proomt

Search

Search posts, papers, and topics

All posts

Hugging Face Daily PapersJulien Siems, Riccardo Grazzi, Korbinian Pöppel1 min readpaperadvanced

Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention

Summary

Complex KDA (CKDA) extends Kimi Delta Attention by widening gate and beta ranges, allowing a single diagonal‑plus‑rank‑one transition to represent any orthogonal matrix, including 2‑D rotations. The paper proves this expressivity, keeps the model non‑expansive, and shows CKDA achieves state‑of‑the‑art length‑extrapolation and competitive language‑model performance.

  • Extending KDA gates to [-1,1] and β to [0,2] creates Complex KDA, enabling any orthogonal diagonal‑plus‑rank‑one matrix in a single transition.
  • CKDA keeps the diagonal‑plus‑rank‑one, non‑expansive update, preserving the efficiency and stability of linear RNNs.
  • Proof that every orthogonal diagonal‑plus‑rank‑one matrix is exactly a CKDA transition; a single layer can track any finite subgroup of SO(3).
  • Empirically, CKDA achieves best length‑extrapolation on S3, S4, periodic audio and matches/exceeds Transformers on language modeling while scaling well.

Developers of long‑context LLMs and efficient sequence models should care because CKDA offers a provably expressive yet stable linear RNN alternative with competitive performance.

8/10

Related reading

  1. SpectralShift: Effective Context Window Extension of Gated DeltaNet via Spectral Reparameterization

    This paper introduces SpectralShift, a spectral reparameterization method for extending the context window of Gated DeltaNet (GDN) linear attention models. It reconfigures the decay spectrum by enhancing slow propagation and preserving fast-decaying modes, consistently improving long-context capabilities during continual pretraining.

    Hugging Face Daily Papersarxiv.org1 minpaper
  2. VC-Attention: Value Smoothing and Softmax Casting for Low-bit Attention

    VC-Attention introduces a training‑free low‑bit attention pipeline for diffusion transformers. It smooths value tensors via lightweight online clustering (V‑Smooth) and quantizes only the residual after subtracting block means, restoring the mean from the softmax row sum. It also replaces the FP32 softmax exponential with a fused FP8 cast (ExpCast‑FP8) that maps log‑scores directly to E4M3 probab…

    Hugging Face Daily Papersarxiv.org1 minpaper
  3. Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation

    Video DeltaNet (VDN) replaces full‑softmax attention in video diffusion models with a hybrid: per‑frame local Softmax for fine detail and a bidirectional linear memory (Video Delta Attention) for long‑range context. A teacher‑alignment schedule injects the linear branch into a pretrained MiniMax H3 model, preserving Softmax for text/audio streams. On eight NVIDIA B200 GPUs VDN‑H3 denoises a 14.3‑…

    Hugging Face Daily Papersarxiv.org1 minpaper
  4. Attention Is All You Need

    The paper proposes the Transformer, a sequence‑to‑sequence model that relies solely on self‑attention, eliminating recurrence and convolutions. It achieves state‑of‑the‑art translation BLEU scores while training orders of magnitude faster.

    Hall of Famearxiv.org27 minpaper
  5. Zing-0.5: Toward Playable Worlds with Real-Time Joint Action and Text Control

    Zing‑0.5 is a 5 B autoregressive world model that lets users control generated environments in real time using both keyboard actions and text prompts. The paper introduces unified action‑text conditioning, segment‑level teacher distillation, and a low‑cost streaming inference pipeline that runs at 24 FPS (832×480) for about $0.009 per minute, achieving 81 % overall and 88.5 % consistency on a nav…

    Hugging Face Daily Papersarxiv.org1 minpaper