Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like Benchmarks

Source

arXiv

arxiv.org/abs/2607.28587v1

Read the source

Source

Plain-Language Brief

The paper proposes a system called PAIChecker to check for misalignment between Pull Requests (PR) and Issues in SWE-bench-like benchmarks used for evaluating Large Language Models' issue resolution capability. The system identifies patterns of misalignment, synthesizes labels across multiple agents, and validates at the code level. Experiments show that PAIChecker outperforms other LLM backbones on SWE-Gym and SWE-bench Multilingual.

What the Document Says

  • SWE-bench-like benchmarks are used for evaluating Large Language Models' (LLM) issue resolution capability.
  • The benchmarks follow a common construction pipeline: each PR is paired with its linked issue, the issue description serves as the problem statement, and the PR patch serves as the test oracle.
  • Despite common practice, PR-Issue pairings are often misaligned in large repositories due to complexity.
  • The paper studies SWE-bench Verified instances and finds 13.6% exhibit misalignment across five patterns in eleven fine-grained scenarios.
  • To ensure reliable and scalable construction of such benchmarks, the authors propose PAIChecker, a multi-agent system for checking PR-Issue misalignment.
  • PAIChecker adopts a three-phase design: pattern identification, cross-agent label synthesis, and code-level validation.
  • The system is designed to increase accuracy, generalizability, and progressively verified detection.
  • Experiments on SWE-Gym and SWE-bench Multilingual show that PAIChecker outperforms other LLM backbones, achieving up to 92.12% and 91.67% binary accuracy respectively.

Key Concepts Mentioned

  • SWE-bench-like benchmarks
  • PR-Issue misalignment
  • Large Language Models (LLM)
  • Pattern identification
  • Cross-agent label synthesis
  • Code-level validation
  • SWE-Gym
  • SWE-bench Multilingual

Claims Made by the Document

  • 13.6% of SWE-bench Verified instances exhibit misalignment
  • PAIChecker achieves the best performance across all four LLM backbones

Evidence Used

  • The paper studies SWE-bench Verified instances and finds 13.6% exhibit misalignment (in what_it_says)
  • Experiments on SWE-Gym and SWE-bench Multilingual show that PAIchecker achieves the best performance (in what_it_says)

Methods / Approach

  • Pattern identification
  • Cross-agent label synthesis
  • Code-level validation

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

SWE-bench-like benchmarks are widely used for evaluating LLM's issue resolution capability. They typically follow a common construction pipeline: each PR (Pull Request) is paired with its linked issue by extracting issue references from the PR description; the issue description is used as the problem statement, and the PR patch serves as the test oracle. However, due to the inherent complexity of developing and maintaining large repositories, such PR-Issue pairings are often misaligned in practice. In this work, we systematically study SWE-bench Verified instances, finding that 13.6% exhibit misalignment across five patterns in eleven fine-grained scenarios. To enable reliable and scalable construction of those benchmarks in the future, we propose PAIChecker, a multi-agent system for checking PR-Issue misalignment in SWE-bench-like benchmarks. Specifically, PAIChecker adopts a three-phase design that combines specific pattern identification, cross-agent label synthesis, and code-level validation, thereby enabling more accurate, generalizable, and progressively verified detection. Experiments on SWE-Gym and SWE-bench Multilingual show that PAIchecker achieves the best performance across all four LLM backbones, reaching up to 92.12% and 91.67% binary accuracy, respectively.