Get Started
Home
Topics
Search
Library
Research questionHow can coding agents repair scientific software when domain guidance may mislead them?Scientific software failures can alter both program behavior and the evidence supporting scientific conclusions. Agents may miss domain abstractions, make surface-level repairs, or fail to integrate and generalize changes, while supplied guidance can either constrain repair or cause anchoring.
AI
AI Agents
Code Generation & Program Synthesis
Evaluation & Benchmarks
Latest papersRecent research connected to this question, newest first.SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?Evidence comes from SWE-bench Science: 119 repository-level tasks from 98 GitHub repositories across 20 scientific domains, covering issue-driven, expert-exploratory, and engineering-integration tasks. The benchmark reports pass@1 below 50% for the best-performing evaluated agent and uses an ablation removing explicit scientific guidance while preserving repository and executable engineering context; it identifies recurring failure mechanisms involving scientific knowledge, exploration, repair coverage, integration, and generalization.research paper · Sep 1, 2026
Related questions
How can repository-level coding-agent benchmarks detect review-constraint failures beyond passing functional tests?How can scientific agents choose domain-specific procedures that make analyses defensible?How can coding agents maintain accurate repository context as codebases evolve without repeated discovery?How can coding agents reliably implement systems-level requirements and detect the defects they introduce?