Just MLPs: Efficient Visual State Reconstruction for Multimodal Language Models
Long visual token sequences often account for a substantial fraction of the computational overhead in multimodal large language models~(MLLMs). Existing approaches reduce this cost by pruning redundant visual tokens, but permanently discard visual evidence that may become useful in subsequent layers. We instead ask whether all visual tokens can be preserved while reducing the cost of repeatedly evolving the representations through the Transformer. To answer this question, we perform low-rank interventions on visual-to-text information flow. We find that, after visual-to-text attention is blocked, restoring only a few directions recovers most of the lost accuracy, suggesting the relevant v...
| Authors | Jingdi lei, Junxian Li, Di Zhang et al. |
|---|---|
| Primary category | cs.CV |
| Published | 28 September 2026 |
| Last updated at source | 28 September 2026 |
Connected in the knowledge graph
- huaXiaKyrie/pixmo-ama-trainDataset · cited by
Models and datasets referencing this paper
- huaXiaKyrie/pixmo-ama-trainDataset
Silytics shows public metadata from arXiv and links to the original. Information may be incomplete or out of date. Last refreshed 17 October 2026. Licenses and terms of the original apply to the work itself. Not financial, legal or professional advice.