ReSI: Recursive Safety Improvement toward Resistant and Resilient AI
Abstract Overview
ReSI is a recursive safety improvement framework that treats model safety alignment as an iterative loop of red-teaming, automated recipe search, and validated model updates. In each round, it identifies vulnerabilities in the current model using multiple red-teaming methods, screens candidate alignment recipes with pilot runs, and advances only updates that improve safety while meeting capability-retention criteria. The framework searches over attack-data sources, alignment methods, training mixtures, and hyperparameters, with a Pareto gate enforcing safety gains alongside retained instruction following and benign compliance. Across four dense and mixture-of-experts models, the paper reports improved in-distribution safety, stronger robustness to unseen attack mechanisms, and mostly preserved general capabilities.
Novelty
The distinctive contribution is framing safety alignment itself as a bounded recursive self-improvement process, operationalized through automated research rather than a fixed training recipe. ReSI also combines iterative red-teaming with a structured search over alignment methods and recipe components, while using a fixed Pareto gate to prevent safety gains from propagating if they degrade utility too much.
Results
Across four evaluated models, ReSI reduces attack success rates on both in-distribution and out-of-distribution safety benchmarks, often outperforming the A3 baseline and matching or exceeding MAGIC in most reported comparisons. The paper highlights a reduction in mean X-Teaming attack success rate from 86.01% to 31.45% across the four models, with all four ReSI-aligned models reported below the evaluated frontier-model range on that benchmark. Reported capability retention is generally strong, with benign compliance improving for some models and instruction-following accuracy remaining close to backbone levels, though some capability losses appear for Ministral.
Key Points
- ReSI organizes safety improvement into repeated rounds of red-teaming assessment, pilot screening, full-scale refinement, and gated model updating.
- The framework explicitly targets both resistance to known attacks and resilience to unseen attacks by training on discovered vulnerabilities and testing on held-out attack mechanisms such as H-CoT and X-Teaming.
- Empirically, ReSI improves safety across four model families while largely preserving benign compliance and general capabilities under its Pareto-based selection rule.