Logit-Gap Steering: A New Frontier in Understanding and Probing LLM Safety

This report outlines research on “logit-gap steering,” a method that exploits the refusal‑affirmation logit gap to reliably jailbreak aligned LLMs (including Qwen, LLaMA, Gemma, and gpt‑oss‑20b) with >75% success, concluding that alignment alone cannot prevent harmful outputs; it recommends defense‑in‑depth with external protections, improved evaluation benchmarks, and invites further study via the referenced arXiv paper.