Princeton researcher proposes carrying decoder state across every token
single source · 1 articles · research · confidence: medium · first seen 2026-09-13 17:02 UTC
A Princeton researcher, Yifan Zhang, has published a technical report proposing a Recurrent Looped Transformer: a causal encoder paired with a decoder that carries its final hidden state and its sliding-window attention cache from token to token, with no reset at the boundary between prompt and response. The reference configuration pairs 48 encoder with 48 decoder layers, running 96 logical blocks per token, and the state path grows to 48t decoder blocks after t tokens. The report also describes hardware-aware execution around the recurrent core and an exact current-policy RL replay contract. No code, weights or measured results have been released.
What this means for you
Nothing to act on. This is a design proposal, not a released model: no code, no weights, no measurements, and no stated plan to publish any. The claim that carrying state across the serving boundary helps is untested.
Key facts
- ·Yifan Zhang, a Princeton researcher, published a technical report dated 13 September 2026 proposing a Recurrent Looped Transformer source
- ·The design pairs a causal encoder with a recurrent decoder that carries its final hidden state and layerwise sliding-window attention cache across every prompt and response token with no reset at the serving boundary source
- ·The reference tied configuration uses 48 encoder and 48 decoder layers, executing 96 logical blocks per token source
- ·The state path grows to 48t decoder blocks after t tokens source
- ·The report specifies hardware-aware execution around the recurrent core and an exact current-policy RL replay contract source
- ·No code, weights or measured results have been released source
What the sources say
- MarkTechPost — Sets out the architecture, layer counts and replay contract, and notes nothing has been released to test.
Sources
The original reporting. Follow these — they did the work.