Lightning Weave lifts Qwen3.5-4B maths accuracy while shortening its answers
single source· 1 articles · confidence: high · first seen 2026-09-12 20:00 UTC
What this means for you
Nothing to act on yet: no code, no weights, no release date, no independent reproduction. When it does ship, the thing worth testing is whether composing cheaply trained specialists beats training one larger model for your task.
A paper introduces Lightning Weave, a post-training method that folds several separately trained specialist models into a single student. It does this by on-policy distillation — training the student on its own outputs, scored against each specialist — rather than serving the specialists side by side when answering. On Qwen3.5-4B the authors report HMMT 2025 maths accuracy rising from 59.2% to 64.0% with 10.7% fewer response tokens, and LiveCodeBench v5 rising from 41.7% to 54.2% with 9.6% fewer. They claim a state-of-the-art accuracy-efficiency frontier across several student models. The paper gives no evaluation harness or scoring dates, and the code is not yet released.
Key facts
- ·Lightning Weave raises HMMT 2025 accuracy on Qwen3.5-4B from 59.2% to 64.0% while emitting 10.7% fewer response tokens, as reported by the paper's authors. source
- ·LiveCodeBench v5 accuracy on Qwen3.5-4B rises from 41.7% to 54.2% with 9.6% fewer response tokens in the same reported results. source
- ·The method composes independently post-trained specialists through on-policy distillation, caching each trajectory once per anchor pair so multiple teacher models need not be served simultaneously during student training. source
- ·Adjusting the relative strength of the anchor signals traces a Pareto frontier between accuracy and token use, according to the paper. source
- ·No code, weights or evaluation harness details are provided; the paper states code will be released soon. source
What the sources say
- Hugging Face Daily Papers (research) — Single paper introducing a distillation method; reports maths and code gains without an evaluation harness or code release.
Sources
The original reporting. Follow these — they did the work.
- Hugging Face Daily PapersLightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition2026-09-12