V-LUMEN introduces a visual lookup memory for vision‑language models by discretizing visual features with a frozen codebook, creating spatially aware memory keys via hashing, and modulating them with text‑conditioned relevance signals. The retrieved embeddings are fused into a frozen Qwen3‑VL‑2B backbone through a gated residual pathway, yielding substantial improvements on five reasoning‑intensive benchmarks and zero‑shot transfer to three disjoint image datasets. At a matched parameter budget, reallocating capacity from LoRA or MoE adapters to V‑LUMEN preserves accuracy on two backbones, demonstrating that external visual memory offers an independent capacity axis. whyItMatters":"The paper shows that adding a non‑parametric visual lookup memory can significantly boost vision‑language model performance without increasing per‑token compute, offering a new way to scale such models."
arXiv:2610.15855v1 Announce Type: new Abstract: We study how to detect and mitigate visual attention sinks in vision-language models (VLMs). Visual attention sinks are image tokens that absorb a larg...
VisCache introduces a two-stage, plug‑and‑play framework for pruning visual key‑value caches in Vision Large Language Models without retraining. The first stage filters out temporally redundant keyframes, while the second stage, PruneKV, applies a parabolic layer‑wise budget and asymmetric update to selectively prune keys and fuse values, preserving essential context. Experiments show up to 2.35× speedup and significant memory savings with only 19–28% of the original cache retained, outperforming existing baselines.
arXiv:2607. 12815v1 Announce Type: new Abstract: Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces.
VisionWeave introduces elastic visual representation weaving, a native capability for multimodal large language models that learns where and at what granularity to encode visual information. The method combines a gated spatial pooler for coarse representations with a granularity router that allocates content‑adaptive token usage, trained end‑to‑end on large‑scale data. Experiments on Qwen3.5‑4B and Qwen3.8‑27B show that VisionWeave can save 43.0% of tokens while preserving 98.9% of performance across eight benchmarks, and delivers significant throughput gains and latency reductions when deployed on the SGLang serving engine.
CatchUp is a training‑free framework that improves efficiency in large vision‑language models by depth‑aware re‑admission of visual tokens. It repeatedly evaluates query‑conditioned relevance, processes only selected tokens per decoder interval, and preserves deferred tokens for later re‑admission, using reference‑to‑current transition signals to approximate missed representation evolution. Experiments on three referring‑expression grounding datasets and six multimodal benchmarks show that CatchUp retains over 92% of full‑model performance with 88.9% fewer active tokens and reduces end‑to‑end latency by 8.0% at a 64‑token budget.
Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces. We ask whether CoT requires continued access to image tokens, or whether it mainly operates over visual information already made available earlier in the forward pass.
arXiv:2606. 11576v1 Announce Type: cross Abstract: Modern Vision-Language Models (VLMs) benefit from chain-of-thought prompting and test-time scaling, but these gains often come with prohibitive inference cost due to large visual contexts and long decoding chains.
The paper introduces the Separable Law, a framework that predicts how vision‑language model performance varies with language backbone size and visual token count. By fitting this law to 26 InternVL and QwenVL models across a range of backbone sizes (1B–72B) and image resolutions (224–8K pixels), the authors show that some question types scale predictably with model capacity while others do not. The law also provides a closed‑form rule for allocating compute between backbone size and visual tokens, helping to choose near‑optimal model and image sizes under a fixed budget.
arXiv:2607. 23913v1 Announce Type: new Abstract: Modern vision-language models (VLMs) increasingly rely on dynamic or high-resolution visual encoding, producing thousands of visual tokens that substantially increase downstream language-model inference cost.
The paper introduces Visual Retrieval Heads (VRHs), a small fraction of attention heads in vision‑language models that are causally responsible for grounding text descriptions to image regions. By recasting head‑scoring methods and evaluating across eleven VLMs and five benchmarks, the authors show that masking the top 20 VRHs can drop grounding accuracy by up to 80 percentage points, while random masking has little effect. VRHs generalize across various visual reference tasks, preserve output format while corrupting localization, and transfer causally across models sharing an LLM backbone.
PACE introduces a training‑free Condense‑and‑Extract framework that speeds up Vision‑Language Model inference by first adaptively downsampling visual inputs before encoding and then selectively retaining essential tokens during decoding. The Adaptive Pixel Compressor (APC) reduces encoder workload while preserving global context, and the Dynamic Dual‑Attention Extractor (DDAE) keeps task‑critical details by fusing visual and language signals. Applied to Qwen2.5‑VL‑7B, PACE maintains 93.8% of performance using only 10% of visual tokens, achieving a 3.1× speedup in time to first token.
Vision-language models commonly project all tokens produced by a pretrained vision encoder into a large language model. However, final-layer features can discard text, local attributes, and spatial relationships, while high-resolution inputs substantially increase context length and inference latency.