Back to AI News
arxiv

GLiDE: Calibrated Decision Models Distilled from Reasoning Teachers

Researchers present GLiDE, a family of decision models from 0.4B to 27B parameters that score candidate answers in a single forward pass without generating text or intermediate reasoning. Built on Qwen hybrid language models, the approach removes causal masking from full-attention layers and combines soft-target distillation from multiple teachers, including a reasoning model, with progressive post-training on decision tasks. The 27B model reports a hidden-set score of 71.52 with 2.0% expected calibration error and 92.9 ms median latency on a single GPU; checkpoint averaging and parameter-free recurrent depth raise the score to 73.37 with 2.1% calibration error. Across five sizes, the authors say GLiDE lies on the performance-parameter Pareto frontier among evaluated alternatives.

AI Brief

It offers a concrete distillation recipe for retaining reasoning-model quality in fast, calibrated single-pass decision models, with reported latency and calibration results relevant to efficient AI deployment.

arXiv:2611.26509v1 Announce Type: new Abstract: Decision models enable fast, confidence-aware decisions by scoring candidate answers in a single forward pass, without generating text or intermediate reasoning. Yet reasoning can substantially improve decision quality, raising the question of whether its benefits can be retained without the cost of generating it. We investigate this question with \textbf{GLiDE}, a family of decision models spanning 0.4B to 27B parameters. Building on Qwen hybrid language models, GLiDE enables bidirectional interactions between the context, question, and candidate answers by removing causal masking from full-attention layers. To transfer reasoning capabilities into single-pass inference, we combine soft-target distillation from multiple teachers, including a reasoning model, with progressive post-training on diverse decision tasks. We further explore parameter-free recurrent depth to increase inference-time computation without generating tokens. Across five model sizes, GLiDE lies on the performance--parameter Pareto frontier of the Hugging Face \DI{}, with each model outperforming evaluated alternatives of equal or smaller size. \dmtwo{} (27B) achieves a hidden-set score of 71.52, with an expected calibration error of 2.0\% and a median latency of 92.9\,ms on a single GPU. Further checkpoint averaging and parameter-free recurrent depth yield GLiDE~27B, which reaches 73.37 with an expected calibration error of 2.1\%.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!