PaperCompiler is a framework that translates research papers into repository-level code by compiling paper-grounded evidence into explicit implementation specifications. It preserves source provenance, distinguishes between paper-supported, inferred, externally delegated, and unresolved information, and encodes non-degradation requirements, ownership, cross-file dependencies, and file-level constraints. The approach outperforms baselines on Paper2CodeBench, improving reference-based fidelity by 13.8% and halving high-severity evaluator critiques.
PaperCompiler is a framework that translates research papers into repository-level code by compiling paper-grounded evidence into explicit implementation specifications. It preserves source provenance, distinguishes between paper-supported, inferred, externally delegated, and unresolved information, and encodes requirements such as non-degradation, ownership, cross-file dependencies, and file-level constraints. The framework improves reference-based fidelity by 13.8% and cuts high-severity evaluator critiques from 13.2% to 6.1% on Paper2CodeBench.
By Yunhao Liu, Hong Phuc Pham, Jaehong Yoon
arXiv:2608. 19799v1 Announce Type: new Abstract: Software increasingly functions as part of the scientific instrument itself, making failures in scientific code capable of compromising not only program behavior but also the evidence underlying scientific conclusions.
By Zhipeng Xu, Jiahao Lu, Yining Zheng, Yuxin Wang, Xipeng Qiu
SkillScriptBench is a 350‑task benchmark that evaluates how well large language models can self‑evolve executable agent skill packages by distinguishing between documentation repair, script repair, and preservation. The benchmark draws from 100 real GitHub skill packages, creating 150 repair tasks with injected script faults and a controlled track of 200 tasks that test clean, documentation‑faulty, script‑faulty, and combined‑faulty states. Experiments show that AST‑Guided Skill Revision—using abstract syntax trees to target relevant code locations—improves repair success by up to 27.7% and yields more consistent outcomes across repeated runs.
whyItMatters":"The benchmark provides a systematic way to measure and improve the reliability of automated skill package maintenance, which is essential for deploying robust LLM agents in real‑world applications."
By Yuxuan Liu, Haoran Li, Yuhao Zhang, Jiahe Guo, Hongyu Luo, Wenbin Hu, Huihao Jing, Kawai Chung, Junle Chen, Changxuan Fan, Qing Zong, Lingyun Xie, Yangqiu Song
SWE‑bench Science is a benchmark for evaluating coding agents on scientific software engineering tasks, comprising 119 tasks from 98 GitHub repositories across 20 scientific domains. The tasks are grouped into Issue‑driven, Expert‑exploratory, and Engineering‑integration paradigms, and even the best agent, Claude Code with Opus‑5 (max), achieves a pass@1 below 50%. The study identifies four common failure mechanisms—lack of scientific knowledge, misguided exploration, incomplete repair coverage, and poor generalization—and shows that providing scientific guidance can both help and hinder repair depending on its alignment with the task.
The paper explores whether natural‑language documentation aids coding agents in fixing software bugs and introduces a roundtrip benchmark that evaluates code descriptions by regenerating code and testing it. It finds that description completeness, not length, determines fidelity, and presents an optimizer that can produce fully faithful descriptions that generalize to new files. However, experiments across two model families and ten repositories show that such compact documentation does not improve an agent’s ability to resolve real repository issues compared to using the issue alone.
By Md Shohel Arman, Igor Molybog