Registers for diffusion LLMs: tokens that let models keep reasoning after history is cleared
charles_irl · x · 2026-09-17
Researchers introduce register tokens for diffusion LLMs: special tokens the model can write to and read from during text generation. The key benefit is that the model can keep its chain of reasoning even after its generation history is cleared, offering a new mechanism for state persistence in dLLMs. Details are laid out in the accompanying thread.
Related event: Register Tokens Let Diffusion LLMs Keep Reasoning After Clearing Context(2 posts)→
More from Research
- JHU's Track, Articulate, Act reconstructs articulated objects and dexterous interactions from casual monocular videos — mangahomanga · 2026-09-17
- Nature's Paper2Agent converts research papers into interactive AI agents via MCP servers — krishnan · 2026-09-17
- Frozen Parnet embeddings match task-specific tools and larger genomic language models without fine-tuning — anshulkundaje · 2026-09-17
- 25 Fields Medalists Sign Declaration Warning Against Measuring AI Progress by Solved Problems — krishnan · 2026-09-17
- PARNET: A CLIP-seq-based foundation model for RNA sequence representation — anshulkundaje · 2026-09-17
- Gemini, Claude and Grok all invent the same "Dr. Elena" — evidence of shared training data — dejanseo · 2026-09-17