SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?

Hugging Face Daily Papers Papers

Summary

SWE-bench Science introduces a repository-level benchmark for evaluating coding agents on scientific software repair tasks, revealing failure mechanisms and mixed effects of scientific guidance.

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. Yet existing evaluations of coding agents largely emphasize aggregate task success, providing limited insight into why agents fail when repairing scientific software. We introduce SWE-bench Science, a repository-level benchmark for scientific software engineering comprising 119 tasks from 98 GitHub repositories across 20 scientific domains. Each task is organized into one of three paradigms: Issue-driven, Expert-exploratory, and Engineering-integration. Even the best-performing agent, Claude Code with Opus-5 (max), achieves a pass@1 below 50\%, highlighting the substantial challenges posed by scientific software engineering. We identify four recurring failure mechanisms: deficits in scientific knowledge or abstraction, misguided exploration or surface-level repair, incomplete repair coverage or system integration, and failures to generalize scientific knowledge beyond observed cases in our analysis. We further conduct a paired ablation that removes explicit scientific guidance while preserving the repository and executable engineering context. The results show that scientific knowledge is not uniformly beneficial: well-grounded information can constrain repair and improve average performance and token efficiency, whereas poorly aligned guidance can induce anchoring and does not necessarily improve exact repair success. Together, SWE-bench Science provides a broad testbed for studying both the capabilities and failure mechanisms of coding agents in scientific software engineering.
Original Article
View Cached Full Text

Cached at: 08/21/26, 08:08 AM

Paper page - SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?

Source: https://huggingface.co/papers/2608.19799

Abstract

SWE-bench Science benchmarks coding agents on scientific software repair, revealing failure mechanisms and mixed effects of scientific guidance.

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. Yet existing evaluations ofcoding agentslargely emphasize aggregate task success, providing limited insight into why agents fail when repairing scientific software. We introduceSWE-bench Science, a repository-level benchmark forscientific software engineeringcomprising 119 tasks from 98 GitHub repositories across 20 scientific domains. Each task is organized into one of three paradigms: Issue-driven, Expert-exploratory, and Engineering-integration. Even the best-performing agent, Claude Code with Opus-5 (max), achieves apass@1below 50\%, highlighting the substantial challenges posed byscientific software engineering. We identify four recurringfailure mechanisms: deficits inscientific knowledgeor abstraction, misguided exploration or surface-level repair, incomplete repair coverage or system integration, and failures to generalizescientific knowledgebeyond observed cases in our analysis. We further conduct a pairedablationthat removes explicit scientific guidance while preserving the repository and executable engineering context. The results show thatscientific knowledgeis not uniformly beneficial: well-grounded information can constrain repair and improve average performance and token efficiency, whereas poorly aligned guidance can induce anchoring and does not necessarily improve exact repair success. Together,SWE-bench Scienceprovides a broad testbed for studying both the capabilities andfailure mechanismsofcoding agentsinscientific software engineering.

View arXiv pageView PDFProject pageGitHub13Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.19799 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.19799 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.19799 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

SWE-INTERACT: Reimagining SWE Benchmarks as User-Driven Long-Horizon Coding Sessions

Hugging Face Daily Papers

SWE-Interact is a new testbed that evaluates coding agents in realistic multi-turn, user-driven software engineering tasks, revealing that strong single-turn benchmark performance does not reliably transfer to interactive, iterative workflows where agents must discover user intent and adapt to evolving requirements.

SWE-Touch: Benchmarking Coding Agents When Users Touch the Code

Hugging Face Daily Papers

Introduces SWE-Touch, a benchmark framework that injects conflicting user edits during agent coding trajectories, showing that current coding agents significantly degrade in collaborative settings despite strong standalone benchmark performance.