Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 28·4 min read
CRaFT: Circuit-Guided Refusal Feature Selection via Cross-Layer Transcoders
TL;DR
Researchers identified a gap in LLM evaluation benchmarks. They built a synthetic dataset with 10k adversarial prompts targeting reasoning failures.
✦ Why It Matters
Use this benchmark to audit LLM robustness before deploying in production reasoning pipelines.
Key Takeaways
- Utilize CRaFT for better feature selection in LLMs.
- Focus on inter-feature relationships for deeper insights.
- Evaluate model performance against jailbreak benchmarks.
- Enhance safety measures in AI applications.
How It Works
CRaFT employs cross-layer transcoders to create a sparse feature circuit graph that illustrates how different features interact and influence the model's refusal decisions. By quantifying these inter-feature influences, CRaFT can effectively rank the most impactful features that contribute to the model's ability to refuse harmful requests.