The long-visual-context bottleneck
ReToken, submitted July 30, addresses the tendency of vision-language models to lose relevant evidence as image collections or videos grow. Processing every visual token together is expensive, while additional distractors degrade retrieval.
The method trains one embedding as an explicit retrieval target, selecting a sparse query-relevant subset from a prefilled visual KV cache. It is trained on a small image-QA dataset and then transferred to image and video benchmarks.
Author-reported results
The authors report gains of 13.4 points for Qwen3VL-8B and 12.4 points for InternVL3.5 on Visual Haystacks. On long-video LVBench, Qwen3VL-8B improved by 8.0 points with zero-shot transfer. They say training and long-video inference each fit on one H100 and have released code.
Use and limitations
The approach could help enterprise video search, surveillance review, and lecture navigation where relevant intervals must be found without rereading everything. This is a preprint, however, and the numbers are author-reported on selected models and benchmarks. Deployment should reproduce miss rates, cache cost, and latency across video lengths and query types.