AI Insight
This study introduces CORE (Collaborative Reasoning), a framework where pairs of language models learn to work together during training by teaching each other. When one model solves a problem successfully, it provides hints to help its peer model that failed on the same problem, creating a two-stage learning process. Testing on mathematical reasoning tasks, a pair of small models (3B and 4B parameters) achieved 99.54% accuracy on GSM8K and 92.08% on MATH with only 1,000 training examples, substantially outperforming models trained individually.
Why it matters
This approach demonstrates that smaller AI models working collaboratively can match or exceed the performance of larger individual models, potentially reducing computational costs and energy consumption in AI deployment. The method could make advanced reasoning capabilities more accessible by enabling effective use of smaller, more efficient models.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Abstract: Large language models exhibit complementary reasoning errors: on the same instance, one model may succeed with a particular decomposition while another fails. We propose Collaborative Reasoning (CORE), a training-time collaboration framework that converts peer success into a learning signal via a cross-teaching protocol. Each problem is solved in two stages: a cold round of independent sampling, followed by a contexted rescue round in which models that failed receive hint extracted from a successful peer. CORE optimizes a combined reward that balances (i) correctness, (ii) a lightweight DPP-inspired diversity term to reduce error overlap, and (iii) an explicit rescue bonus for successful recovery. We evaluate CORE across four standard reasoning datasets GSM8K, MATH, AIME, and GPQA. With only 1,000 training examples, a pair of small open source models (3B+4B) reaches Pass@2 of 99.54% on GSM8K and 92.08% on MATH, compared to 82.50% and 74.82% for single-model training. On harder datasets, the 3B+4B pair reaches Pass@2 of 77.34% on GPQA (trained on 348 examples) and 79.65% on AIME (trained on 792 examples), using a training-time budget of at most 1536 context tokens and 3072 generated tokens. Overall, these results show that training-time collaboration can reliably convert model complementarity into large gains without scaling model size.