SAIL: Self-Improving Efficient Online Alignment of Large Language Models (2406.15567v1)

Abstract: Reinforcement Learning from Human Feedback (RLHF) is a key method for aligning LLMs with human preferences. However, current offline alignment approaches like DPO, IPO, and SLiC rely heavily on fixed preference datasets, which can lead to sub-optimal performance. On the other hand, recent literature has focused on designing online RLHF methods but still lacks a unified conceptual formulation and suffers from distribution shift issues. To address this, we establish that online LLM alignment is underpinned by bilevel optimization. By reducing this formulation to an efficient single-level first-order method (using the reward-policy equivalence), our approach generates new samples and iteratively refines model alignment by exploring responses and regulating preference labels. In doing so, we permit alignment methods to operate in an online and self-improving manner, as well as generalize prior online RLHF methods as special cases. Compared to state-of-the-art iterative RLHF methods, our approach significantly improves alignment performance on open-sourced datasets with minimal computational overhead.

SAIL: Self-Improving Efficient Online Alignment of LLMs

This paper introduces SAIL (Self-Improving Efficient Online Alignment) as a sophisticated framework for the online alignment of LLMs. Existing methodologies for Reinforcement Learning from Human Feedback (RLHF) largely depend on offline datasets of fixed preferences, which can lead to suboptimal performance due to insufficient coverage and misalignment of response-query pairs. The paper critiques the current literature's dual focus on both offline alignment approaches, such as Direct Preference Optimization (DPO), Iterative Preference Optimization (IPO), and Self-Labeled Inner-loop Critic (SLiC), and nascent online RLHF methods for their lack of a unified conceptual framework and issues related to distribution shift.

Key Contributions

1. Bilevel Optimization Framework: The authors argue that online LLM alignment can be formalized within a bilevel optimization framework. This approach incorporates the interdependencies between the reward learning phase and policy optimization phase. The bilevel formulation ensures proper characterization of these dependencies, which is often ignored in existing methodologies. The proposed approach reduces bilevel optimization to a single-level first-order method through the reward-policy equivalence.
2. Self-Improvement Mechanisms: SAIL introduces self-improving mechanisms to reduce dependency on oracle preference functions by iteratively refining model alignment through emergent responses and regulated preference labels. This facilitates alignment methods to work in an online and self-improving manner, enhancing their ability to generalize previous online RLHF methods.
3. Adaptive Direct Preference Optimization: By leveraging the reward-policy equivalence, the paper proposes a new unified optimization framework that mitigates distribution shift issues through efficient single-level DPO-style analysis. This approach extends the adaptive direct preference optimization to harness the LLM policy itself to generate preference feedback, therefore, reducing exhaustive dependency on human-generated datasets.
4. Experimental Validation: The proposed SAIL framework demonstrates substantial improvements over state-of-the-art iterative RLHF methods in terms of alignment performance. The authors evaluated it on open-sourced datasets and reported significantly better results with minimal computational overhead.

Methodology

The work hinges on transforming the bilevel optimization problem, representing the entanglement between reward learning and policy update, into an efficient single-level optimization problem. This mathematical transformation relies on the unique equivalence between the reward function and the LLM policy. The paper provides detailed theoretical foundations for this equivalence and derives gradient expressions that are computationally efficient to evaluate.

• Gradient Evaluation: The gradient terms derived for the optimization problem focus on both the existing and newly introduced terms due to the self-improving preference optimization framework. The added gradients involve the log-probabilities of the response policies, facilitating their efficient computation within existing DPO pipelines.
• Self-Improving LLMs: The proposed self-improving mechanism operates under the mixture distribution approach, which combines generated responses and preferences from both offline datasets and the policy itself. This mitigates the reliance on a preference oracle and iteratively improves model performance while maintaining computational tractability.

Experimental Results

The paper’s experiments focus on two primary aspects: improving DPO training efficiency and performance, and practical application of SAIL to state-of-the-art LLM alignment.

• Experimental Setup: The authors conducted experiments on models such as Qwen1.5-0.5B, Phi-3 (3.8B), and Llama-3 (8B) using datasets like PKU-SafeRLHF and UltraFeedback. They assessed performance improvements in reward margin, evaluation reward, pairwise winrate, and MT-Bench scores.
• Findings: SAIL variants, including DDP (Dataset-Driven Preference), DPP (Policy-Driven Preference), and DPR (Reward-Driven Preference), consistently outperformed the standard DPO in winrate and evaluation reward with varying computational overheads. The SAIL-DPP and SAIL-DPR designs showed robust performance across different evaluation criteria, especially in mitigating distribution shift issues.

Implications and Future Work

The theoretical implications of this research are significant as it presents a unified framework for RLHF, underpinning various online and offline methods as special cases of the proposed bilevel optimization structure. Practically, SAIL's self-improving mechanism could lead to more efficient and reliable alignment processes, which is critical given the growing deployment of LLMs in real-world applications.

Future developments could explore alternative utility functions for preference modeling beyond the Bradley-Terry framework. Additionally, scaling evaluations to even larger models may provide more comprehensive insights into SAIL's advantages, thus further bolstering the framework's robustness and applicability in diverse AI use-cases.

In conclusion, the paper provides a valuable contribution to the field of LLM alignment, presenting an innovative and efficient framework that addresses key limitations of existing methodologies while opening avenues for future research.