Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Convex Optimization for Alignment and Preference Learning on a Single GPU

About

Fine-tuning large language models (LLMs) to align with human preferences has driven the success of systems such as Gemini and ChatGPT. However, approaches like Reinforcement Learning from Human Feedback (RLHF) remain computationally expensive and complex. Direct Preference Optimization (DPO) offers a simpler alternative but has limitations such as inconsistent ranking accuracy, high dependence on GPU resources, and expensive hyperparameter tuning. We propose the Convex Optimization for Alignment and Preference Learning Algorithm (COALA): a novel lightweight strategy with strong theoretical guarantees. By leveraging the convex optimization reformulation of neural networks, COALA eliminates the need for a reference model and obtains significant reduction in both training time and VRAM consumption, thus enabling efficient training on a single GPU. Experiments across four datasets--including a 26621-sample synthetic Educational Feedback dataset--and six models (including Llama-3.1-8B) demonstrate COALA's competitive performance and efficiency while utilizing as little as ~17.6% of DPO's total TFLOPs. COALA exhibits stable, monotonically increasing rewards and reaches peak margins in significantly shorter time in comparison to traditional methods such as DPO and ORPO. To the best of our knowledge, this is the first time convex optimization has been effectively applied to preference fine-tuning of LLMs.

Miria Feng, Mert Pilanci• 2026

Related benchmarks

TaskDatasetResultRank
LLM AlignmentEduFeedback
TFLOPs152.6
20
Multi-turn Conversation EvaluationUltraFeedback
MT-Bench Score6.1
20
Multi-turn Conversation EvaluationIMDB
MT-Bench Score5.6
20
Multi-turn Conversation EvaluationEduFeedback alternate
MT-Bench Score8.1
20
LLM AlignmentAlpacaEval EDU 2.0 (test)
LC WR40.9
12
LLM AlignmentAlpacaEval IMDB 2.0 (test)
LC WR39.72
12
LLM AlignmentAlpacaEval ULTRA 2.0 (test)
LC WR (%)31.58
12
Preference AlignmentEDU
Human Win Rate39.1
8
Preference AlignmentIMDB
Human Win Rate42.7
8
Preference AlignmentIMDB (test)
Pairwise Win Rate (Excl. Ties)60.91
5
Showing 10 of 13 rows

Other info

Follow for update