Adaptive Decoding via Latent Preference Optimization
This addresses the limitation of one-size-fits-all temperature settings in language model decoding for general instruction following, though it appears incremental as it builds on existing temperature sampling concepts.
The paper tackles the problem of language models using fixed temperature sampling for both creative and factual tasks by introducing Adaptive Decoding, which dynamically selects temperature at inference time, and Latent Preference Optimization to train it. The method outperforms fixed temperatures on tasks like UltraFeedback, Creative Story Writing, and GSM8K.
During language model decoding, it is known that using higher temperature sampling gives more creative responses, while lower temperatures are more factually accurate. However, such models are commonly applied to general instruction following, which involves both creative and fact seeking tasks, using a single fixed temperature across all examples and tokens. In this work, we introduce Adaptive Decoding, a layer added to the model to select the sampling temperature dynamically at inference time, at either the token or example level, in order to optimize performance. To learn its parameters we introduce Latent Preference Optimization (LPO) a general approach to train discrete latent variables such as choices of temperature. Our method outperforms all fixed decoding temperatures across a range of tasks that require different temperatures, including UltraFeedback, Creative Story Writing, and GSM8K.