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

Sahibzada Allahyar, Julia White, Dhruv Atreja, George Hurn-Maloney, Urchade ZaratianaarXiv 2026 · arXiv

TL;DR

GLiDE is a new family of "decision models" that delivers the high-quality performance of multi-step reasoning models at the lightning-fast speed of a single forward pass. By distilling the "thinking" process of large reasoning teachers into a specialized bidirectional architecture, GLiDE achieves state-of-the-art accuracy and industry-leading calibration (knowing when it’s likely to be wrong) without generating a single token of intermediate text.

What Problem Does This Paper Address?

In the world of LLMs, there is a persistent "reasoning gap." We know that forcing a model to "think step-by-step" (Chain-of-Thought) significantly boosts its accuracy on complex tasks. However, this comes at a massive cost: generating reasoning tokens is slow and computationally expensive.

For many applications—like routing API requests, selecting tools, or classifying sensitive content—we need a "Decision Model" that is fast, reliable, and provides a clear confidence score. Traditional encoders are fast but often lack the "smarts" of reasoning LLMs. The authors of GLiDE asked: Can we capture the benefits of reasoning while retaining the speed of a single-pass prediction?

Methods

The researchers at Fastino Labs developed a multi-pronged approach to bridge this gap:

  1. Hybrid Bidirectional Architecture: They started with Qwen hybrid language models, which mix standard self-attention with efficient Gated DeltaNet layers. Crucially, they removed the causal mask from the full-attention layers. This allows the model to look at the context, the question, and all candidate answers simultaneously before making a choice, rather than just looking "backwards" as standard GPT-style models do.
  2. Reasoning Distillation: Instead of just training on "right/wrong" labels, they used a "Reasoning Teacher." They let a larger model think through a problem and then recorded its final probability distribution over the answers. GLiDE was trained to match this "soft target" distribution. This teaches the model not just the right answer, but also the nuances and uncertainties the teacher found during its reasoning process.
  3. Dedicated Option Readout: Rather than predicting the next word in a vocabulary of 100k+ tokens, GLiDE uses a specialized 255-way readout that scores candidate options (like A, B, C, D) directly from the final hidden state.
  4. Parameter-Free Recurrent Depth: For their largest model (GLiDE 27B), they introduced a "loop" where certain middle layers are applied twice during inference. This adds extra "thinking time" (computation) without adding any new parameters to the model.
Figure 2: GLiDE V1.2 architecture. Each decision is encoded as a single sequence containing the context, question, coded candidate answers, and an answer prefix. Starting from a 64-layer Qwen hybrid decoder, GLiDE removes causal masking from its 16 full-attention layers, enabling bidirectional interactions while retaining causal recurrence in the 48 Gated DeltaNet layers. A dedicated 255-way readout scores the final hidden state, and a softmax over the offered options produces a normalized probability distribution. Decisions are computed in a single forward pass, without generating tokens or intermediate reasoning. Appendix C details the output semantics and deployment policies.

Datasets

GLiDE was trained on a massive mixture of 1.6 million distinct questions across 500+ task families, totaling roughly 2.3 billion tokens. This included classification, tool use, safety, and reasoning tasks.

For evaluation, the authors relied on independent, third-party benchmarks:

  • Hugging Face Decision Index: A leaderboard using hidden test sets that developers cannot see.
  • Artificial Analysis Decision Index: An externally administered benchmark assessing knowledge, reasoning, and safety via public APIs.
  • Complementary Sets: JevBench, SciFact, and internal clinical/regression suites.

Key Insights and Ideas

  • Bidirectional Interaction is a Superpower: Removing the causal mask allows the model to compare options against each other and the context more effectively. The ablation studies showed this was a primary driver of performance.
  • Distilling "Uncertainty" Matters: By training on the teacher's full probability distribution (soft targets) rather than just the gold label, the student model becomes much better calibrated.
  • Calibration as a Feature: GLiDE doesn't just guess; it knows how sure it is. With an Expected Calibration Error (ECE) of just 2.0%, if the model says it's 90% sure, it's actually right about 90% of the time. This allows developers to set "confidence thresholds" to defer difficult cases to humans.
  • Recurrent Depth Works: Re-applying layers (Equation 15 in the paper) is a viable way to scale performance at inference time without the memory overhead of a larger model.

Key Results

GLiDE V1.2 (27B) dominated the leaderboards at the time of publication:

  • #1 Ranking: It took first place on the Hugging Face Decision Index with a score of 71.52, outperforming the runner-up by 1.77 points.
  • Speed: It achieved a median latency of 92.9 ms on a single GPU—significantly faster than competitors like Jeffan-31B or mammoth-1.
  • Pareto Frontier: Across five sizes (0.4B to 27B), every GLiDE model outperformed every other evaluated alternative of equal or smaller size.
  • Stability: On the Artificial Analysis index, GLiDE remained within 0.6 points of the highest score across three different versions of the benchmark without being retrained.
GLiDE V1.2 shows a superior balance of high scores and low latency compared to other models.

Limitations and Future Work

  • Option Limit: The current readout is capped at 255 candidate options. While sufficient for most tasks, it limits extremely large-scale retrieval or multi-choice scenarios.
  • Hidden Set Gap: While GLiDE generalizes well, there is still a small gap between public benchmark scores and hidden test set scores, a common challenge in the field.
  • Future Directions: The authors suggest exploring even deeper recurrent loops and more sophisticated multi-teacher distillation strategies to further close the gap with "slow" reasoning models.

Summary updated

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\%.
BibTeX
@misc{allahyar2026glide,
  author = {Sahibzada Allahyar and Julia White and Dhruv Atreja and George Hurn-Maloney and Urchade Zaratiana},
  title = {{GLiDE: A Fast, Calibrated Decision Model Distilled from Reasoning Teachers}},
  year = {2026},
  eprint = {2611.26509},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.26509},
}
Figure