ECHO Achieves Efficient Causal Self-Recursion in Model Training
October 16, 2026
ECHO implements Efficient Causal Self-Recursion by using the backbone's final layer to predict multiple horizons, integrating future prediction into the autoregressive objective. It employs Dynamic Depth Sampling to keep deeper prediction horizons computationally affordable during training.
HOW THIS AFFECTS YOU
●
builderYou can benefit from improved downstream accuracy by incorporating multi-horizon objectives more efficiently.
●
researcherThis allows for multi-token prediction benefits without the overhead of auxiliary modules.