Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 26·5 min read
Guarded Repair for Harm-Aware Post-hoc Replacement of LLM Mathematical Reasoning
TL;DR
Large Language Models (LLMs) can mistakenly replace correct mathematical reasoning with incorrect solutions, posing a risk. GuardedRepair is a framework that selectively repairs reasoning traces, ensuring replacements are only made when verified as safer.
✦ Why It Matters
Engineers can implement GuardedRepair to enhance the reliability of LLMs in mathematical tasks.
Key Takeaways
- Implement GuardedRepair for safer LLM reasoning corrections.
- Use lightweight symbolic checks to evaluate repair safety.
- Focus on selective replacement to avoid introducing new errors.
- Analyze the fixed/broken tradeoff in LLM outputs.
Full Summary
Large Language Models (LLMs) often face challenges in mathematical reasoning, where incorrect answers can be mistakenly corrected, leading to further errors. GuardedRepair is a new framework designed to address this issue by implementing a selective replacement strategy for reasoning traces.
It employs lightweight symbolic checks and surface semantic-risk diagnostics to evaluate whether a repair is safer than keeping the original answer. In tests on the GSM8K dataset, GuardedRepair increased accuracy from 95.60% to 96.89%, successfully correcting 17 out of 58 errors without introducing new mistakes.
Additionally, in a weaker reasoning scenario, accuracy improved from 78.40% to 87.60%. The findings indicate that this method significantly enhances the balance between fixing errors and maintaining correct answers, suggesting a more cautious approach to post-hoc repairs in LLMs.