GLiDE: distilled single-pass decision models hit 73.37 on HF Decision Index
kalyan_kpl · x · 2026-11-22
GLiDE is a family of single-pass decision models (0.4B–27B) distilled from reasoning teachers:
- Architecture: modified Qwen hybrid with bidirectional full-attention layers and a dedicated option readout to jointly process context, questions, and candidate answers.
- Training: distilled from multiple teachers on 1.6M+ questions across 500+ task families.
- Novelty: parameter-free recurrent depth adds inference-time compute by repeatedly applying selected layers — no extra parameters or tokens.
- Results: GLiDE V1.2 27B scores 71.52 on the Hugging Face Decision Index with 2.0% ECE and 92.9 ms median latency; an extended 27B variant reaches 73.37 with 2.1% ECE.
More from Research
- Gordic: interpretability of looped transformers likely no harder than deep models — gordic_aleksa · 2026-11-22
- Soon the interesting bit in math is minimizing Kolmogorov complexity of proofs — gordic_aleksa · 2026-11-22
- Russian philosophy journal tackles 'Do LLMs think?' with a computational answer — abenitezburraco · 2026-11-22
- OpenAI's own data suggests hidden CoT barely matters for distillation at medium reasoning effort — adam_karvonen · 2026-11-22
- Sampling the SK Model with Vanishing TV Error Now Extends to the Conjectured β<1 — LucaAmb · 2026-11-22
- Speed-to-power dominates: on-site solar-plus-storage cheapest for AI data centers in 85% of US locations — mattbeane · 2026-11-22