AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

V-LUMEN: Visual Lookup Memory for Embedding Scaling in Vision-Language Models

Miso Choi, Daekeun Kim, Eunji Kim, Jungbeom Lee

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes V-LUMEN, an extension of embedding scaling to vision-language models using visual lookup memory, based on the abstract only.

Problem

Existing embedding scaling methods target discrete text tokens; it is unclear whether such mechanisms extend to vision-language models whose visual representations differ from text tokens.

Approach

Extends embedding scaling from text tokens to visual representations in vision-language models via a visual lookup memory, increasing model capacity at the embedding level rather than by expanding transformer depth or width. Details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
attach V-Lumen visual lookup memory to a frozen backboneQwen3-VL-2B-Instruct (frozen) · five reasoning benchmarks (avg)accuracy: 2.18 -> 19.97Improved
add V-Lumen without any in-domain adaptationQwen3-VL-2B (frozen) · TextVQA, OCR-VQA, VizWiz-VQA (zero-shot)accuracy: TextVQA 11.00 -> 37.17; OCR-VQA 13.80 -> 47.67; VizWiz 2.00 -> 9.97Improved
move part of LoRA capacity into V-Lumen at matched total budget (ablation)LoRA adapter on Qwen3-VL-2B · five reasoning benchmarks (avg)accuracy: 22.76 -> 23.33No clear effect
move part of MoE capacity into V-Lumen at matched total budget (ablation)MoE adapter on Qwen3-VL-2B · five reasoning benchmarks (avg)accuracy: 19.03 -> 19.10No clear effect
add V-Lumen on top of identical LoRA(0.5) adapter (rank-matched) (ablation)LoRA(0.5) on Qwen3-VL-2B · five reasoning benchmarks (avg)accuracy: 23.79 -> 23.33No clear effect
add V-Lumen to LoRA at matched budget (ablation)LoRA adapter · MMMU-Proaccuracy: 26.50 -> 29.20Improved
add V-Lumen to LoRA at matched budget (ablation)LoRA adapter · VisResaccuracy: 22.03 -> 21.70Hurt
add V-Lumen to MoE at matched budget (ablation)MoE adapter · MMMU-Proaccuracy: 20.10 -> 19.00Hurt
grow memory table from 16,384 to 655,360 rows per cell (ablation)V-Lumen lookup · lookup latency (one fusion layer)lookup total (ms): 11.22 -> 11.23No clear effect
add more memory rows (ablation)V-Lumen · accuracyaccuracyNo clear effect
widen retrieved memory vectors (ablation)V-Lumen · accuracyaccuracyImproved
add V-Lumen memory to adapter, probed with logit lensMoE (VisRes) and LoRA (MMMU-Pro) · final-layer top-1 accuracytop-1 accuracy: VisRes 0.236 -> 0.251; MMMU-Pro 0.277 -> 0.289Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy