Discovering Preference Optimization Algorithms with and for Large Language Models

TL;DR

This work iteratively prompts an LLM to propose and implement new preference optimization loss functions based on previously-evaluated performance metrics, which leads to the discovery of previously-unknown and performant preference optimization algorithms.

Abstract

Offline preference optimization is a key method for enhancing and controlling the quality of Large Language Model (LLM) outputs. Typically, preference optimization is approached as an offline supervised learning task using manually-crafted convex loss functions. While these methods are based on theoretical insights, they are inherently constrained by human creativity, so the large search space of possible loss functions remains under explored. We address this by performing LLM-driven objective discovery to automatically discover new state-of-the-art preference optimization algorithms without (expert) human intervention. Specifically, we iteratively prompt an LLM to propose and implement new preference optimization loss functions based on previously-evaluated performance metrics. This process leads to the discovery of previously-unknown and performant preference optimization algorithms. The best performing of these we call Discovered Preference Optimization (DiscoPOP), a novel algorithm that adaptively blends logistic and exponential losses. Experiments demonstrate the state-of-the-art performance of DiscoPOP and its successful transfer to held-out tasks.

Figures (14)

• Figure 1: Left. Conceptual illustration of LLM-driven discovery of objective functions. We prompt an LLM to output new code-level implementations of offline preference optimization losses $\mathbb{E}_{(y_w,y_l,x)\sim \mathcal{D}}\left[f\left(\beta \rho \right)\right]$ as a function of the policy ($\pi_\theta$) and reference model's ($\pi_\text{ref}$) likelihoods of the chosen ($y_w$) and rejected ($y_l$) completions. Afterwards, we run an inner loop training procedure and evaluate the resulting model on MT-Bench. The corresponding performance is fed back to the language model, and we query it for the next candidate. Right. Performance of discovered objective functions on Alpaca Eval.
• Figure 2: LLM-driven objective discovery for CIFAR-10 classification. Left. Performance across LLM-discovery trials. The proposals alternate between exploring new objective concepts, tuning the components, and combining previous insights. Right. The best three discovered objectives transfer to different network architectures and longer training runs (100 epochs).
• Figure 3: Examples of LLM Objective Discovery improvement across generations. The first and second runs are shown left and right respectively.
• Figure 4: MT-Bench Discovered Objective Evaluations
• Figure 5: Frontiers of expected reward vs KL divergence for converging models for the LRML against DPO and SLiC objective function. The rewards and KL-divergence values are averaged over 10 generations with different seeds. The sweep is done over $\beta \in \{0.025, 0.05, 0.1, 0.25, 0.5, 1.0\}$. The optimal point is the top left corner, where the perfect reward is achieved with minimal divergence from the reference model.