Papers

GLiDE:从推理教师蒸馏的单次前向决策模型,0.4B 到 27B

GLiDE: A Fast, Calibrated Decision Model Distilled from Reasoning Teachers

Why we picked it

适合做低延迟决策场景的工程师看看。GLiDE 用蒸馏把推理模型的能力塞进单次前向,27B 版延迟不到 100ms 还自带校准分数,比跑完整推理链省太多算力了。做分类、路由这类要快速出答案的活儿可以参考它的做法。

IBM Research 与红杉资本联合推出的 GLiDE 是一个面向单次前向推理的决策模型家族,参数量从 0.4B 到 27B 共五个规格。它在 Decision-Making Benchmark 上进入性能-参数量 Pareto 前沿,27B 版本在隐藏集上拿到 71.52 分,期望校准误差仅 2.0%,单 GPU 中位延迟 92.9 ms。

Body

技术上,GLiDE 基于 Qwen 混合语言模型,通过移除全注意力层的因果掩码,让上下文、问题和候选答案之间实现双向交互。为了把推理教师的能力迁移到单次推理,研究团队采用多教师软目标蒸馏(教师中包含推理模型),并在多样化决策任务上做渐进式后训练。此外,GLiDE 27B 引入免参数的循环深度机制,在不生成 token 的前提下增加推理时计算量。

结果上,加入 checkpoint 平均与循环深度的 GLiDE 27B 将隐藏集分数提升到 73.37,期望校准误差 2.1%。对需要在毫秒级延迟下做出带置信度决策的应用(如分类、路由、审核)而言,这条不生成文本、却保留推理收益的路线提供了一个比直接调用推理模型更省算力的选择。

Original · arXiv cs.AI

GLiDE: A Fast, Calibrated Decision Model Distilled from Reasoning Teachers

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\%.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include GLiDE, Qwen, distillation, reasoning models.

Source attribution: curated from arXiv cs.AI (@Sahibzada Allahyar, Julia White, Dhruv Atreja, George Hurn-Maloney, Urchade Zaratiana).