AIAny
Icon for item

SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring

Provides a curated benchmark of 170 real-world, multilingual code-refactoring instances to evaluate AI coding agents on large-scale, behavior-preserving, cross-file refactors. Each task includes rewritten issue descriptions and manually reviewed test suites to avoid over- and under-constraining evaluations.

Introduction

Most existing code-evaluation suites emphasize single-file fixes or short edits that can be overfit or validated by flawed tests. SWE-Bench ProMax reframes the problem around refactoring: coordinated, behavior-preserving changes across many files and languages that better reflect professional, long-horizon engineering work.

Key Findings
  • SWE-Bench ProMax contains 170 human-curated refactoring instances drawn from real commits across seven languages (Python, Java, TypeScript, Go, C, C++, Rust). Instances average 11.4 modified files and 261.6 lines changed, substantially larger than prior benchmarks.
  • Curation targets known benchmark failure modes: issue descriptions are rewritten from scratch for clarity, test suites are manually reviewed to remove overly narrow or overly broad checks, and trivial or single-file tasks are filtered out to preserve difficulty and realism.
  • Empirical evaluation shows frontier models under common agent scaffolds achieve a peak resolve rate of 41.2%, confirming the benchmark is substantially unsaturated and stresses long-horizon coordination and behavior-preservation capabilities.
Who it's for and trade-offs

Great fit if you are evaluating or developing AI coding agents, autonomous refactoring tools, or benchmarks that measure cross-file, behavior-preserving edits across multiple programming languages. It highlights real-world failure modes (test quality, specification ambiguity, large-scope edits) that short bug-fix benchmarks often miss. Look elsewhere if you need micro-benchmarks focused on single-line fixes, unit-level algorithmic problems, or extremely high-volume automated scoring: ProMax prioritizes curation and realism over sheer instance count and may require longer evaluation cycles per instance.

Information

  • Websitearxiv.org
  • AuthorsYuling Shi, Jinghan Xu, Kelin Fu, Wenhao Zeng, Shilin He, Lei Zhang, Yue Liu, Zelin Zhao, Terry Yue Zhuo, Jialun Cao
  • Published date2026/08/10

More Items

Enables a coding agent to self-develop by evolving its harness, prompts, tools, and core code via reviewed commits — supporting recursive free evolution and experience-driven evolution. Demonstrated a 161-day live lineage and state-of-the-art scores on multiple coding benchmarks while foregrounding operational safety.

Transfers hierarchical, structured memory from a large teacher agent into small LLM agents to improve tool-use success. Constructs Workflow, Subtask and Function memories with proactive/reactive injection; training-free and validated on multiple tool-use benchmarks.

A research report proposing a continual-learning agent workflow that pairs recursive self-improvement with a Mixture-of-LoRA design: freeze a foundation model, compose specialist LoRA adapters routed per user turn, and support them with long-context RL and post-training infrastructure.