Unpacking DPO and PPO: Disentangling Best Practices for Learning from Preference Feedback

Learning from preference feedback has emerged as an essential step for improving the generation quality and performance of modern language models (LMs). Despite its widespread use, the way preference-based learning is applied varies wildly, with differing data, learning algorithms, and evaluations used, making disentangling the impact of each aspect difficult. In this work, we identify four core aspects of preference-based learning: preference data, learning algorithm, reward model, and policy training prompts, systematically investigate the impact of these components on downstream model performance, and suggest a recipe for strong learning for preference feedback. Our findings indicate that all aspects are important for performance, with better preference data leading to the largest improvements, followed by the choice of learning algorithm, the use of improved reward models, and finally the use of additional unlabeled prompts for policy training. Notably, PPO outperforms DPO by up to 2.5% in math and 1.2% in general domains. High-quality preference data leads to improvements of up to 8% in instruction following and truthfulness. Despite significant gains of up to 5% in mathematical evaluation when scaling up reward models, we surprisingly observe marginal improvements in other categories. We publicly release the code used for training (https://github.com/hamishivi/EasyLM) and evaluating (https://github.com/allenai/open-instruct) our models, along with the models and datasets themselves (https://huggingface.co/collections/allenai/tulu-v25-suite-66676520fd578080e126f618).

Paper

References (55)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer P8L34/10 · confidence 4/52024-06-13

Summary

The paper identifies four core components in training from preference feedback (RLHF): labeled preference data, learning algorithm, reward model, and unlabeled training prompts; and conducts a study on the effect of each component separately to disentangle their contributions to performance. The authors experiment with variations of each component and make several claims: 1. There is a clear ordering of the impact of each component on performance: preference data quality > algorithm choice > reward model quality > targeted policy prompts. 2. Synthetic data with (aggregated) per-aspect preferences works best. 3. PPO outperforms DPO 4. Increasing the reward model size or its training dataset size improves its performance on reward benchmarks, but has marginal effects on downstream tasks overall. 5. targeted online training prompts can improve performance in domain-specific settings, but not overall performance. Hence, they suggest a recipe with the best of each component.

Strengths

The paper is well presented: the claims are clear, the experiments backing them are easy to identify, and the narrative is easy to follow. The description of the experimental setup is comprehensive and very detailed, helping the applicability of the paper. The authors experiment with a comprehensive collection of preference datasets and benchmarks. The authors make several interesting observations from their experiments such as the effects f the preference datasets used showing mostly on dimensions such as truthfulness and instruction following, or the improvement of reward modeling showing in rewards benchmarks but not translating to downstream tasks.

Weaknesses

I find that the claims made in the paper often deviate from the experimental support provided, they are either overstated compared to the scope or the experiments, or inaccurately reflect the scope of the experiments. Specifically, **Algorithm choice:** Although comparing components such as prompt datasets is possible when keeping all else the same (algo, hyperparameters, etc), the claim "PPO outperforms DPO" cannot be made without carefully considering hyperparameters. The authors have a single hyperparameter configuration for each model, and although they may have been validated in other settings, they have not been optimized (within a fixed but same budget) for each algorithm on the author's setting. The DPO hyperparameters were originally used by Zephyr-Beta, and taken without tuning for TULU 2 [Appendix F.1, 22]. For PPO, apart from the number of epochs that have been cut to 1, it's not clear how the authors obtained the rest of the hyperparameters. Furthermore, from Table 2, "PPO outperforms DPO" _on average_ on _pre-selected_ datasets. When looking at individual benchmarks DPO models outperform PPO models in quite a number of tasks, and Table 2 only includes a subset of datasets which has not been motivated in the paper. Therefore I believe that the claim is misleading. **Preference dataset**: Only Table 1 with DPO is used to formulate the conclusions on the best preference dataset. For the claim to hold in the scope of the paper (considering its limitations with a single pre-trained model) the preference datasets have to be compared with at least multiple algorithms (and ideally hyperparameters). Table 2 provides some results for PPO but on all datasets, so it fails to complement Table 1. The authors do not provide insights about why synthetic data performs best. It is appreciated to observe such difference in performance, but the claim would have more value and impact with some analysis or intuition as to why it would hold. The claim "Preference-based learning has the strongest effect on instruction following and truthfulness performance." may only be valid in when the preference data itself mainly includes preferences based on instruction following and truthfulness, which is the case in the datasets considered in the paper. However, it may be very valid, but inefficient, to train factuality using preferences. The scope of such a claim should be carefully explicited. **Reward model**: Although the paper aims to disentangle the effects of each component in RLHF, the experiments it conducts on some components tend to depend on the results from the previous components. E.g. the reward model experiments are only carried out on the preference datasets providing the best performance, giving a biased view on the effects of the reward model and does not allow to formulate an ordering on the performance contribution of each component independently. Also, the average improvement of 0.6 from the 13B to the 70B UltraF. reward model has been qualified as marginal while the improvement of 0.7 from DPO to PPO has been qualified as significant.

Questions

How did the authors select the preference datasets to compare DPO vs PPO in Table 2? One would expect to have them compared on all datasets as in Table 1 or to have a justification for the datasets that have been dropped/kept.

Rating

4

Confidence

4

Soundness

2

Presentation

4

Contribution

2

Limitations

Although in the appendix, the authors adequately state the limitations of their work and its broader impacts.

Reviewer wF895/10 · confidence 4/52024-07-05

Summary

The work summarizes the area of learning from preferences for optimizing language models. Specifically, the analyze four aspects: preference data, learning algorithm, reward model, and policy training prompts. They empirically answer questions on the downstream improvement by improvements in each of these axis. Overall, they observe that the largest improvement arises from higher quality preference data, followed by choice of learning algorithm (PPO over DPO) with relatively smaller gains on reward model capabilities followed by training prompts.

Strengths

* Clarity of writing: The work is well written, concise and clear to follow. * Significance: We gain practical insights from this work, including the use of synthetic datasets and large reward models, which can be valuable for practitioners in the field. * Quality: Empirical results are well documented, methods are sound * Originality: While this offers new insights, the work extends existing methodologies rather than introducing novel concepts.

Weaknesses

* Novelty: While the work helps answers some excellent questions, the overall scope of the work seems limited. It would be interesting if this could be expanded to a more comprehensive study, see the question below.

Questions

How do these insights scale to other paradigms of language models such as in the multimodal regime?

Rating

5

Confidence

4

Soundness

3

Presentation

4

Contribution

2

Limitations

Yes, Section A in the Appendix.

Reviewer WxZd5/10 · confidence 4/52024-07-07

Summary

This work concentrates on methods for LLM learning from preference feedback and conducts a lot of experiments to identify key aspects of the preference-based methods. The work gives an ordering for the importance of the core aspects: preference data quality, algorithm choice, reward model quality, and finally targeted policy training prompt.

Strengths

Since the pipeline of preference-based methods has many details, it is hard to see works like this one to have done plenty of ablation experiments. This work brings a lot of observations for this community and can help researchers to have more understanding on different parts of preference-based methods. This work gives an order for the importance of four key elements of preference learning. This is a relatively systematical investigation for reference-based learning. This work also gives a recipe for learning from preferences and this may guide better performance for LLM.

Weaknesses

Although a lot of observations are provided in this work, it seems that authors don't provide a deeper and more systematic understanding of the whole process of preference learning. Indeed, this is not an easy task, but I think it would be really helpful if we could draw some deeper insights from all these observations. Since there are many elements involved in the preference training process, it is not easy to control variates. I found that some results might be improved with more ablations.

Questions

Here I have some detailed questions and I would be thankful if authors can give some further explanations. 1. On page 5 line 153, the quality of pairwise judgements is mentioned but there is not much explanation for this. Can we give a relatively quantitative method to judge the quality of pairs? 2. On page 5 line 161, it is mentioned that PPO outperforms DPO. Since there is a lot of DPO related research recently, I'm willing to see some more discussion about PPO and DPO. Is the offline property of DPO the main reason? 3. On page 6 table 3, it seems quite strange that 70B Mix RM has a lower RewardBench Score than 13B Mix RM. Is there any reason for this result? 4. On page 7 line 201, it is mentioned that there are no improvements in overall downstream performance. I noticed that the PPO is trained on UltraF prompts. Since the two Mix RMs are trained on more data, I might expect the two models can achieve better performance if the PPO is trained on the Mix prompts. In Fig.3, we observed that Mix RMs gain higher acc on GSM train. Also, I noticed that in table 4, Mix RMs trained on the code-and-math-mixed prompts have poor performance. Can we get some principles for dataset selection for RM? 5. On page 7 table 4, I found that the set of UF prompts and the set of Mixed prompts have the size. Since more code and math prompts are used in the Mixed prompts, some UF prompts are not used. Could this lead to the degradation of the Avg. Across all Evals? What if we use all UF prompts and some further code and math prompts to train a PPO? 6. On page 18 table 7, the UF dataset is split into Strong, Middle and Weak sets. Are the four datasets (UF all, Strong, Middle and Weak) downsampled to the same size for comparision? 7. On page 19 equation (5), is there any typo with the \pi_{ref}?Does it refer to the SFT policy? 8. The work shows that preference data is most important. I noticed that effect of dataset size is not well discussed. Is there any observation about this? Can more data provide more improvements? 9. I am curious about the training process during these experiments. For example, how the training reward changes, how the KL divergence between PPO policy and SFT policy changes, how the entropy of the policy changes, and how the length of the responses changes during the whole training processes. I would be grateful if authors could share some interesting observations during training.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A

Reviewer mnCL6/10 · confidence 3/52024-07-10

Summary

This paper disentangles core components of current learning from preference feedback algorithms in alignment, conducts comprehensive experiments on the individual effect of each component, and provides a recipe of learning from preference feedback based on experiment results.

Strengths

1. This paper aims to understand PPO and DPO from a practical perspective by conducting comprehensive experiments on core components of the RLHF pipeline. The results enhance the understanding of RLHF and provide valuable reference data for the community. 2. This paper is well-written and clearly presented.

Weaknesses

1. The experiment testing the influence of training dataset size for the reward model is not sound enough. The dataset is a mixture of different high-quality datasets, which may affect the reward model's performance on different tasks, making it difficult to attribute changes in performance solely to the dataset size. Besides, the conclusion that "both increasing the reward model dataset (‘Mix’) and reward model size (from 13B to 70B) improve performance" is not evident, as the dataset size shows little influence in the 70B model. 2. The limited number of models tested in the paper may restrict the impact of this work, as results could vary among different models, as the authors have mentioned in their "Limitations" section.

Questions

Here are some questions that I'd like to discuss with the authors to enhance my understanding of this work: 1. What is the data ratio of Mix and UltraF in the reward model experiment? 2. What is the codebase used in this paper? 3. On the right side of Table 3, the GSM scores indicate that the 70B UltraF.RM with UltraF.prompts performs best, whereas Figure 3 shows a different result. Is this discrepancy due to different experimental settings?

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

none

Reviewer wF892024-08-11

Acknowledging Rebuttal

Thanks for taking the time answering the questions, keeping my score.

Reviewer P8L32024-08-11

I thank the authors for the clarifications. Their reply addresses all my points but also acknowledges a major limitation due to computational costs. I understand that the authors aimed to provide the best empirical recommendations they could given a limited budget and acknowledge that many of the numbers in the paper will be relevant to readers. However, I believe that under this limited computational budget, not enough research questions have been tackled and answered in satisfactory manner. I find this a necessary condition for a publication at the venue. I, therefore, increase my score but maintain it below the acceptance threshold.

Authorsrebuttal2024-08-13

Thank you very much for your response and raising your score! As for research questions, we believe that we have examined a number of useful and interesting areas, including (a) what public datasets works well for DPO, (b) relative performance of DPO and PPO across different datasets (and models, with our Llama 3 results), (c) the effect (or rather, surprising lack thereof) of the size of the RM during PPO training (with similar results from increasing dataset size using well-performing datasets from prior steps), (d) the effect of using more domain-specific prompts during PPO training. We hope and believe these are interesting and useful results for researchers in the RLHF space, and supported by the experimental results we have reported. While it would be great to further explore these effects by doing a more thorough grid search over {dataset, RM size, prompt} combinations, we note that running a PPO experiment takes \~54 hours on a v3-256, and so running all possible dataset and RM size combinations (14 x 2) would take 1,512 hours (ignoring extra experiments exploring mixing datasets or varying prompt sets, which would incur further costs). Based on the google cloud calculator (https://cloud.google.com/products/calculator?hl=en) as of the 12th of august, 700 hours on a v3-256 (rather than newer, more expensive options) costs \\$394,240 in europe-west4, and so running these additional experiments would take \~\$800,000, and take 63 days to run (and if we wanted to e.g. explore the effect of using a different prompt set in each case, this would further multiply the cost). As such, we first ablated datasets in cheaper experiments (DPO), and then examined the more promising ones in PPO, and then additionally further explored key aspects of PPO using the best dataset found. This allows us to explore interesting and promising aspects of PPO without extreme computational costs.

Reviewer P8L32024-08-13

I totally understand the computational budget limits and did not imply that the budget was objectively low, nor did I imply that its use was not efficient. What I mean is that there is a misfit between the budget and the initial research question addressed or the claims expected from them, especially the "Disentangling" part, which didn't really end up being a disentangling. (a), (b), (c), and (d) can all be answered by observations and that's why I said that the numbers will be useful to readers, but I believe that they lack some research depth providing an understanding of why a number is better than another. The authors blamed the computational budget which I then raised as a limitation and a misfit, but I also believe that the current results can be analyzed further to answer other research questions with enough depth. For this reason, I prefer to keep my score.

Authorsrebuttal2024-08-13

Thank you for responding! We are happy to clarify any further questions or address concerns if you have other reasons for not raising the score beyond the original weaknesses and questions asked (which we hope we have addressed appropriately above).

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC