Understanding Likelihood Over-optimisation in Direct Alignment Algorithms

• URL: http://arxiv.org/abs/2410.11677v2
• Date: Fri, 18 Oct 2024 09:41:53 GMT
• Title: Understanding Likelihood Over-optimisation in Direct Alignment Algorithms
• Authors: Zhengyan Shi, Sander Land, Acyr Locatelli, Matthieu Geist, Max Bartolo,
• Abstract summary: Direct Alignment Algorithms (DAAs) have emerged as alternatives to online Reinforcement Learning from Human Feedback. These algorithms aim to increase the likelihood of generating better (preferred) completions while discouraging worse (non-preferred) ones. This work explores the relationship between completion likelihood and model performance in state-of-the-art DAAs.
• Score: 20.043560907227018
• License:
• Abstract: Direct Alignment Algorithms (DAAs), such as Direct Preference Optimisation (DPO) and Identity Preference Optimisation (IPO), have emerged as alternatives to online Reinforcement Learning from Human Feedback (RLHF) algorithms such as Proximal Policy Optimisation (PPO) for aligning language models to human preferences, without the need for explicit reward modelling. These methods generally aim to increase the likelihood of generating better (preferred) completions while discouraging worse (non-preferred) ones, while staying close to the original model's behaviour. In this work, we explore the relationship between completion likelihood and model performance in state-of-the-art DAAs, and identify a critical issue of likelihood over-optimisation. Contrary to expectations, we find that higher likelihood of better completions and larger margins between better and worse completion likelihoods do not necessarily lead to better performance, and may even degrade it. Our analysis reveals that while higher likelihood correlates with better memorisation of factual knowledge patterns, a slightly lower completion likelihood tends to improve output diversity, thus leading to better generalisation to unseen scenarios. Moreover, we identify two key indicators that signal when over-optimised output diversity begins to harm performance: Decreasing Entropy over Top-k Tokens and Diminishing Top-k Probability Mass. Our experimental results validate that these indicators are reliable signs of declining performance under different regularisations, helping prevent over-optimisation and improve alignment with human preferences.

Related papers

• Towards Improved Preference Optimization Pipeline: from Data Generation to Budget-Controlled Regularization [14.50339880957898]
  We aim to improve the preference optimization pipeline by taking a closer look at preference data generation and training regularization techniques. For preference data generation, we propose an iterative pairwise ranking mechanism that derives preference ranking of completions using pairwise comparison signals. For training regularization, we observe that preference optimization tends to achieve better convergence when the LLM predicted likelihood of preferred samples gets slightly reduced.
  arXiv  Detail & Related papers  (2024-11-07T23:03:11Z)
• Uncertainty-Penalized Direct Preference Optimization [52.387088396044206]
  We develop a pessimistic framework for DPO by introducing preference uncertainty penalization schemes. The penalization serves as a correction to the loss which attenuates the loss gradient for uncertain samples. We show improved overall performance compared to vanilla DPO, as well as better completions on prompts from high-uncertainty chosen/rejected responses.
  arXiv  Detail & Related papers  (2024-10-26T14:24:37Z)
• Ordinal Preference Optimization: Aligning Human Preferences via NDCG [28.745322441961438]
  We develop an end-to-end preference optimization algorithm by approxing NDCG with a differentiable surrogate loss. OPO outperforms existing pairwise and listwise approaches on evaluation sets and general benchmarks like AlpacaEval.
  arXiv  Detail & Related papers  (2024-10-06T03:49:28Z)
• Preference Optimization as Probabilistic Inference [21.95277469346728]
  We propose a method that can leverage unpaired preferred or dis-preferred examples, and works even when only one type of feedback is available. This flexibility allows us to apply it in scenarios with varying forms of feedback and models, including training generative language models.
  arXiv  Detail & Related papers  (2024-10-05T14:04:03Z)
• Self-supervised Preference Optimization: Enhance Your Language Model with Preference Degree Awareness [27.43137305486112]
  We propose a novel Self-supervised Preference Optimization (SPO) framework, which constructs a self-supervised preference degree loss combined with the alignment loss. The results demonstrate that SPO can be seamlessly integrated with existing preference optimization methods to achieve state-of-the-art performance.
  arXiv  Detail & Related papers  (2024-09-26T12:37:26Z)
• Bridging and Modeling Correlations in Pairwise Data for Direct Preference Optimization [75.1240295759264]
  We propose an effective framework for Bridging and Modeling Correlations in pairwise data, named BMC. We increase the consistency and informativeness of the pairwise preference signals through targeted modifications. We identify that DPO alone is insufficient to model these correlations and capture nuanced variations.
  arXiv  Detail & Related papers  (2024-08-14T11:29:47Z)
• Discovering Preference Optimization Algorithms with and for Large Language Models [50.843710797024805]
  offline preference optimization is a key method for enhancing and controlling the quality of Large Language Model (LLM) outputs. We perform objective discovery to automatically discover new state-of-the-art preference optimization algorithms without (expert) human intervention. Experiments demonstrate the state-of-the-art performance of DiscoPOP, a novel algorithm that adaptively blends logistic and exponential losses.
  arXiv  Detail & Related papers  (2024-06-12T16:58:41Z)
• Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer [52.09480867526656]
  We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences. To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model. Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
  arXiv  Detail & Related papers  (2024-05-26T05:38:50Z)
• Towards Efficient Exact Optimization of Language Model Alignment [93.39181634597877]
  Direct preference optimization (DPO) was proposed to directly optimize the policy from preference data. We show that DPO derived based on the optimal solution of problem leads to a compromised mean-seeking approximation of the optimal solution in practice. We propose efficient exact optimization (EXO) of the alignment objective.
  arXiv  Detail & Related papers  (2024-02-01T18:51:54Z)
• Functional Graphical Models: Structure Enables Offline Data-Driven Optimization [111.28605744661638]
  We show how structure can enable sample-efficient data-driven optimization. We also present a data-driven optimization algorithm that infers the FGM structure itself.
  arXiv  Detail & Related papers  (2024-01-08T22:33:14Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.