GLiDE: A Fast, Calibrated Decision Model Distilled from Reasoning Teachers
Organizations: Fastino Labs
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 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%.
Figures & tables
| Model | Full | Same-skill | New-domain | Gen. gap | ECE (%) | Latency (ms) |
|---|---|---|---|---|---|---|
| GLiDE V1.2 | 71.52 | 68.31 | 76.32 | 2.0 | 92.9 | |
| Jeffan-31B | 69.75 | 65.86 | 75.58 | 4.7 | 121.0 | |
| mammoth-1 | 68.44 | 65.98 | 72.14 | 4.1 | 220.1 | |
| Decider v1.2 | 67.98 | 63.99 | 73.98 | 5.5 | 144.4 | |
| d3-turbo v1.1 | 67.31 | 63.98 | 72.31 | 5.2 | 190.5 | |
| Jev 1.14 | 60.24 | 58.91 | 62.24 | 8.0 | 196.0 † |
| Training characteristic | Value |
|---|---|
| Final-stage examples | 374,720 |
| Distinct questions | 1.6M |
| Training tokens | 2.3B |
| Task families | 500 |
| Screened domains | 267 |
| Configuration | Dev. | Acc. | Sens. |
|---|---|---|---|
| Stage 1 (no training) | 59.36 | – | 1.36 |
| Causal + gold labels | 60.42 | 87.01 | 1.25 |
| Non-causal + gold | 61.78 | 88.38 | 0.22 |
| Non-causal + soft targets | 62.59 | 89.16 | |
| + 30% format augmentation | 62.22 | 88.67 | – |
| Benchmark | GLiDE V1.2 | Best open | Jev 1.13 |
|---|---|---|---|
| JevBench (accuracy) | 93.07 | 93.94 | 86.58 |
| Clinical held-out | 97.01 | 96.56 | – |
| SciFact (macro-F1) | 85.63 | 84.05 | – |
| Regression watch | 84.88 | 84.10 | – |
| Model | Developer | v1 | v2 | v2.1 |
|---|---|---|---|---|
| Drex Max v1.1 | Nace.AI | 63.78 | 73.39 | 72.62 |
| SubQ Snap 1.2 | Subquadratic | – | – | 72.58 |
| GLiDE V1.2 | Fastino | 65.01 | 72.83 | 72.52 |
| Mammoth 27B | Baseten | – | 72.45 | 71.36 |
| d1-ultra | Liquid AI | 65.25 | 70.45 | 70.58 |
| SubQ Snap 1.1 | Subquadratic | 65.00 | 69.17 | 67.92 |
| Confidence at least | Answers kept (%) | Accuracy (%) |
|---|---|---|
| 0 (all answers) | 100.0 | 81.5 |
| 0.5 | 90.8 | 86.2 |
| 0.6 | 82.9 | 89.5 |
| 0.7 | 76.1 | 92.0 |
| 0.8 | 67.9 | 95.0 |
| 0.9 | 56.0 | 97.4 |
| Model | ECE | Conf. acc. | Wrong, conf. 0.95 | Brier |
|---|---|---|---|---|
| GLiDE V1.2 | 2.0 | +1.1 | 0.72 | 0.257 |
| StartLux v1.1 | 2.6 | +2.5 | 1.28 | 0.239 |
| mammoth-1 | 4.1 | +3.9 | 1.37 | 0.238 |
| Jeffan-31B | 4.7 | +4.7 | 1.64 | 0.249 |
| FlexDM | 4.4 | +3.5 | 0.25 | 0.355 |
| Jev 1.14 | 8.0 | +7.3 | 2.65 | 0.357 |
| Model | Params | Full | Latency | ECE | |
|---|---|---|---|---|---|
| GLiDE 27B | 26.1B | 73.37 | 122 ms | 2.1% | |
| GLiDE 9B | 7.9B | 62.48 | 86 ms | 1.7% | |
| GLiDE 2B | 1.9B | 43.72 | 37 ms | 6.6% | |
| GLiDE 0.8B | 0.75B | 36.36 | 37 ms | 7.6% | |
| GLiDE 0.4B | 0.39B | 33.84 | 18 ms | 4.9% |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Component | Base LM | GLiDE V1.2 |
|---|---|---|
| Input | Text prefix | Context, question, options |
| Full attention (16) | Causal | Bidirectional |
| Gated DeltaNet (48) | Causal | Causal |
| Readout | 248K token logits | 255 option logits |
| Prediction | Next token | Option probabilities |
| Forward passes | One per token | One per decision |
| Model | Developer | v1 | v2 | v2.1 |
|---|---|---|---|---|
| Drex Max v1.1 | Nace.AI | 63.78 | 73.39 | 72.62 |
| SubQ Snap 1.2 | Subquadratic | – | – | 72.58 |
| GLiDE V1.2 | Fastino | 65.01 | 72.83 | 72.52 |
| Mammoth 27B | Baseten | – | 72.45 | 71.36 |
| d1-ultra | Liquid AI | 65.25 | 70.45 | 70.58 |
| Decider v1.3 | Perplexity | – | – | 70.41 |
LLM Decision-Making
LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.
See all 279 papers