Fusion Training for Mathematical Generalization in Large Language Models
A systematic study of Thinking Mode Fusion (TMF) analyzes training schedules and data ratios between thinking and non-thinking modes for mathematical problem solving. Increasing non-thinking supervision reduces thinking-mode accuracy, revealing an asymmetric interaction. Different training schedules modulate this trade-off, and the optimal schedule depends on the data ratio. A negative correlation between non-thinking and thinking mode supervision highlights inherent tension.
Thinking Mode Fusion (TMF) unifies concise and long-form reasoning in a single large language model. This study systematically examines TMF training dynamics, focusing on data ratio and training schedule between thinking and non-thinking modes. Using a mathematical problem-solving benchmark, experiments show that higher non-thinking supervision degrades thinking-mode accuracy, and training schedules can modulate this trade-off. The findings quantify a negative correlation between the two modes, offering practical guidance for TMF training design.
The asymmetric interaction between thinking and non-thinking modes suggests that joint optimization is non-trivial; the negative correlation implies a Pareto frontier where improving one mode degrades the other. Future work may explore adaptive scheduling or architectural modifications to mitigate this tension.
For AI developers aiming to deploy versatile models that handle both quick responses and deep reasoning, these findings highlight the need for careful data balancing and schedule tuning. Products requiring reliable mathematical reasoning may need to prioritize thinking-mode performance at the cost of non-thinking efficiency.
Optimizing TMF can lead to more efficient training of multi-capability models, reducing the need for separate specialized models. This can lower deployment costs and improve user experience in applications like tutoring, coding assistants, and enterprise analytics where both quick answers and detailed reasoning are required.
Next signals include empirical validation on broader reasoning tasks beyond math, exploration of dynamic data mixing during training, and potential integration with reinforcement learning from human feedback to align mode selection with user intent.