The Flow has not summarised this story yet — read it at arXiv Computer Vision.
The paper introduces PACE, a Pareto‑adaptive compression framework designed for native multimodal large language models (MLLMs) that directly tokenize pixel patches without external vision encoders. By analyzing token discriminability and heterogeneous redundancy across spatial dimensions, PACE decomposes attention into height‑ and width‑specific saliency maps and applies a Pareto‑optimization strategy to allocate token budgets, achieving near‑lossless performance with only 25‑50% of visual tokens and more than 2× FLOP savings. Experiments demonstrate that PACE not only preserves but sometimes improves benchmark results compared to existing ViT‑based compression methods.
arXiv:2608. 19825v1 Announce Type: cross Abstract: Medical image captioning is a technique that accelerates early-stage diagnostic workflows and enhances the interpretability of medical diagnostic AI systems.
arXiv:2608.29663v1 Announce Type: new Abstract: Remote photoplethysmography (rPPG) enables contactless physiological measurement from facial videos, yet its subtle pulse-related variations are easily...
Hypnos is a multi‑modal sleep foundation model that learns generalisable representations of physiological signals by predicting the next token across eight sensing modalities (EEG, ECG, respiratory, etc.) using an auto‑regressive RQ‑Transformer. Trained on over 20,000 overnight polysomnography recordings, it outperforms existing foundation models on most benchmarks, matching strong supervised baselines for sleep‑stage classification while requiring 100× less labelled data. Additionally, Hypnos generalises beyond sleep, surpassing a dedicated ECG model in atrial fibrillation detection.
arXiv:2606. 09605v1 Announce Type: new Abstract: Foundation models offer a promising route to compress multi-modal physiological signals into compact representations of human health, with broad applications across sleep medicine, cardiology, neurology and other healthcare domains.
VIVAS is a new Vision‑Language Model pre‑training framework that addresses the lack of fine‑grained visual perception in existing VLMs. It introduces a unified token space and a dense‑structural‑semantic vision tokenizer that expands the textual vocabulary with visual tokens, enabling vision‑language unified autoregressive supervision over both visual details and linguistic content. Trained on 12.4 T tokens, VIVAS achieves state‑of‑the‑art results on 7 tasks and 39 multimodal benchmarks.