Allspark trains a small weak teacher with while a frozen copy of the same model provides alternating reasoning segments, producing the final answer at training time. At inference, the frozen partner is replaced by a stronger student, and the teacher continues to generate reasoning that steers the student, eliminating the need for strong-model rollouts during training. Experiments on Qwen models (math/reasoning) and Inkling models (ARC-AGI-2) show accuracy gains in within-family and cross-family transfers, including to Kimi and Nemotron, with benefits varying across inference settings. The results suggest that a single trained weak teacher can be reused to improve multiple strong models, reducing the cost of improving frontier reasoning.
Small models can steer stronger ones through shared reasoning traces
Allspark trains a weak model to contribute reasoning segments that improve larger, fixed models without using their outputs during training.
Big Tech
Kaizhao Liang · Junxiong Wang · Chen Liang · Zhendong Wang · Qiang Liu
UT Austin · Together AI · Microsoft
Research Digest··3 min read
In one line
Weak teacher trained with RL improves strong student accuracy by alternating reasoning chains without strong model rollouts.
Before you read this
Two ideas this paper builds onRead it as
Liang and colleagues introduce Allspark, a framework for transferring reasoning behavior from a reinforcement-learned small model to stronger models through alternating text-based chains of thought.
During training, two copies of the same small model alternately extend a shared reasoning trace.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.