Low-bit quantization is a common way to reduce the cost of serving large reasoning models, but it comes with a hidden penalty: long chains of thought become noticeably less reliable. The problem, according to a new arXiv paper, is that each reasoning step reads cached keys and values spanning thousands of tokens, and quantization errors from those reads pile up over time.

The paper introduces RECAP, a calibration approach that explicitly accounts for the history a model has already seen. By focusing on the cached context that the model actually reads, RECAP aims to keep errors from compounding during extended reasoning.

This is a single-source report based on the abstract only; details of the calibration mechanism and experimental results are not yet available. The work highlights a practical trade-off between efficiency and accuracy in low-bit reasoning models, and suggests that history-aware calibration may be a way to ease that tension.