GLiDE decision models distill reasoning into single-pass inference
November 20, 2026
GLiDE is a family of decision models from 0.4B to 27B parameters that removes causal masking to allow bidirectional attention between context and candidate answers. The method uses soft-target distillation from reasoning teachers to provide fast, calibrated scores without generating text.
HOW THIS AFFECTS YOU
●
builderYou can get reasoning-level decision quality without the latency or cost of generating long text chains.
●
researcherYou can use parameter-free recurrent depth to scale inference-time computation without adding new tokens.