ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping

Shuang Chen, Yue Guo, Yimeng Ye, Shijue Huang, Wenbo Hu, Haoxi Li, Manyuan Zhang, Jiayu Chen, Song Guo, and Nanyun Peng, in Proceedings of the International Conference on Learning Representations (ICLR), 2026.

Download the full text

Abstract

The paper presents a unified open-source framework designed to enhance the adaptive reasoning capabilities of Multimodal Large Reasoning Models (MLRMs). It addresses the imbalance in how MLRMs tackle problems by dynamically allocating exploration effort based on the difficulty of the task. The framework uses High window-entropy (HWE) tokens to identify critical reasoning moments and implements a two-stage training pipeline: Adaptive Cold-Start (AdaCS) Fine-Tuning and Adaptive Entropy Policy Optimization (AEPO).

Bib Entry

@inproceedings{chen2026ares,
  title = {ARES: Multimodal Adaptive Reasoning via Difficulty-Aware Token-Level Entropy Shaping},
  author = {Chen, Shuang and Guo, Yue and Ye, Yimeng and Huang, Shijue and Hu, Wenbo and Li, Haoxi and Zhang, Manyuan and Chen, Jiayu and Guo, Song and Peng, Nanyun},
  booktitle = {Proceedings of the International Conference on Learning Representations (ICLR)},
  year = {2026}
}