A Common Pitfall of Margin-based Language Model Alignment: Gradient Entanglement

• URL: http://arxiv.org/abs/2410.13828v1
• Date: Thu, 17 Oct 2024 17:52:01 GMT
• Title: A Common Pitfall of Margin-based Language Model Alignment: Gradient Entanglement
• Authors: Hui Yuan, Yifan Zeng, Yue Wu, Huazheng Wang, Mengdi Wang, Liu Leqi,
• Abstract summary: Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for language model alignment. In this paper, we identify a common pitfall of margin-based methods. We demystify the reasons behind these problematic behaviors.
• Score: 47.95776810771774
• License:
• Abstract: Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for language model (LM) alignment. At its core, RLHF uses a margin-based loss for preference optimization, specifying ideal LM behavior only by the difference between preferred and dispreferred responses. In this paper, we identify a common pitfall of margin-based methods -- the under-specification of ideal LM behavior on preferred and dispreferred responses individually, which leads to two unintended consequences as the margin increases: (1) The probability of dispreferred (e.g., unsafe) responses may increase, resulting in potential safety alignment failures. (2) The probability of preferred responses may decrease, even when those responses are ideal. We demystify the reasons behind these problematic behaviors: margin-based losses couple the change in the preferred probability to the gradient of the dispreferred one, and vice versa, often preventing the preferred probability from increasing while the dispreferred one decreases, and thus causing a synchronized increase or decrease in both probabilities. We term this effect, inherent in margin-based objectives, gradient entanglement. Formally, we derive conditions for general margin-based alignment objectives under which gradient entanglement becomes concerning: the inner product of the gradients of preferred and dispreferred log-probabilities is large relative to the individual gradient norms. We theoretically investigate why such inner products can be large when aligning language models and empirically validate our findings. Empirical implications of our framework extend to explaining important differences in the training dynamics of various preference optimization algorithms, and suggesting potential algorithm designs to mitigate the under-specification issue of margin-based methods and thereby improving language model alignment.

Related papers

• Understanding Likelihood Over-optimisation in Direct Alignment Algorithms [20.043560907227018]
  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.
  arXiv  Detail & Related papers  (2024-10-15T15:14:22Z)
• Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization [60.176008034221404]
  Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences. Prior work has observed that the likelihood of preferred responses often decreases during training. We demonstrate that likelihood displacement can be catastrophic, shifting probability mass from preferred responses to responses with an opposite meaning.
  arXiv  Detail & Related papers  (2024-10-11T14:22:44Z)
• Geometric-Averaged Preference Optimization for Soft Preference Labels [78.2746007085333]
  Many algorithms for aligning LLMs with human preferences assume that human preferences are binary and deterministic. In this work, we introduce the distributional soft preference labels and improve Direct Preference Optimization (DPO) with a weighted geometric average of the LLM output likelihood in the loss function.
  arXiv  Detail & Related papers  (2024-09-10T17:54:28Z)
• On the Algorithmic Bias of Aligning Large Language Models with RLHF: Preference Collapse and Matching Regularization [33.331389392270665]
  preference matching (PM) RLHF is a novel approach that aligns large language models with the preference distribution of the reward model under the Bradley--Terry--Luce/Plackett--Luce model. Central to our approach is a PM regularizer that takes the form of the negative logarithm of the LLM's policy probability distribution over responses. For practical implementation, we introduce a conditional variant of PM RLHF that is tailored to natural language generation.
  arXiv  Detail & Related papers  (2024-05-26T07:00:05Z)
• 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)
• Controllable Preference Optimization: Toward Controllable Multi-Objective Alignment [103.12563033438715]
  Alignment in artificial intelligence pursues consistency between model responses and human preferences as well as values. Existing alignment techniques are mostly unidirectional, leading to suboptimal trade-offs and poor flexibility over various objectives. We introduce controllable preference optimization (CPO), which explicitly specifies preference scores for different objectives.
  arXiv  Detail & Related papers  (2024-02-29T12:12:30Z)
• Smoothing the Edges: Smooth Optimization for Sparse Regularization using Hadamard Overparametrization [10.009748368458409]
  We present a framework for smooth optimization of explicitly regularized objectives for (structured) sparsity. Our method enables fully differentiable approximation-free optimization and is thus compatible with the ubiquitous gradient descent paradigm in deep learning.
  arXiv  Detail & Related papers  (2023-07-07T13:06:12Z)
• Provable Offline Preference-Based Reinforcement Learning [95.00042541409901]
  We investigate the problem of offline Preference-based Reinforcement Learning (PbRL) with human feedback. We consider the general reward setting where the reward can be defined over the whole trajectory. We introduce a new single-policy concentrability coefficient, which can be upper bounded by the per-trajectory concentrability.
  arXiv  Detail & Related papers  (2023-05-24T07:11:26Z)

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.