Verified, then logged.
Sun 22 Nov 2026 · 00:16 KST / 15:16 UTC한국어 | ENSearch /

Breaking01 / AI

Natural safety responses improve retention after training

Researchers propose NEST, a training approach that uses context-aware safety responses instead of formulaic refusals.

Seen from behind in an AI research lab, an anonymous researcher compares repetitive and context-dependent language-model responses on two monitors with unreadable screen content. AI-generated image

Researchers reported on 10 Nov in the preprint When Safety Becomes An Outlier: Understanding the Retention of LLM Safety Behaviors that comparisons of refusal responses found safety improvements; the paper is not peer-reviewed, and its abstract presents no independent replication results.

The authors argue that formulaic refusals can be easy to learn yet vulnerable to later model updates. In controlled experiments, they reported that models learned fixed refusals much like arbitrary constant strings, unlike natural responses grounded in the request.

Their proposed approach, Naturalness-Enhanced Safety Targets (NEST), constructs natural safety responses suited to each request. It uses perplexity—a measure of how difficult a response is for a model to predict—calculated with the original model.

According to the abstract, the researchers applied NEST to supervised training, which learns from example answers, and preference-based training, which learns from preferences between responses. They reported better initial safety and better retention of safe behaviour after further training, without sacrificing helpfulness (sample sizes not reported; multiple models and training approaches).

The authors explain that earlier research already examined how further training can weaken safe behaviour even when safety data are included. This study instead examines the format of refusal responses and their connection to the meaning of the request as a possible cause.

The abstract does not report the size of the improvements for individual models or the evaluation sample sizes.