Breaking01 / AI
SPECTRA proposes restoring model refusals without retraining
Researchers report that transferring a base model’s refusal direction improves safety while preserving specialist knowledge.

Researchers behind the SPECTRA paper, which has not been peer-reviewed, reported on 10 Nov that safety evaluations showed restored refusal behaviour in fine-tuned language models; the abstract does not report whether the work has been independently replicated.
The researchers describe transferring a base model’s refusal direction—a pattern associated with rejecting harmful requests—into the weights of a model further trained for a specialist domain. The method uses a low-rank update, which changes only a limited set of directions in the model’s weights, informed by its internal responses.
The authors say the update targets the space most directly involved in responses to harmful inputs. They describe this restricted intervention as designed to preserve responses to ordinary requests.
In safety evaluations, the researchers report improved refusal behaviour while retaining general response coherence and specialist knowledge. They also report negligible changes in the rate at which the model incorrectly rejects benign requests.
The authors contrast SPECTRA with preference optimisation, which they describe as computationally expensive and dependent on domain-specific preference data that can be difficult to obtain. SPECTRA is designed to work without gradient-based optimisation—the process of calculating how to adjust model weights—or domain-specific safety data.
The abstract does not identify the models evaluated or disclose the sample size, leaving the scope of the reported results unspecified.