Generation through Real-World Risk Scenarios", "authors": ["Lixun Ma", "Ruolong Ma", "Bei Wang", "Feng Wei", "Zhenguang Liu", "Lorenzo Cavallaro", "Wentao Chen"], "venue": "arXiv 2026/07", "category": "", "published": "2026-07-25", "links": {"paper": "https://arxiv.org/abs/2607.23088", "github": "", "website": ""}}, "category": "software_security", "tags": ["benchmark"], "summary": "Introduces a 2,700-case benchmark of realistic ambiguous/under-specified prompts to evaluate security vulnerabilities in LLM-generated code, finding over 56% average vulnerability rates across eight LLMs. Shows security-aware prompting can mitigate risks by up to 45%.", "reason": "A benchmark resource serving secure code generation, routed to software_security per code-generation security scope.", "source": "crawl"}]