AIR: Adaptive Interleaved Reasoning with Code in MLLMs

The paper introduces AIR (Adaptive Interleaved Reasoning with Code) to enhance multimodal large language models (MLLMs) by integrating adaptive interleaved reasoning capabilities for complex numerical computations. It outlines a three-component solution involving a cold-start data construction pipeline, data filtering for reinforcement learning (RL) dataset curation, and an adaptive tool-invocation strategy using a group-constrained reward function, resulting in an average performance improvement of 6.1 percentage points on benchmarks, with interleaved reasoning accuracy increasing by 9.9 percentage points. This advancement is significant for practitioners as it enables MLLMs to effectively handle numerical tasks and tool-use scenarios, expanding their practical applications in complex problem-solving.