Weekly AI Paper Digest: Attention, Nvidia VLA, TTS, and Graph Neural Networks

This week’s AI paper roundup highlights five recent studies that push the boundaries of long‑context modeling, autonomous‑driving reasoning, non‑autoregressive speech synthesis, latent‑space multi‑agent collaboration, and graph neural network depth.

1. Every Token Counts: Generalizing 16M Ultra‑Long Context in Large Language Models

The authors define ultra‑long context modeling as a memory challenge and propose Hierarchical Sparse Attention (HSA), which satisfies sparsity, random‑access flexibility, and length‑generalization. Integrated into a Transformer, HSA builds an 8‑billion‑parameter mixture‑of‑experts model called HSA‑UltraLong. https://go.hyper.ai/axKy6

2. Alpamayo‑R1 (AR1): Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail

AR1 is a visual‑language‑action (VLA) model that combines causal reasoning with trajectory planning. Compared with a trajectory‑only baseline, AR1 improves planning accuracy by up to 12% in complex scenarios, reduces road‑deviation rate by 35%, and cuts near‑collision incidents by 25% in closed‑loop simulation, offering a practical path toward Level‑4 autonomous driving. https://go.hyper.ai/Q15y9

3. F5‑TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching

F5‑TTS is a fully non‑autoregressive text‑to‑speech system built on flow matching and a Diffusion Transformer (DiT). Trained on a 100‑k‑hour multilingual dataset, it demonstrates highly natural zero‑shot generation, seamless code‑switching, and efficient speed‑control performance. https://go.hyper.ai/Q15y9

4. Latent Collaboration in Multi‑Agent Systems (LatentMAS)

LatentMAS proposes an end‑to‑end, training‑free framework for pure latent‑space collaboration among LLM agents. Each agent generates a latent thought representation from its final hidden state; a shared latent working memory stores and propagates these representations, ensuring lossless information exchange across agents. https://go.hyper.ai/M587U

5. Deeper‑GXX: Deepening Arbitrary GNNs

Deeper‑GXX introduces two core modules: Weight‑Decaying Graph Residual Connection (WDG‑ResNet) to alleviate gradient vanishing and suppress shadow‑neighbor effects, and Topology‑Guided Graph Contrastive Loss (TGCL) that leverages graph topology for contrastive learning, enhancing node‑representation discrimination and mitigating over‑smoothing. https://go.hyper.ai/gwM7J

The full list of papers and additional AI research can be found on HyperAI’s “Latest Papers” section, and researchers are invited to submit high‑quality work for future newsletters.