Oracle-Robust Online Alignment for Large Language Models
This work is significant for researchers and practitioners aligning large language models, as it provides a method to improve robustness against misspecified preference feedback, a common challenge in real-world applications.
This paper addresses the online alignment of large language models when preference feedback is misspecified, meaning the observed oracle differs from the true one. The authors formulate an oracle-robust online alignment objective as a worst-case optimization problem, which for log-linear policies decomposes into the original loss plus a sensitivity penalty. They develop projected stochastic composite updates and prove an oracle complexity of \u00d5(\u03b5^-2) for approximate stationarity.
We study online alignment of large language models under misspecified preference feedback, where the observed preference oracle deviates from an ideal but unknown ground-truth oracle. The online LLM alignment problem is a bi-level reinforcement problem due to the coupling between data collection and policy updates. Recently, the problem has been reduced to tractable single-level objective in the SAIL (Self-Improving Efficient Online Alignment) framework. In this paper, we introduce a pointwise oracle uncertainty set in this problem and formulate an oracle-robust online alignment objective as a worst-case optimization problem. For log-linear policies, we show that this robust objective admits an exact closed-form decomposition into the original loss function plus an explicit sensitivity penalty. We develop projected stochastic composite updates for the resulting weakly convex objective and prove $\widetilde{O}(\varepsilon^{-2})$ oracle complexity for reaching approximate stationarity.