AIAny
Icon for item

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

Evaluates whether coding agents can modify real scientific software while preserving domain-specific scientific contracts. Contains 119 repository-level tasks across 98 GitHub projects and 20 scientific domains, measures reproducible edits in pinned Docker images, and analyzes recurring failure modes.

Introduction

Scientific software encodes domain knowledge and data semantics that make simple patch-based evaluation misleading: a passing unit test can still break scientific contracts. SWE-bench Science surfaces this gap by forcing coding agents to edit real repositories while preserving units, file formats, numerical invariants, and end-to-end behavior.

Key Findings
  • Benchmark scope and setup: 119 tasks drawn from 98 GitHub repositories spanning 20 scientific domains, organized into three paradigms (Issue-driven, Expert-exploratory, Engineering-integration). Each task ships a baseline environment image and a separate verifier image; images are immutable and pinned by digest for reproducible evaluation.
  • Performance and difficulty: Even the best agent evaluated (Claude Code with Opus-5 (max)) achieves pass@1 below 50%, showing that repository-level scientific repairs remain challenging for current coding agents.
  • Failure mechanisms: Four recurring modes are identified — lack of scientific abstraction/knowledge, surface-level or misguided exploration, incomplete system-level integration or coverage, and failures to generalize scientific reasoning beyond observed examples.
  • Ablation insight: Removing explicit scientific guidance (while preserving repository and execution context) yields mixed effects — accurate, well-aligned scientific guidance can constrain search and improve efficiency, but misaligned guidance can anchor agents and hurt exact-repair success.
Who it's for and tradeoffs

Great fit if you research coding-agent capabilities, evaluate LLM-driven code repair in domain-rich settings, or need a repository-level benchmark that stresses scientific correctness beyond unit tests. Look elsewhere if you need micro-level function-completion tasks or synthetic toy benchmarks; SWE-bench Science emphasizes end-to-end engineering context, reproducible images, and held-out verifiers rather than isolated function synthesis. The benchmark favors experiments that integrate execution, pinned environments (Docker images), and domain-aware verification over purely prompt-based code generation.

Information

  • Websitearxiv.org
  • AuthorsZhipeng Xu, Jiahao Lu, Yining Zheng, Yuxin Wang, Xipeng Qiu
  • Published date2026/08/20

Categories

More Items

Proposes FACET, a framework that synthesizes verifiable terminal tasks by reconstructing scenario intent and grounding instruction, solution, and verifier in a shared executable container state. Key features include environment-first generation, execution-based validation, and targeted repair to preserve source intent and cross-artifact consistency.

Wraps static, hand-built environments with a programmable plug-in harness that reshapes environment behavior without changing underlying logic. EnvRigger automates diagnosis and synthesis of harness components from agent failure trajectories, validating edits via fresh rollouts to improve agent success and efficiency.

Conducts end-to-end multidisciplinary research directly from heterogeneous raw evidence using lifecycle-wide perception and three autonomous agents (Ideation, Experiment, Writeup). Integrates perceptual analysis, execution provenance, and code-enforced checks to produce executable analyses, validated results, and compiled manuscripts across many modalities.