GLiDE distills reasoning-teacher knowledge into fast single-pass decision models
A new arXiv paper describes a family of decision models from 0.4B to 27B parameters that score candidate answers in one forward pass, trained with distillation from reasoning teachers.
In plain words
Decision models pick between candidate answers without writing out reasoning, which makes them fast but can cost accuracy. The GLiDE authors train such models to inherit some benefits of reasoning by learning from reasoning-capable teacher models. The result is a set of models that score answers in a single pass while reporting how confident they are.
What was announced
GLiDE builds on Qwen hybrid language models. The authors remove causal masking from full-attention layers so that context, question, and candidate answers can interact bidirectionally, and they transfer reasoning ability into single-pass inference using soft-target distillation from multiple teachers, including a reasoning model, combined with progressive post-training on diverse decision tasks. They also test parameter-free recurrent depth, which adds inference-time computation without generating tokens.
The reported results come from the paper's abstract. Across five model sizes, the authors place GLiDE on the performance-parameter Pareto frontier of a Hugging Face benchmark, with each model beating evaluated alternatives of equal or smaller size. The 27B version scored 71.52 on a hidden set with a 2.0% expected calibration error and 92.9 ms median latency on one GPU; with checkpoint averaging and recurrent depth, GLiDE 27B reached 73.37 with 2.1% calibration error.
What it does not show
- Evidence is limited to the paper's abstract; the full paper's methods, benchmarks, and comparisons are not covered, and no independent confirmation of the reported scores exists.
For security and IT teams
- Status
- The work appears as a new arXiv abstract dated 20 November 2026; the abstract does not state release of the models or code.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.