Logits are All We Need to Adapt Closed Models

• DOI: 10.48550/arxiv.2502.06806
• Publication Date: 2025-02-03
• Authors: Gaurush Hiranandani; H. Z. Wu; Sourav Mukherjee; Sanmi Koyejo
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Computer Science - Computation and Language; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Computation and Language (cs.CL); Machine Learning (cs.LG)

Many commercial Large Language Models (LLMs) are often closed-source, limiting developers to prompt tuning for aligning content generation with specific applications. While these models currently do not provide access token logits, we argue that if such were available, it would enable more powerful adaptation techniques beyond engineering. In this paper, propose a token-level probability reweighting framework that, given logits and small amount of task-specific data, can effectively steer black-box LLMs toward application-specific generation. Our approach views next-token prediction through the lens supervised classification. We show data be formulated as label noise correction problem, leading \emph{Plugin} model -- an autoregressive operates solely on logits. theoretical justification why alone is sufficient task adaptation. Extensive experiments multiple datasets, LLMs, demonstrate effectiveness our method, advocating broader in closed-source models.