Breaking01 / AI
Safety score calibration cuts harmful responses in VLM test
Researchers report fewer harmful responses after training a vision-language model with calibrated scores and limits on harmful assistance and unnecessary refusals.

Researchers reported on 13 November in Safety as a Budget, a preprint that has not been peer-reviewed, that score calibration reduced harmful responses to 4.8% (n=1,037, calibrated method) in a comparative evaluation; the paper reports no independent replication.
The paper describes training Qwen3-VL-2B, a vision-language model that handles images and text, with limits on the rates of harmful responses and unnecessary refusals. Scores were calibrated using evaluators' judgements of training responses, then incorporated into the training constraints.
The researchers reported a harmful response rate of 8.9% (n=1,037, standard GRPO), compared with 4.8% (n=1,037, calibrated method). GRPO is a reinforcement learning method that compares response scores within a group. Harmful response rates fell across all experimental seeds, which control randomness in the runs.
The paper reported unnecessary refusals on benign questions at no more than 2% (n=400, calibrated method). The researchers also reported that the results held under judgements from other evaluators and human review.
The paper argues that standard GRPO cancels out the value controlling constraint strength when it standardises rewards within each group. The proposed method retains that value in the learning signal, allowing it to control how strongly training discourages harmful responses.
The abstract reports neither independent replication nor performance with other policy models—the models being trained to generate responses.