A prominent issue with such methods is reward over-optimization or reward hacking, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates.

We introduce ConfPO, a method for preference learning in Large Language Models (LLMs) that identifies and optimizes preference-critical tokens based solely on the training policy's confidence, without requiring any auxiliary models or compute. Unlike prior Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO), which uniformly adjust all token probabilities regardless of their relevance to preference, ConfPO focuses optimization on the most impactful tokens. This targeted approach improves alignment quality while mitigating overoptimization (i.e., reward hacking) by using the KL divergence budget more efficiently. In contrast to recent token-level methods that rely on credit-assignment models or AI annotators, raising concerns about scalability and reliability, ConfPO is simple, lightweight, and model-free. Experimental results on challenging alignment benchmarks, including AlpacaEval 2 and Arena-Hard, demonstrate that ConfPO consistently outperforms uniform DAAs across various LLMs, delivering better alignment with zero additional computational overhead.

Reinforcement Learning from Human Feedback (RLHF)has been crucial to the recent success of Large Language Models (LLMs), however it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimized the LLM. A prominent issue with such methods is reward over-optimization or reward hacking, where the performance as measured by the learned proxy reward model increases, but the true model quality plateaus or even deteriorates. Direct Alignment Algorithms (DDAs), such as Direct Preference Optimization (DPO) have emerged as alternatives to the classical RLHF pipeline. However, despite not training a separate proxy reward model or using RL, they still commonly deteriorate from over-optimization.

Direct Preference Optimization (DPO) demonstrates the advantage of aligning a large language model with human preference using only an offline dataset. However, DPO has the limitation that the KL penalty, which prevents excessive deviation from the reference model, is static throughout the training process. Several methods try to turn this static KL penalty into a dynamic one, but no approach can adaptively assign different KL penalties for each preference pair. In this paper, we propose $\varepsilon$-Direct Preference Optimization ($\varepsilon$-DPO), which allows adaptive control of the KL penalty strength $\beta$ for each preference pair. Specifically, $\varepsilon$-DPO adaptively controls $\beta$ for each preference pair based on the monotonicity of logits as a preference model under the perturbation of $\beta$ during training by simply reusing the logit of the current policy and the reference policy. Experimental results show that $\varepsilon$-DPO outperforms existing direct alignment algorithms and KL penalty relaxation methods on general chatbot benchmarks, highlighting the significance of adaptive KL penalty relaxation at the instance-level in DPO.

Direct Alignment Algorithms (DAAs) simplify language model alignment by replacing reinforcement learning (RL) and reward modeling (RM) in Reinforcement Learning from Human Feedback (RLHF) with direct policy optimization. DAAs can be classified by their ranking losses (pairwise vs. pointwise), by the rewards used in those losses (e.g., likelihood ratios of policy and reference policy, or odds ratios), or by whether a Supervised Fine-Tuning (SFT) phase is required (two-stage vs. one-stage). We first show that one-stage methods underperform two-stage methods. To address this, we incorporate an explicit SFT phase and introduce the $\beta$ parameter, controlling the strength of preference optimization, into single-stage ORPO and ASFT. These modifications improve their performance in Alpaca Eval 2 by +$3.46$ (ORPO) and +$8.27$ (ASFT), matching two-stage methods like DPO. Further analysis reveals that the key factor is whether the approach uses pairwise or pointwise objectives, rather than the specific implicit reward or loss function. These results highlight the importance of careful evaluation to avoid premature claims of performance gains or overall superiority in alignment algorithms.

Preference alignment in Large Language Models (LLMs) has significantly improved their ability to adhere to human instructions and intentions. However, existing direct alignment algorithms primarily focus on relative preferences and often overlook the qualitative aspects of responses. Striving to maximize the implicit reward gap between the chosen and the slightly inferior rejected responses can cause overfitting and unnecessary unlearning of the high-quality rejected responses. The unawareness of the reward scores also drives the LLM to indiscriminately favor the low-quality chosen responses and fail to generalize to responses with the highest rewards, which are sparse in data. To overcome these shortcomings, our study introduces reward-conditioned LLM policies that discern and learn from the entire spectrum of response quality within the dataset, helping extrapolate to more optimal regions. We propose an effective yet simple data relabeling method that conditions the preference pairs on quality scores to construct a reward-augmented dataset. This dataset is easily integrated with existing direct alignment algorithms and is applicable to any preference dataset. The experimental results across instruction-following benchmarks including AlpacaEval, MT-Bench, and Arena-Hard-Auto demonstrate that our approach consistently boosts the performance of DPO by a considerable margin across diverse models. Additionally, our method improves the average accuracy on various academic benchmarks. When applying our method to on-policy data, the resulting DPO model achieves SOTA results on AlpacaEval. Through ablation studies, we demonstrate that our method not only maximizes the utility of preference data but also mitigates the issue of unlearning, demonstrating its broad effectiveness beyond mere dataset expansion. Our code is available at https://github.com/shenao-zhang/reward-augmented-preference.

Reinforcement Learning from Human Feedback (RLHF) has been crucial to the recent success of Large Language Models (LLMs), however, it is often a complex and brittle process. In the classical RLHF framework, a reward model is first trained to represent human preferences, which is in turn used by an online reinforcement learning (RL) algorithm to optimize the LLM. A prominent issue with such methods is \emph{reward over-optimization} or \emph{reward hacking}, where performance as measured by the learned proxy reward model increases, but true quality plateaus or even deteriorates. Direct Alignment Algorithms (DDAs) like Direct Preference Optimization have emerged as alternatives to the classical RLHF pipeline by circumventing the reward modeling phase. However, although DAAs do not use a separate proxy reward model, they still commonly deteriorate from over-optimization. While the so-called reward hacking phenomenon is not well-defined for DAAs, we still uncover similar trends: at higher KL budgets, DAA algorithms exhibit similar degradation patterns to their classic RLHF counterparts. In particular, we find that DAA methods deteriorate not only across a wide range of KL budgets but also often before even a single epoch of the dataset is completed. Through extensive empirical experimentation, this work formulates and formalizes the reward over-optimization or hacking problem for DAAs and explores its consequences across objectives, training regimes, and model scales.