![[2602.19049] IAPO: Information-Aware Policy Optimization for Token-Efficient Reasoning](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19049] IAPO: Information-Aware Policy Optimization for Token-Efficient Reasoning
Summary
The paper presents IAPO, a novel framework for token-efficient reasoning in large language models, enhancing accuracy while reducing inference time by optimizing token-wise advantages based on mutual information.
Why It Matters
As large language models grow in complexity, the need for efficient reasoning methods becomes critical. IAPO addresses the challenge of balancing reasoning accuracy with token efficiency, making it relevant for researchers and practitioners in AI and machine learning fields focused on improving model performance without increasing computational costs.
Key Takeaways
- IAPO optimizes token efficiency by assigning advantages based on conditional mutual information.
- The framework reduces reasoning verbosity by up to 36% without compromising accuracy.
- Empirical evaluations show IAPO outperforms existing token-efficient reinforcement learning methods.
- The approach provides a principled mechanism for identifying informative reasoning steps.
- IAPO represents a significant advancement in post-training methods for large language models.
Computer Science > Computation and Language arXiv:2602.19049 (cs) [Submitted on 22 Feb 2026] Title:IAPO: Information-Aware Policy Optimization for Token-Efficient Reasoning Authors:Yinhan He, Yaochen Zhu, Mingjia Shi, Wendy Zheng, Lin Su, Xiaoqing Wang, Qi Guo, Jundong Li View a PDF of the paper titled IAPO: Information-Aware Policy Optimization for Token-Efficient Reasoning, by Yinhan He and 7 other authors View PDF HTML (experimental) Abstract:Large language models increasingly rely on long chains of thought to improve accuracy, yet such gains come with substantial inference-time costs. We revisit token-efficient post-training and argue that existing sequence-level reward-shaping methods offer limited control over how reasoning effort is allocated across tokens. To bridge the gap, we propose IAPO, an information-theoretic post-training framework that assigns token-wise advantages based on each token's conditional mutual information (MI) with the final answer. This yields an explicit, principled mechanism for identifying informative reasoning steps and suppressing low-utility exploration. We provide a theoretical analysis showing that our IAPO can induce monotonic reductions in reasoning verbosity without harming correctness. Empirically, IAPO consistently improves reasoning accuracy while reducing reasoning length by up to 36%, outperforming existing token-efficient RL methods across various reasoning datasets. Extensive empirical evaluations demonstrate that information...