Collaborative smaller models improve reasoning and efficiency in large tasks
Collaborative Learning and Reasoning over Model Slices
Artificial Intelligence
Summary
Large language models can make mistakes when solving complex problems step-by-step because they rely on their own earlier answers. The authors found that splitting a big model into smaller, different models that work together lets these models double-check each other’s reasoning. This teamwork helps catch errors and improves performance on tasks like math problems, code writing, and answering tricky questions. Their approach matches or beats using one big model or other multi-agent methods, while also running faster.
What this means in practice
- •For software developers: Use collaborative smaller models to improve code generation quality and catch errors during development.
- •For automated customer support teams: Enhance multi-step question answering systems by integrating model slices that validate and refine responses for better accuracy.
Authors
Yuheng Wang, Runde Yang, Zhuohan Bao, Qiran Zhang, Huatao Li, Yiyang Zheng, Lin Wu, Jingru Fan, Shu Yao, Yihan Li, Zhi Li, Chen Qian
Abstract
Large language models (LLMs) excel across diverse tasks, benefiting from knowledge priors, in-context reasoning, and test-time computation. Despite these strengths, single-model reasoning remains vulnerable to error propagation, as later steps depend on the model's earlier outputs, leaving reasoning blind spots difficult to independently reassess. To address this limitation, we develop \textit{collaborative learning}, encompassing the construction of heterogeneous collaborators and the learning of effective cross-model collaboration. Specifically, \textit{model slicing} replaces a large model with smaller, heterogeneous models under a comparable parameter budget, enabling prior reasoning to be reassessed across distinct model parameterizations. \textit{Pyramid supervision} combines protocol distillation with step-level supervision to refine reasoning and interaction decisions throughout collaboration. We evaluate our approach across four task categories: multi-hop question answering, mathematical reasoning, code generation, and open-ended constrained generation. The trained heterogeneous models achieve comparable or better task performance than the original large model and established multi-agent baselines, including reinforcement-learning-based methods, while also achieving higher inference efficiency than the multi-agent baselines and exhibiting differentiated collaboration patterns across tasks. These findings show that learned collaboration turns model heterogeneity into complementary reasoning capabilities.