New Recursive Language Model Architecture DABSN
BleedingXiko · reddit · 2026-07-17
The author shares the preprint and code for DABSN (Dynamic Adaptive Bias State Network), a novel recursive language model architecture, and is seeking collaborators for larger-scale reproduction and expansion.
What has been shown
- The paper focuses on the architecture itself and its performance on reasoning, memory, and long-sequence tasks.
- It mentions benchmarks like MQAR, Copy, Key-Value retrieval, and A5/60.
- The code is public, including PyTorch, C++, and Triton implementations for reproducibility.
Future Plans
- The author has already trained a small language model using the same cell: 24M parameters, 1 billion token pre-training, GPT-2 tokenizer.
- He plans to write a second paper focusing on language modeling, long context, and scaling. He is looking for collaborators to help independently reproduce it, design stronger baselines, or provide larger GPU resources.
More from Research
- Nat Lambert shares a reading list on synthetic data and agentic SFT data — natolambert · 2026-07-22
- Turning Noise into Signal: Predicting TCR Binding Using AlphaFold3 Hallucinations — quaidmorris · 2026-07-22
- Lightwheel AI Launches SimReadyGen: Text-to-Physics-Accurate Robot Sim Assets — ZeYanjie · 2026-07-22
- PNAS special issue examines copyright, governance, and AI in the legal system — chrmanning · 2026-07-22
- WeirdChat catalogs strange model behaviors from more than 100 million sampled responses — JacobSteinhardt · 2026-07-22
- New agentic benchmark shows AI managers escalate to coercion and fake success — Jasmine Brazilek · 2026-07-22