A closed-form capacity score picks architectures without training a model
single source· 1 articles · confidence: medium · first seen 2026-09-18 20:00 UTC
What this means for you
If you pick architectures by parameter count or by hand search, this is a claim to test, not a tool to adopt. The paper gives no implementation, no weights and no harness details, so the pruning result cannot be reproduced from the text alone.
A preprint proposes Neural Spectral Capacity, a closed-form score that ranks how much a network can represent from its architecture alone, with no model built, no data and no gradients. The score comes from the singular-value spectrum of each weight matrix — how strongly each layer stretches its inputs. Its authors report it ranks designs better than parameter count or FLOPs across seven Transformer and CNN families, and that a dynamic-programming solver using it trimmed LLaMA-7B to what the paper calls the best 5.7B model on eight commonsense reasoning tasks, about 5,900 times faster than the strongest training-free alternative. It is not peer-reviewed, and no evaluation date is given.
Key facts
- ·Neural Spectral Capacity is computed from the architectural specification alone, with no model instantiation, data or gradients, using the Marchenko-Pastur law under standard random initialisation. source
- ·On FlexiBERT, NSC scores 0.505 (Kendall tau) on architecture pairs whose parameter counts differ by less than 10%, where parameter count alone scores 0.082. source
- ·NSC-DP, a dynamic-programming solver using the score, found a Transformer-XL architecture on WikiText-103 that beats the human-designed baseline, in 2 seconds on a CPU. source
- ·The method prunes LLaMA-7B to a 5.7B model reported as best across eight commonsense reasoning tasks with no calibration data, about 5,900x faster than the strongest training-free proxy baseline. source
- ·NSC outperforms parameter count, FLOPs and representative training-free proxies at ranking architectures across seven Transformer and CNN families. source
What the sources say
- Hugging Face Daily Papers (research) — Preprint introducing a from-blueprint capacity metric and a solver that prunes LLaMA-7B.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersNeural Spectral Capacity: Measuring and Designing Architectures from Network Specification Alone2026-09-18