An official website of the United States government
Direct Preference Optimization (DPO) and its variants are increasingly used for aligning language models with human preferences. Although these methods are designed to teach a model to generate preferred responses more frequently relative to dispreferred responses, prior work has observed that the likelihood of preferred responses often decreases during training. The current work sheds light on the causes and implications of this counterintuitive phenomenon, which we term likelihood displacement. We demonstrate that likelihood displacement can be catastrophic, shifting probability mass from preferred responses to responses with an opposite meaning. As a simple example, training a model to prefer over can sharply increase the probability of . Moreover, when aligning the model to refuse unsafe prompts, we show that such displacement can unintentionally lead to unalignment, by shifting probability mass from preferred refusal responses to harmful responses (e.g., reducing the refusal rate of Llama-3-8B-Instruct from 74.4% to 33.4%). We theoretically characterize that likelihood displacement is driven by preferences that induce similar embeddings, as measured by a centered hidden embedding similarity (CHES) score. Empirically, the CHES score enables identifying which training samples contribute most to likelihood displacement in a given dataset. Filtering out these samples effectively mitigated unintentional unalignment in our experiments. More broadly, our results highlight the importance of curating data with sufficiently distinct preferences, for which we believe the CHES score may prove valuable
more »
« less
D2PO: Discriminator-Guided DPO with Response Evaluation Models
Varied approaches for aligning language models have been proposed, including supervised fine-tuning, RLHF, and direct optimization methods such as DPO. Although DPO has rapidly gained popularity due to its straightforward training process and competitive results, there is an open question of whether there remain practical advantages of using a discriminator, like a reward model, to evaluate responses. We propose D2PO, discriminator-guided DPO, an approach for the online setting where preferences are being collected throughout learning. As we collect gold preferences, we use these not only to train our policy, but to train a discriminative response evaluation model to silver-label even more synthetic data for policy training. We explore this approach across a set of diverse tasks, including a realistic chat setting, we find that our approach leads to higher-quality outputs compared to DPO with the same data budget, and greater efficiency in terms of preference data requirements. Furthermore, we show conditions under which silver labeling is most helpful: it is most effective when training the policy with DPO, outperforming traditional PPO, and benefits from maintaining a separate discriminator from the policy model.
more »
« less
- Award ID(s):
- 2505865
- PAR ID:
- 10631149
- Publisher / Repository:
- https://doi.org/10.48550/arXiv.2405.01511
- Date Published:
- ISSN:
- 2405.01511
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
We study the problem of aligning large language models (LLMs) with human preference data. Contrastive preference optimization has shown promising results in aligning LLMs with available preference data by optimizing the implicit reward associated with the policy. However, the contrastive objective focuses mainly on the relative values of implicit rewards associated with two responses while ignoring their actual values, resulting in suboptimal alignment with human preferences. To address this limitation, we propose calibrated direct preference optimization (Cal-DPO), a simple yet effective algorithm. We show that substantial improvement in alignment with the given preferences can be achieved simply by calibrating the implicit reward to ensure that the learned implicit rewards are comparable in scale to the ground-truth rewards. We demonstrate the theoretical advantages of Cal-DPO over existing approaches. The results of our experiments on a variety of standard benchmarks show that Cal-DPO remarkably improves off-the-shelf methods.more » « less
-
Aligning large language models (LLMs) to human preferences is a crucial step in building helpful and safe AI tools, which usually involve training on supervised datasets. Popular algorithms such as Direct Preference Optimization (DPO) rely on pairs of AI-generated responses ranked according to human annotation. The response pair annotation process might bring human bias. Building a correct preference dataset is the costly part of the alignment pipeline. To improve annotation efficiency and quality in the LLMs alignment, we propose REAL:Response Embedding-based Alignment for LLMs, a strategy for constructing a high-quality training dataset that focuses on acquiring the less ambiguous preference pairs for labeling out of a set of response candidates. Our selection process is based on the similarity of embedding responses independently of prompts, which guarantees the selection process in an off-policy setting, avoiding adaptively measuring the similarity during the training. Experimental results on real-world dataset SHP2 and synthetic HH-RLHF benchmarks indicate that choosing dissimilar response pairs enhances the direct alignment of LLMs while reducing inherited labeling errors. The model aligned with dissimilar response pairs obtained a better margin and win rate on the dialogue task. Our findings suggest that focusing on distinct pairs can reduce the label error and improve LLM alignment efficiency, saving up to 65% of annotators’ work. The code of the work can be found https://github.com/ honggen-zhang/REAL-Alignment.more » « less
-
null (Ed.)This paper presents a policy-driven sequential image augmentation approach for image-related tasks. Our approach applies a sequence of image transformations (e.g., translation, rotation) over a training image, one transformation at a time, with the augmented image from the previous time step treated as the input for the next transformation. This sequential data augmentation substantially improves sample diversity, leading to improved test performance, especially for data-hungry models (e.g., deep neural networks). However, the search for the optimal transformation of each image at each time step of the sequence has high complexity due to its combination nature. To address this challenge, we formulate the search task as a sequential decision process and introduce a deep policy network that learns to produce transformations based on image content. We also develop an iterative algorithm to jointly train a classifier and the policy network in the reinforcement learning setting. The immediate reward of a potential transformation is defined to encourage transformations producing hard samples for the current classifier. At each iteration, we employ the policy network to augment the training dataset, train a classifier with the augmented data, and train the policy net with the aid of the classifier. We apply the above approach to both public image classification benchmarks and a newly collected image dataset for material recognition. Comparisons to alternative augmentation approaches show that our policy-driven approach achieves comparable or improved classification performance while using significantly fewer augmented images. The code is available at https://github.com/Paul-LiPu/rl_autoaug.more » « less
-
Preferences within a group of people are not uniform but follow a distribution. While existing alignment methods like Direct Preference Optimization (DPO) attempt to steer models to reflect human preferences, they struggle to capture the distributional pluralistic preferences within a group. These methods often skew toward dominant preferences, overlooking the diversity of opinions, especially when conflicting preferences arise. To address this issue, we propose Group Distributional Preference Optimization (GDPO), a novel framework that aligns language models with the distribution of preferences within a group by incorporating the concept of beliefs that shape individual preferences. GDPO calibrates a language model using statistical estimation of the group's belief distribution and aligns the model with belief-conditioned preferences, offering a more inclusive alignment framework than traditional methods. In experiments using both synthetic controllable opinion generation and real-world movie review datasets, we show that DPO fails to align with the targeted belief distributions, while GDPO consistently reduces this alignment gap during training. Additionally, our evaluation metrics demonstrate that GDPO outperforms existing approaches in aligning with group distributional preferences, marking a significant advance in pluralistic alignment.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
