From Routing Signals to Selective Review: Visual regrounding in MoE VLMs
MoE routing signals can flag missing visual targets before a VLM starts answering.
The paper tests whether expert-routing probabilities inside MoE vision-language models reveal when a question assumes an object that is not present. Its detectors use routing from target tokens in Qwen3-VL-30B-A3B-Instruct and Gemma-4-26B-A4B-it, then trigger a target-aware review prompt only when needed. On GQA-Inpaint, routing alone reaches ROC-AUCs of 0.9988 for Qwen and 0.9956 for Gemma, with weaker but still useful transfer to OBER. The authors say threshold shifts across datasets need recalibration, but false-positive review caused limited harm overall. ArXiv · AI/CL/LG's note
The paper tests whether expert-routing probabilities inside MoE vision-language models reveal when a question assumes an object that is not present. Its detectors use routing from target tokens in Qwen3-VL-30B-A3B-Instruct and Gemma-4-26B-A4B-it, then trigger a target-aware review prompt only when needed. On GQA-Inpaint, routing alone reaches ROC-AUCs of 0.9988 for Qwen and 0.9956 for Gemma, with weaker but still useful transfer to OBER. The authors say threshold shifts across datasets need recalibration, but false-positive review caused limited harm overall. ArXiv · AI/CL/LG's note
score 4