A Zeroth-Order Paradigm for LLM Preference Alignment
ComPO is a zeroth-order preference alignment method using comparison oracles to mitigate likelihood displacement across Mistral, Llama, Gemma, and Qwen3 models.
The paper proposes Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method that extracts directional information from preference pairs with small likelihood margins without directly optimizing a differentiable preference loss. The authors prove convergence guarantees for the offline scheme and performance guarantees for a constrained online variant with reverse-KL control. Experiments on Mistral, Llama, Gemma-2, Qwen3, and Gemma-3 show improvements over existing direct alignment methods, including length-controlled win rates, with pair-level diagnostics consistent with mitigating likelihood displacement.
- Likelihood displacement motivates alternatives to direct preference losses.
- Online ComPO uses unlabeled policy generations for reverse-KL control.
- Guarantees assume smoothness, gradient sparsity, and oracle-objective compatibility.
- Coverage-based analysis supports the constrained scheme.
Full article176 words · extracted from huggingface.co · click to collapse
Direct preference alignment methods are widely used to align large language models (LLMs) with human preferences because of their computational and memory efficiency. However, likelihood displacement motivates alternative ways to extract information from preference pairs with small likelihood margins. In this paper, we propose and analyze Comparison-based Preference Optimization (ComPO), a zeroth-order alignment method based on comparison oracles. ComPO extracts directional information from these pairs without directly optimizing a differentiable preference loss on them. We establish a convergence guarantee for its basic offline scheme under smoothness, gradient sparsity, and compatibility between the oracle and a latent objective. We further introduce online ComPO, which retains the offline comparison mechanism and uses unlabeled policy generations for reverse-KL control relative to a reference policy. Following the coverage perspective of preference fine-tuning, we establish a performance guarantee for a basic constrained scheme under local coverage and in-distribution pairwise reward accuracy. Experiments on Mistral, Llama, Gemma-2, Qwen3, and Gemma-3 models demonstrate improvements over existing direct alignment methods, including length-controlled win rates, with pair-level diagnostics providing evidence consistent with mitigating likelihood displacement.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.19144