Lightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition

Lightning Weave. Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition. By Yecheng Wu and two others, from the Massachusetts Institute of Technology and NVIDIA. Published on the thirteenth of September. Its code repository is worked out: about one hundred and ninety source files, with a readme and a licence.

A core goal of efficient reasoning is to improve the accuracy-efficiency frontier. Improving reasoning accuracy and inference efficiency together is hard, though, because the two objectives can favour different reasoning behaviours. At the same time, models that were post-trained independently already offer distinct strengths, some in accuracy and some in efficiency. The strengths exist; they sit in different models. Pushing a single model toward one objective can pull it away from the other, which is why the two are hard to improve together.

Lightning Weave is a post-training framework that extracts those independently learned capabilities and composes them in a single student, through on-policy distillation. Each capability is represented by a policy shift: the difference between the model before post-training and the specialist that post-training produced. The specialist is not copied whole. What is carried over is the change that made it a specialist. One specialist contributes its gain in accuracy, another its gain in efficiency, and the student is trained to carry both.

The framework combines aligned log-ratio shifts at token states the student shares, and it uses Tilted-Target DOPD to convert the cached signals into a stable learning target. Each anchor pair scores the cached trajectories only once. After that, the student can be trained without serving several live anchor models at the same time, which is what keeps the composition practical. The costly part, scoring with the anchor models, happens once up front. Training the student afterwards works from the cache.

Across diverse student models and benchmarks in mathematics and code, Lightning Weave substantially improves on the base students and achieves a state-of-the- art accuracy-efficiency frontier. On Qwen3.5-4B, it raises HMMT two thousand twenty-five accuracy from fifty-nine point two percent to sixty-four percent, while using ten point seven percent fewer response tokens. Both results are on the same four billion parameter student.

On the same model, LiveCodeBench v5 accuracy rises from forty-one point seven percent to fifty-four point two percent, with nine point six percent fewer response tokens. In both cases accuracy goes up and length goes down together. Adjusting the relative strengths of the anchor signals yields a strong empirical Pareto frontier between the two. The code will be released soon.

Rather than training one model to be accurate and brief at once, Lightning Weave borrows each quality from a specialist that already has it. The trade-off between them then becomes a matter of how strongly each anchor is weighted, which the authors present as a new practical route to efficient reasoning. Composition, rather than a new objective, is the lever.