Allspark: Weak to Strong Transfer via Alternating Chain of Thought
Allspark transfers weak-model reasoning to stronger students via alternating chains of thought, including Kimi and Nemotron.
Allspark transfers reasoning improvements from a small weak model to a stronger model without using the strong model's rollouts during training. A weak teacher is trained beside a frozen copy; the two alternate chain-of-thought segments and the frozen model produces the final answer. At inference a stronger student replaces that partner while both models stay fixed, communicating only through text so families and tokenizers can differ. Controlled Qwen math and reasoning tests plus larger Inkling runs on ARC-AGI-2 show within-family and cross-family gains, including transfer to Kimi and Nemotron, with an accuracy-token tradeoff.
- A weak teacher alternates reasoning with a frozen copy, which gives the final answer.
- At inference a stronger student replaces the frozen partner; both stay fixed.
- Text-only communication supports different families and tokenizers.
- Inkling tests show within- and cross-family gains, including Kimi and Nemotron.
Full article187 words · extracted from huggingface.co · click to collapse
Recent progress in frontier models has renewed interest in large-scale reinforcement learning (RL), but the cost of generating large-model rollouts makes even testing RL recipes expensive. We ask whether reasoning improvements learned by a small, weak model can benefit a larger, stronger model without using the strong model's rollouts during training. We introduce Allspark, a training and inference framework for weak-to-strong transfer through alternating chains of thought. A weak teacher is trained alongside a frozen copy of the same model; the two alternate reasoning segments, and the frozen model produces the final answer. At inference time, a stronger student replaces the frozen training partner, while both models remain fixed. Because they communicate through text, the teacher can steer students from different model families and with different tokenizers. We study Allspark at two scales: controlled Qwen experiments across math and reasoning, and larger-scale Inkling experiments on ARC-AGI-2. The Inkling experiments show accuracy gains in within-family and cross-family settings, including transfer to Kimi and Nemotron, with benefits that vary across inference settings. These findings motivate reusing a trained weak teacher across strong students and examining the resulting accuracy--token tradeoff.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.32913