PPS: Plug-and-Play Saccadic Vision for Fine-Grained Classification proposes a lightweight extension that transforms any pretrained backbone with spatial feature maps into a saccadic classifier. The method adds only a small priority head and a single cross-attention fusion block, using a weight-tied encoder to process a downsampled peripheral glance and high-resolution glimpse crops selected by a priority distribution. A sequential sampler with non-maximum suppression draws a configurable number of glimpses, and a residual multi-head cross-attention block fuses the glance with the glimpses; the priority head is trained via a per-glimpse information-gain target from a lightweight memory bank, avoiding reinforcement learning. The approach achieves linear complexity in the number of glimpses and decouples peak memory from input resolution, consistently improving performance across nine fine-grained classification benchmarks.
arXiv:2607.09086v2 Announce Type: replace Abstract: We present Subtoken Vision Transformer (SubViT), a selective image tokenization method for fine-grained visual recognition. Standard Vision Transfo...
Recent advancements in MLLM-based long-form video understanding have mitigated inference-time computational cost and limited context lengths by selecting query-relevant frames. However, existing approaches predominantly rely on external proxy scorers and rigid heuristic rules, inevitably suffering from misalignment with the target MLLM's intrinsic evidence and failing to accommodate the non-uniform spatiotemporal information density.
arXiv:2605.15195v2 Announce Type: replace Abstract: Recent feed-forward reconstruction models, such as VGGT, have proven competitive with traditional optimization-based reconstructors while also prov...
Gaze Attention is a new mechanism for multimodal large language models that selectively focuses on relevant visual regions during each generation step, rather than attending to all visual tokens. By grouping tokens into spatial regions and using learnable context tokens to retain global information, it reduces attention computation and visual key‑value entries by up to 90%. Experiments on 13 image and 6 video benchmarks show that Gaze Attention matches or outperforms dense‑attention baselines while using fewer visual resources.
arXiv:2608.22869v1 Announce Type: cross Abstract: While Vision-Language-Action (VLA) models have leveraged internet-scale pretraining and task-focused finetuning to achieve strong performance on long...
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.
UniMem is a framework that unifies visual and textual memory with low‑level control in a single Vision‑Language‑Action model, using an event classifier, keyframe encoder, and caching to manage memory efficiently. It addresses non‑Markovian tasks that require remembering past scenes, avoiding the memory bottleneck of separate VLMs and the performance drop from fixed‑interval frame sampling. Across five simulation and four hardware tasks, UniMem outperforms fixed‑interval sampling (93.4% vs. 68.2%) and hierarchical VLM‑VLA baselines (80.0% vs. 43.5%) while offering faster inference and a simpler training pipeline.
arXiv:2508.03351v3 Announce Type: replace-cross Abstract: Large language models (LLMs) have demonstrated remarkable capabilities across diverse language tasks, motivating their extension to vision-la...
arXiv:2605.12491v2 Announce Type: replace Abstract: Vision Transformers (ViTs) learn rich visual-semantic representations through all-to-all self-attention among patch tokens. However, this design im...
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.
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.
MapLightning introduces a new approach to online vectorized HD map construction that replaces dense bird’s‑eye‑view grids with a compact set of 1D learnable map tokens. Using a transformer‑based mapper that applies full self‑attention to combined image and map tokens, the method achieves efficient inference, lower memory usage, and improved global context without relying on camera projection parameters. Experiments on nuScenes and Argoverse 2 show state‑of‑the‑art accuracy, with a lightweight variant outperforming MapTRv2 by over 10 mAP and running 1.73× faster at 40+ FPS while using 53 % less memory.