The developers say the tokens let a model keep reasoning even after its generation history is cleared.
A developer announced register tokens for diffusion large language models, saying a model can write to and read from them while generating text. The claimed benefit is that the model can keep reasoning even after its generation history is cleared.
—
2 Sources, first seen ago
—
Not ranked yet
—
Not ranked yet
@albert_ge_95Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read from while generating text. They let the model keep reasoning even after its generation history is cleared 🧵
@charles_irlRT @albert_ge_95: Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read fr…@albert_ge_95Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read from while generating text. They let the model keep reasoning even after its generation history is cleared 🧵@charles_irlRT @albert_ge_95: Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read fr…@albert_ge_95Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read from while generating text. They let the model keep reasoning even after its generation history is cleared 🧵@charles_irlRT @albert_ge_95: Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read fr…@albert_ge_95Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read from while generating text. They let the model keep reasoning even after its generation history is cleared 🧵@charles_irlRT @albert_ge_95: Excited to introduce registers for diffusion LLMs! TL;DR: we develop register tokens that a dLLM can write to and read fr…
