← Jucheng Shen

Blog

One Model, Two Roles

One Transformer updates two latent states, and only one update sees the input. That single difference is enough for the two states to take on distinct roles.

NeurIPS 2026

CadLLM

A training-free controller uses a diffusion LLM's own confidence to spend compute where the model is uncertain and save it where predictions have settled.

Findings of ACL 2026