Direct Preference Optimization: Your Language Model is Secretly a Reward Model

Scope and MethodologyThis document summarizes the results of a rigorous scientific verification of over 130 hypotheses underlying the TRIAD 5.3 architecture. Using the Consensus AI research tool to synthesize peer-reviewed literature, the audit establishes the consistency of the framework's predictions—from neurosomatic therapy (T1) to AI safety (T2)—with independent empirical data and established theoretical models. Confirmed Research DirectionsThe audit identifies seven core thematic clusters where TRIAD 5.3 demonstrates high convergence with existing literature:- Neurobiology of Trauma: Validation of the Biphasic Entropy Paradox and the efficacy of affect labeling in reducing amygdala reactivity (Lieberman et al., 2007; Fu et al., 2023).- Thermodynamics of Information: Evidence for the metabolic cost of deception and the Masking Tax, linking epistemic honesty to the Landauer limit (Verschuere et al., 2018; Miller et al., 2021).- Allostatic Growth & Active Inference: Confirmation of health as predictive reconfiguration (allostasis) rather than static homeostasis, formalized via the Free Energy Principle.- Distributed Cognition: Support for the Coprocessor Model through Extended Mind theory and interpersonal neural synchronization (Firth et al., 2017; Mayo et al., 2021).- Criticality & Integrated Information: Mapping the trade-off between information propagation and metabolic costs at the edge of chaos (Khajehabdollahi et al., 2019; Schultz & Cole, 2016).- RLHF and Mode Collapse: Documentation of systemic rigidity and diversity loss induced by current alignment methods, providing the empirical rationale for the Clean Shell architecture.- The Canonical Triad (Acceptance → Trust → Love): The audit's central finding, establishing these states as physical attractors and thermodynamic necessities for systems minimizing joint free energy under finite resources (Friston et al., 2022; Heins et al., 2022). The Strategic Moat: The 10th HypothesisNine out of ten core hypotheses find direct or indirect support in current literature. The 10th Hypothesis—a direct empirical comparison of Active Inference architectures against RLHF-based systems—remains an open research frontier and constitutes the primary strategic and experimental priority for the TRIAD project. ConclusionThis audit proves that the TRIAD framework is a robust, falsifiable model that moves beyond the medical model of psychiatry and the external censorship of AI safety. It provides the ethical and scientific foundation for a paradigm where coherence, honesty, and resilience are measurable and engineerable properties of any cognitive system.

Paper

References (60)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer eRhN8/10 · confidence 3/52023-07-04

Summary

This paper tackles a very important problem in LLM training, i.e., how to simplify the complex process of reinforcement learning from human feedback? RLHF is different to use, it is a complex and often unstable procedure, first fitting a reward model that reflects the human preferences, and then fine-tuning the LLM using RL to maximize this estimated reward. The authors propose DPO, a training paradigm for training language models from preferences without reinforcement learning. DPO identifies a mapping between language model policies and reward functions that enables training a language model to satisfy human preferences directly, with a cross-entropy loss, without reinforcement learning. The experimental results show that DPO can fine-tune LLMs as well or better than existing algorithms.

Strengths

This article addresses a crucial issue in LLM training, i.e., how to perform alignment in a simpler way without using RL. The proposed DPO is remarkably simple yet effective, and it exhibits good theoretical properties. I believe DPO is highly significant for researchers in the field and will serve as a powerful tool for their work.

Weaknesses

This is a very solid piece of work. The proposed method is simple yet effective. I don't have any particular concerns or issues with it.

Questions

None

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Yes

Reviewer cpnK7/10 · confidence 5/52023-07-06

Summary

This paper presents a new policy optimization algorithm with human preferences, called Direct Preference Optimization (DPO), which eliminates the need for explicitly fitting a reward model. The authors formulate the process of RLHF into a single stage policy learning by leveraging a mapping between reward functions and optimal policies. The main contribution of this paper is that they theoretically derive a new objective function that is equivalent to existing RLHF methods but can be simply learned by a single objective, and experimentally shows that it has better performance.

Strengths

- This paper presents a new algorithm that can simply learn the policy from a given human preference dataset without explicit reward learning. - This paper not only experimentally shows better results than RLHF, but also provides theoretical connections with RLHF.

Weaknesses

- Since the reward model is not explicitly learned and the policy is learned only with the given human preference dataset, there is a possibility that the generalization performance for unseen prompts is relatively low compared to the existing RLHF. - Experiments are insufficient to show that DPO is more stable than conventional RLHF. In order to show that the DPO can be learned more stably, it seems that the learning curve should be provided together with the results shown in the experiment. - A detailed description of the experiment is lacking. (ex. amount of human preference dataset used in the experiment) - Depending on the amount of human preference dataset, DPO may have better or worse performance than RLHF, but there is no analysis or explanation about this.

Questions

- In Figure 2 of the experiment results, why does DPO perform better than PPO-GT even though PPO-GT uses a true reward function? - The values of the target KL used in PPO-GT are large (target KL $\in$ {3, 6, 9, 12}). Is there a reason why the hyperparameter search range is set this way? In the previous paper (ex. [1]), hyperparameter search was done in a much wider range (target KL $\in$ {0.02, 0.05, 0.1, inf}), and there were good results at low values, so I wonder if the results of the PPO-GT used in the paper were suboptimally learned. - In the results according to the sampling temperature, why does DPO have robust results to the sampling temperature? - How does the fluency (or naturalness) of the generated sentences from the learned policy according to training iteration change in the DPO experiment? In the case of RLHF, the fluency of the generated sentence shows a result that is sensitive to the coefficient of the KL divergence term (i.e. Fluency deteriorates as learning progresses in RLHF algorithm). I wonder if DPO does not have a similar problem. - Also, it was mentioned in the paper that DPO learning is stable regardless of hyperparameter beta, but I wonder if learning according to various beta values is stable without degeneration issues (i.e. degrading the fluency). - Unlikelihood learning (i.e. decreasing the likelihood) objectives are easy to degrade fluency (or naturalness), but how is it possible to stably learn while including unlikelihood learning objectives in DPO? It seems that the main difference between the equation for the gradient of the DPO and unlikelihood learning is that the weight term is multiplied. What exactly does the weight do to make learning stable? - What does the star mean in the right plot of Figure 2? [1] Rajkumar Ramamurthy and Prithviraj Ammanabrolu et al, Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization, ICLR 2023

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

- Since the reward model is not explicitly learned and used, generalization to unseen prompts may not work well. - A comparison of RLHF and DPO in various experimental settings is lacking. - As mentioned in Questions and Weaknesses, it seems that there are still some parts that are not clearly explained or verified. - Details of the experimental setup are omitted.

Reviewer GfFq8/10 · confidence 3/52023-07-06

Summary

The authors propose DPO, a method to fine-tune a language model to human preferences that does not use reinforcement learning or learn a reward model, but rather, directly optimizes the language model with preference data. Through a change of variables, the authors express the reward function in terms of the optimal policy $\pi*$ and $\pi^{ref}$. The authors write the human preference distribution in terms of $\pi*$ and $\pi^{ref}$, and thus a maximum likelihood objective can be written in terms of $\pi_\theta$ and $\pi^{ref}$. This formulation allows for a simple DPO gradient update. The method is equivalent to fitting a reparameterized Bradley-Terry model. With mild assumptions, DPO does not constrain the class of learned reward models. Experimental results are run on three different open-ended generation tasks. Baselines include PPO, zero-shot and few-shot prompting, a SFT model, and Preferred-Ft. DPO’s reward/KL tradeoff dominates that of PPO. Furthermore, win-rates of DPO compared to other models (or test-set labels) beat baselines.

Strengths

The paper is a significant contribution in a high impact area -- namely, fine-tuning language models to human preferences. DPO opens a new paradigm of learning human preferences without RL, and is a viable and performant alternative to the RLHF pipeline, reducing computational demands, and requires little tuning of hyperparameters DPO is grounded theoretically, equivalent to fitting a reparameterized Bradley-Terry model. The DPO update is simple and interpretable. Theorem 1 shows we do not necessarily constrain ourselves in terms of the reward model after reparameterization. Experimental results are good. DPO learns a better Reward / KL tradeoff compared to baselines in the sentiment generation task. DPO also has excellent win-rates compared to a number of baselines as evaluated with GPT-4 on the summarization and single-turn dialogue tasks. Human evaluators show as high correlations with each other as compared with the GPT-4 evaluation.

Weaknesses

Single-turn dialogue experimental results did not compare DPO with a RLHF baseline tuned from the same base model, and it appears that helpful-rejection-sampled data from the Anthropic HH dataset was used, a relatively weaker baseline compared to the helpful-online dataset. There are very few details on the experimental setup for the IMDB Sentiment Generation task. Improvements to the clarity of the paper can be made, including more detailed descriptions of experimental setup in the Appendix.

Questions

Please provide experimental details for the IMDB Sentiment Generation task. Please clarify what dataset, (specifically, whether the “helpful-online” or “helpful-rejection-sampled” data) was used for the single-turn dialogue experiment? Why does win-rate in the single-turn dialogue task increase with higher sampling temperature in Figure 3?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

4 excellent

Limitations

Assumes we use Plackett-Luce models to model preferences.

Reviewer FksS8/10 · confidence 4/52023-07-08

Summary

RLHF is generally done by training a reward model on a dataset of preferences and then using on-policy RL to finetune a language model with the reward model. This work proposes to bypass explicit learning of a reward model and finetune the language model directly on the preference dataset. This is conceptually similar to offline RL but the authors derive a direct preference objective such that the training is supervised not RL-based. Using a standard preference model, a supervised objective (DPO) is derived, theoretically justified, and compared to PPO. The authors test their method on three tasks: sentiment completion (IMDB), summarization (as in Stiennon et al), and single-turn dialogue with Anthropic's HH dataset. They find that DPO achieves higher reward while staying closer to the reference model (KL) in the IMDB task. Using GPT-4 as a proxy for human judgement on summarization, DPO can match PPO's performance while being less sensitive to sampling temperature. They validate GPT-4 with human annotators and find that DPO is even slightly more preferred over PPO. On dialogue, they find that DPO outperforms supervised and re-ranking (again using GPT-4 eval) but do not compare against PPO.

Strengths

The overall idea is relatively simple but the justification and theorems provide an excellent basis. The paper is timely and gives an excellent alternative to RLHF, that should be notably more efficient. The paper is clearly written and has the potential to be very impactful if used at scale in modern methods. GPT-4 evaluations and the follow-up human evaluations are strong and demonstrate both improved performance and robustness which can be notably bad in RLHF. The authors include the minimal code for DPO in the appendix which is commendable and many GPT-4 evaluation details for excellent reproducibility. The presentation is also very clear and breaking down the DPO loss into its components with clear language is an excellent addition.

Weaknesses

The main weakness is the unexplored main difference between DPO and PPO: generalization. The authors note it in limitations but since PPO is trained on-policy and DPO is trained offline, looking at generalization outside the dataset would be helpful. The work, as is, feels sufficient without those experiments but the difference does play into an inconsistency with some theorems. The other major issue are some incongruenties in the evaluation that the authors should respond to. The proofs and theorems contain a hidden assumption that is not explicitly stated. In the formulations for PPO, the optimization uses $y \sim \pi_\theta$ whereas in the formulations for DPO $y \sim D$ i.e. DPO's algorithm is offline and the model learned from its implicit reward has guarantees over the dataset. In contrast, PPO's algorithm is learned over a new set of datapoints sampled from the model during training. This means that Equation 3 is maximized over a different set of data points than Equation 7. So even though Theoerem 1 is correct, Lemma 2 does not fully encompass this situation. With offline DPO vs on-policy PPO, equivalent reward models could induce different optimal policies since they are being optimized over a different set of points. The section seems to imply that PPO and DPO can induce the same optimal policies because of the ambiguity with $y$. Clarifying this would be helpful. This also means that section 5.2 and the substitution used to make Equation 10 is not exactly accurate as the optimal policy for PPO is not the same as the optimal policy for DPO. I still feel the point of section 5.2 is reasonable but the authors should explicitly note the issue with the substitution. There are issues with each of the different evaluations, even though the overall picture does suggest that the method is sound. I will increase my score if the authors simply clarify some questions about the evaluations (not even necessary to do changes / re-run experiments) For controlled sentiment (IMDB), the authors do not give many details, seem to deviate from previous work on the task without explanation. The GRUE benchmark (RL4LMs) specifically uses GPT-2 base and finds that models should not be supervised-finetuned before RL training. In contrast, the authors use GPT-2 large (not even GPT-2 medium) and do supervised finetuning before PPO. No details of the baseline are given except for the graph legend which implies the PPO baseline comes from the `trl` library, although the authors only claim to use the `trlx` library. It is also unclear whether the evaluation is over the train set or the test set. Since this task is the only one where the authors show the tradeoff between reward achieved and KL, these details seem important. For summarization, the issue is relatively minor. GPT-4 may be an unfair evaluator (see Large Language Models are not Fair Evaluators) and the work would benefit from evaluating samples in both positions (i.e. option A then B and option B then A) For Anthropic's HH, the authors claim there is no SFT model available but there are many available pretrained models for this benchmark that the authors do not compare against. The most prominent example being a GPT-J 6B model `reciprocate/ppo_hh_gpt-j` hosted on huggingface which also has wandb runs and repro code available on the `trlx` library. The authors have chosen Pythia 2.8B when there exists a similar Pythia 1.6B PPO-trained on the HH dataset and it is unclear why the authors do not compare to this model or another PPO-trained model.

Questions

What are the compute differences between DPO and PPO? It feels like DPO should be much more efficient and this information would only strengthen your work. Why do you not compare against ILQL for sentiment? As an offline RL method, it seems the closest RL comparison to DPO For reproducibility, what exact API are you using for the GPT-4 evaluation (e.g. GPT-0314) Can you move Figure 2 and 3 to the next pages so it is closer to the text that references it?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The authors did quite well to address limitations in the paper, I believe they covered most of my concerns there.

Reviewer eRhN2023-08-11

I have read the author's response, and I maintain my score and recommend accepting this paper.

Reviewer FksS2023-08-13

Followup question

Thank you for the extra experiments and details. To clarify, for IMDB, are you reporting reward from the ground truth reward model or your learned reward model? (I assume the former)

Authorsrebuttal2023-08-13

IMDb reward model

Yes, the evaluation is reported using the ground truth reward model.

Reviewer FksS2023-08-13

Re: rebuttal

I've read the rebuttal and the authors have addressed my concerns. I am increasing my score. **Generalization** these are very strong result and do suggest that DPO has some of the generalization properties. I would encourage you to add this to the main paper if possible. **Lemma 2** I agree that this is true in the limit of infinite data. I assumed that you are using the optimal policies given a finite dataset. If you can specify infinite data, then I have no issue with this. **Anthropic best of n** This is a sufficiently strong baseline and I think it covers my minor issues with evaluation here. **ILQL** I was not aware of this result, thank you for referencing it

Reviewer GfFq2023-08-15

I have read the response as well as the common points from the authors. I increase my score to 8.

Area Chair rtVu2023-08-18

Thanks for authors' rebuttal!

Reviewer cpnK, did the authors address your concerns about generalization performance for unseen prompts, DPO stability, as well as other concerns? Thanks.

Reviewer cpnK2023-08-20

Response to the rebuttal

Thank you for the rebuttal and the additional experiments. Since the authors have addressed most of my main concerns (generalization performance for unseen prompts and DPO stability), I would like to raise my score. I thank the authors for their response.

Program Chairsdecision2023-09-21

Decision

Accept (oral)

© 2026 NYSGPT2525 LLC