arXiv:2609.05916v1 Announce Type: cross Abstract: Large vision-language models (LVLMs) achieve strong multimodal understanding, but the hundreds to thousands of visual tokens they process impose subs...
FOVEA introduces a cache‑friendly, on‑demand visual evidence adaptation for multimodal speculative decoding, enabling a lightweight draft model to dynamically retrieve a bounded subset of visual memory based on a cumulative‑mass rule. The retrieved visual readout is fused with the draft hidden state via a lightweight gated residual correction, avoiding the insertion of visual tokens into the autoregressive context. Experiments on various vision‑language backbones and benchmarks show that FOVEA improves draft acceptance and speeds up end‑to‑end decoding by up to 2.13× compared to traditional autoregressive decoding.
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.
The paper investigates how attention sparsity behaves in autoregressive image generation, finding a distinct diagonal sparsity pattern due to spatial locality of visual tokens. It introduces a diagonal‑aware sparse attention mechanism that skips KV entries along the diagonal within a recent window, achieving up to 3.1× higher throughput and 1.19× lower latency with less than 2% quality loss compared to dense inference.
The paper introduces the concept of multi-view hallucination (MVH), where large vision-language models produce incorrect answers when processing images from multiple viewpoints. It presents MVH-Bench, a benchmark of 4.8k question-answer pairs that target cross-instance and cross-view hallucinations, and shows that MVH is common across recent models. The authors propose Reference Shift Contrastive Decoding (RSCD), a training-free decoding method that mitigates visual interference, achieving significant performance gains on MVH-Bench with LLaVA-OneVision and Qwen2.5-VL.
arXiv:2609.37042v1 Announce Type: cross Abstract: Video Large Language Models (VideoLLMs) have achieved strong video understanding capabilities but incur substantial inference overhead due to the lar...
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.
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.
arXiv:2608.20473v1 Announce Type: new Abstract: Video language models process videos as dense visual-token sequences with substantial representational redundancy. Compressing these sequences is there...
GeoGR introduces a two‑stage framework for efficient 3D vision‑language models. The first stage, GeoSemZip, compresses multi‑view visual tokens into compact 3D voxel representations while preserving important scene information through contextual anchors. The second stage, GroupRoute, dynamically routes tokens through the language model, allowing low‑relevance tokens to skip layers and rejoin later, thereby reducing computation without significant loss of performance.
The paper investigates whether keeping non‑spatial register tokens in vision transformer decoders improves image reconstruction and generation. Using representation autoencoders with frozen register‑equipped encoders, the authors show that joint self‑attention over both patch and register tokens yields better reconstruction quality on ImageNet‑1K, with lower FID scores and higher PSNRs compared to patch‑only baselines. For generation, a multi‑head attention pooling predictor recovers register features from sampled patches, further reducing FID by nearly 10% over the patch‑only approach.
LongVU‑TTT is a causal test‑time training method for long‑video multimodal large language models that inserts a convolutional resampler with fast‑weight updates between the vision encoder and the LLM. The fast weights adapt per video and contextualize frame features before compression, while a hybrid selector keeps explicit visual evidence for downstream reasoning. Experiments show that TTT‑Conv outperforms TTT‑MLP and bidirectional Mamba2 on MLVU, and beats attention‑ and fixed‑state recurrent resamplers on three benchmarks, achieving competitive results on five video‑understanding tasks after reducing 512 frames to 128 LLM frames.