ReSI: Recursive Safety Improvement toward Resistant and Resilient AI
ReSI recursively red-teams and retrains models, cutting mean X-Teaming success from 86.01% to 31.45%.
ReSI is a recursive safety-improvement framework that repeatedly red-teams a target model, develops training recipes, and promotes the update with the largest safety gain that still passes a capability-retention Pareto gate. Across four dense and mixture-of-experts models it matches or exceeds evaluated frontier models on in-distribution and out-of-distribution safety benchmarks and beats alignment baselines on nearly all safety tests while largely keeping general capability. Mean X-Teaming attack success across the four models drops from 86.01% to 31.45%, below GPT-5.6-Luna's reported 56.69%.
- Each round red-teams the model and promotes the best capability-safe update.
- Matches or beats frontier models on in- and out-of-distribution safety benchmarks.
- Mean X-Teaming attack success falls from 86.01% to 31.45% across four models.
- That result is below GPT-5.6-Luna's reported 56.69% attack success rate.
- General capabilities are largely preserved while safety improves.
From the original article
1 min read at arxiv.orgRecursive self-improvement, the participation of AI systems in improving their own capabilities, is beginning to move from theoretical prospect to practice, posing both challenges and opportunities for safety alignment. Models evolve through frequent updates, and their safety alignment requires continual adaptation to each new checkpoint.
By Jingnan Zheng · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.