One Model for All: Multi-Objective Controllable Language Models

TMLR Paper5684 Authors

Abstract: Aligning large language models (LLMs) with human preferences is critical to enhancing LLMs' safety, helpfulness, humor, faithfulness, and other desirable properties. Current reinforcement learning from human feedback (RLHF) mainly focuses on a fixed reward learned from average human ratings, which may weaken the adaptability and controllability of varying preferences. However, creating personalized LLMs requires aligning LLMs with individual human preferences, which is non-trivial due to the scarce data per user and the diversity of user preferences in multi-objective trade-offs, such as prioritizing humor and empathy in one context, while seeking efficiency and precision in another. Can we train one LLM to produce personalized outputs for different user preferences on the Pareto front? In this paper, we introduce Multi-Objective Control (MOC), which trains a single LLM to directly generate responses in the preference-defined regions of the Pareto front. Our approach introduces multi-objective optimization (MOO) principles into RLHF to train an LLM as a preference-conditioned policy network. We improve the computational efficiency of MOC by applying MOO at the policy level, enabling us to fine-tune a 7B-parameter model on a single A6000 GPU. Extensive experiments demonstrate the advantages of MOC over baselines in three aspects: (i) controllability of LLM outputs with respect to user preferences on the trade-off among multiple rewards; (ii) quality and diversity of LLM outputs, measured by the hyper-volume of multiple solutions achieved; and (iii) generalization to unseen preferences. These results highlight MOC’s potential for real-world applications requiring scalable and customizable LLMs.

Submission Length: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Sungwoong_Kim2

Submission Number: 5684