Clear Preferences Leave Traces: Reference Model-Guided Sampling for Preference Learning
This addresses the need for efficient preference learning in language model alignment, offering a resource-saving method that is incremental but impactful for technical domains.
The paper tackles the problem of improving Direct Preference Optimization (DPO) by identifying high-quality training samples without extra resources, achieving consistent improvements of +0.1 to +0.4 on MT-Bench with 30-50% less data and substantial gains of +0.4 to +0.98 on technical tasks.
Direct Preference Optimization (DPO) has emerged as a de-facto approach for aligning language models with human preferences. Recent work has shown DPO's effectiveness relies on training data quality. In particular, clear quality differences between preferred and rejected responses enhance learning performance. Current methods for identifying and obtaining such high-quality samples demand additional resources or external models. We discover that reference model probability space naturally detects high-quality training samples. Using this insight, we present a sampling strategy that achieves consistent improvements (+0.1 to +0.4) on MT-Bench while using less than half (30-50%) of the training data. We observe substantial improvements (+0.4 to +0.98) for technical tasks (coding, math, and reasoning) across multiple models and hyperparameter settings.