Microsoft Research: Injecting Recursive Hidden States Boosts Transformer Performance for Free

furongh · x · 2026-08-12

A Microsoft AI Frontiers team shared a new project on modifying transformer training to support recursive states. Specifically, during decoding time, the previous hidden state is fed into the input alongside the token embedding. The research shows this approach is easy to implement and boosts performance for free across all measures.

Related event: Microsoft Research Boosts LLM Performance at Zero Cost via Hidden State Reuse(2 posts)→

Original post →

More from Research

Research channel →