The paper introduces a method to retrofit depth recurrence into pretrained vision‑language models, reducing the language backbone’s parameter count by pruning layers and looping over the remaining ones with linear adapters. A three‑stage recovery training—language‑only pretraining, vision‑language projector alignment, and full fine‑tuning—helps recover performance, with distillation from the non‑recurrent model further improving results. Experiments on TinyLLaVA‑1.4B show the retrofitted recurrent model outperforms the unpruned baseline while saving 23.3% of language parameters and performs better on reasoning‑oriented benchmarks.
arXiv:2610.03389v1 Announce Type: cross Abstract: Vision language models (VLMs) incur substantial inference cost because every visual token is processed by the attention and MLP projections of every...
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. 26596v1 Announce Type: cross Abstract: Multimodal large language models (MLLMs) have demonstrated remarkable capabilities by integrating visual and textual understanding within a unified transformer architecture.
AdaLoop introduces a question‑conditioned visual recurrence framework for vision‑language models, allowing the vision encoder to adapt its computation to each question while keeping pretrained weights frozen. The method reuses encoder transformations across multiple iterations, improving performance over standard fine‑tuning and single‑pass baselines. Experiments on the VisRO dataset demonstrate gains across three VLM families on both in‑distribution and out‑of‑distribution examples.
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.
arXiv:2507.00754v3 Announce Type: replace Abstract: The integration of Large Language Model (LLMs) blocks with Vision Transformers (ViTs) holds immense promise for vision-only tasks by leveraging the...
arXiv:2607. 00434v1 Announce Type: cross Abstract: Vision-language models (VLMs) have become a paradigm for multimodal learning, yet remain unstable due to object hallucination, weak visual grounding, and catastrophic forgetting after full-parameter instruction tuning.
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...
arXiv:2603. 00198v2 Announce Type: replace-cross Abstract: Token reduction accelerates long-video vision--language models (VLMs), but existing methods target Transformers, where reduction is treated as token pruning.
arXiv:2605. 18160v2 Announce Type: replace-cross Abstract: In recent years, multimodal large language models (MLLMs) have achieved remarkable progress, primarily attributed to effective paradigms for integrating visual and textual information.
LoopVLA introduces a recurrent Vision‑Language‑Action architecture that learns to refine multimodal representations, predict actions, and estimate when further refinement is unnecessary. By iteratively applying a shared Transformer block and producing a sufficiency score at each step, it decouples refinement from fixed layer indices and aligns confidence scores with action quality through a self‑supervised objective. Experiments on LIBERO, LIBERO‑Plus, and VLA‑Arena demonstrate that LoopVLA reduces model parameters by 45% and boosts inference throughput up to 1.7× while matching or surpassing strong baselines in task success.
LearnPruner rethinks token pruning in vision‑language models by addressing shortcomings of attention mechanisms in vision encoders and large language models. The method uses a two‑stage approach: first a learnable pruning module removes redundant vision tokens after the vision encoder, then only task‑relevant tokens are kept in the LLM’s middle layer. Experiments show that LearnPruner retains about 95% of performance while using only 5.5% of vision tokens, achieving a 3.2× speed‑up.