Overview
Lightning Weave is introduced as a post-training framework designed to enhance the accuracy and inference efficiency of reasoning models by composing independently learned capabilities. The framework addresses the challenge of jointly improving these two objectives, which can favor distinct reasoning behaviors.
Research Context
Efficient reasoning aims to improve the accuracy-efficiency frontier. Independently post-trained models are noted to possess distinct strengths in terms of accuracy and efficiency. The framework's core premise involves leveraging these pre-existing specialist capabilities.
Approach
Lightning Weave functions by extracting and composing independently learned capabilities within a single student model. This process occurs through on-policy distillation. Each acquired capability is represented as the policy shift from the model prior to post-training to the resulting specialist model. The framework combines aligned log-ratio shifts at shared student token states. To convert cached signals into a stable learning target, Lightning Weave utilizes Tilted-Target DOPD. A key operational aspect is that each anchor pair scores cached trajectories only once, which enables subsequent student training without requiring multiple live anchor models to be served concurrently.
Findings
Across diverse student models and benchmarks in mathematics and code, Lightning Weave demonstrated substantial improvements over base students and achieved a state-of-the-art accuracy-efficiency frontier. Specific results include:
- On Qwen3.5-4B, Lightning Weave raised HMMT 2025 accuracy from 59.2% to 64.0%. This improvement was accompanied by a 10.7% reduction in response tokens.
- For the same model, LiveCodeBench v5 accuracy increased from 41.7% to 54.2%. This was achieved with 9.6% fewer response tokens.
Empirical observations indicated that adjusting the relative strengths of the anchor signals yields a strong empirical accuracy-efficiency Pareto frontier.
Why This Matters
These results establish Lightning Weave as a new practical route to efficient reasoning through the composition of capabilities. The framework's ability to improve both accuracy and reduce response tokens suggests potential for more efficient deployment and operation of reasoning models.