Search papers, labs, and topics across Lattice.
This paper introduces ResponseGuard, a vision-language guard that efficiently detects harmful responses in real-time without relying on step-by-step reasoning. By utilizing a single pooled representation of the request, response, and image, ResponseGuard achieves superior performance on harmfulness detection compared to a more complex reasoning-based guard, while being approximately 150 times faster. The findings suggest that a simplified approach can effectively mitigate risks in AI-generated content, challenging the necessity of reasoning chains in safety mechanisms.
ResponseGuard outperforms traditional reasoning-based guards in harmfulness detection while operating at a fraction of the time cost, reshaping our understanding of safety in AI responses.
A vision-language AI assistant returns its answer as a stream of generated tokens. Therefore, a safety guard that watches that answer has to keep up with the stream and stop a harmful reply before a user reads it. Recent vision-language guardrails instead generate a chain of thought before they issue a verdict. They believe that step-by-step reasoning yields a safer guard. This design makes the guard heavy and slow, since the model must decode many tokens for harmfulness detection. We pose the question of whether a vision-language guard really needs to reason in order to screen a response. We answer with a guard that has no chain. ResponseGuard reads a harmful verdict from a single pooled representation of the request, the response, and the image in one forward pass. Across a standard multimodal guardrail benchmark, our 2B ResponseGuard outperforms a recent 3B reasoning-based vision-language guard on response harmfulness detection, without any reasoning and at about 150 times lower time cost. On request harmfulness the reasoning guard retains an overall lead, and the remaining gap on both tracks sits on the image-only cells. We observe that the gap may stem from the frozen vision encoders that both designs use rather than from the missing chain. We have also found the reasoning guard directs almost none of its verdict attention to the image. Based on a single-pass detection, ResponseGuard can screen an answer sentence by sentence as it streams and stop a harmful answer before it finishes. For guarding the response of a vision-language model, a calibrated single-pass label may provide a sufficient safety signal. We fully release all source code, trained models, and datasets at https://github.com/ndb796/ResponseGuard.
