Switch-Reasoner: Learn When to Think in Multitask Mixtures via Reinforcement Learning

2026-07-09

AI Summary

This work addresses the inefficiency of current large language models in multimodal, multitask settings, where the rigid “think-then-answer” paradigm leads to redundant reasoning and unstable post-training adaptation for deciding when to invoke explicit reasoning. To overcome this, the authors propose Switch-Reasoner, a framework that formulates reasoning as a virtual tool call, enabling the model to adaptively choose between direct answering and explicit reasoning. The approach introduces a two-level control mechanism—sample-level supervision based on relative reasoning gains and global balancing of reasoning mode usage—combined with GRPO-based reinforcement learning. This is the first method to achieve stable, adaptive reasoning control in multimodal multitask scenarios. Experiments across 11 tasks demonstrate that Switch-Reasoner significantly reduces unnecessary reasoning while maintaining high performance, yielding a superior accuracy–efficiency trade-off.

Abstract

Multimodal Large Language Models (MLLMs) often follow a fixed Think-then-Answer paradigm, which is inefficient in heterogeneous multitask settings because simple inputs may not require explicit reasoning while difficult ones can benefit substantially from it. Learning when to think is also unstable during post-training, where imbalanced rollouts can drive the model toward always-thinking or always-direct behavior. We propose Switch-Reasoner, a GRPO-based framework that learns to adaptively select reasoning modes for MLLMs. It treats thinking as a virtual tool invocation and allows the model to either answer directly or invoke explicit reasoning before answering. To stabilize this decision, we introduce a dual-level regulation mechanism that balances the overall use of Thinking Mode and Direct Mode while providing sample-level supervision based on the relative benefit of the two choices. Experiments on 11 multimodal tasks show that Switch-Reasoner reduces unnecessary reasoning while maintaining strong performance, achieving a better accuracy-efficiency trade-off.

Problem

• multitask mixtures
• reasoning efficiency
• thinking paradigm
• post-training instability
• multimodal reasoning

Innovation

• Switch-Reasoner
• adaptive reasoning
• multimodal LLMs
• reinforcement learning
• reasoning mode selection