DABSN recurrent LM architecture preprint and code released

Research OpenSource

TL;DR: A new recurrent architecture called DABSN achieves strong results on reasoning and long-context benchmarks; the author seeks collaborators for scaling.

Summary: The DABSN (Dynamic Adaptive Bias State Network) architecture is detailed in a new preprint, with PyTorch, C++, and Triton implementations publicly available. A 24M-parameter language model trained on 1B tokens shows promising results, and the author is opening collaboration for scaling and independent evaluation.

Why it matters: For AI builders exploring alternatives to transformers, this open-source recurrent architecture offers a potential direction for efficient long-context models. Those with GPU clusters or evaluation expertise can contribute to scaling and benchmarking.

Source: reddit