VisionZip: Longer is Better but Not Necessary in Vision Language Models

VLM quality kept improving with longer visual token sequences, multiplying prefill cost - but visual tokens from CLIP/SigLIP encoders carry large redundancy.

Observes that vision-encoder tokens contain substantial redundancy, then selects an informative subset to pass to the LM using a dominance metric computed in the vision encoder's own feature space (retained tokens are those that dominate the attention distribution of the ones dropped), plus a balancing mechanism so dropped-token information is not lost entirely.

8x improvement in PREFILL TIME versus the unpruned baseline, and at least 5% performance gain over the previous state-of-the-art pruning method across nearly all settings; the pruned LLaVA-NeXT 13B infers FASTER than the unpruned LLaVA-NeXT 7B while scoring better. Both the baseline and the direction are named - this is the headline vision-token number in the literature.

Venue CVPR 2025 (crossref container '2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition'). Read the 8x as a PREFILL number: that is the vision-encoder forward plus the LM's attention over the retained tokens, and it lands on the cost that dominates VLM TTFT. It also gives the crispest demonstration of the small-model-beats-large-model effect - pruning 13B below 7B's cost - which is the practical argument for why token budget, not parameter count, is the VLM serving lever. Post-VisionZip work has pushed on multi-stage dropping, token recycling and 2026-era training-free pruning variants. VENUE_TRACK BACKFILL 2026-10-04 (paper-fields pass, agent w3p): venue_track 'main'. This is STRUCTURAL rather than researched: cvpr has no Findings volume and no workshop track in which a paper could appear, so 'main' is the only possible honest value and no per-paper evidence was needed. The field is annotated here so that 'findings' and 'workshop' remain MEANINGFUL - a filter for findings=true must return only the papers that actually are findings, which requires every other paper to be explicitly NOT findings.