Noise Contrastive Alignment of Language Models with Explicit Rewards

User intentions are typically formalized as evaluation rewards to be maximized when fine-tuning language models (LMs). Existing alignment methods, such as Direct Preference Optimization (DPO), are mainly tailored for pairwise preference data where rewards are implicitly defined rather than explicitly given. In this paper, we introduce a general framework for LM alignment, leveraging Noise Contrastive Estimation (NCE) to bridge the gap in handling reward datasets explicitly annotated with scalar evaluations. Our framework comprises two parallel algorithms, NCA and InfoNCA, both enabling the direct extraction of an LM policy from reward data as well as preference data. Notably, we show that the DPO loss is a special case of our proposed InfoNCA objective under pairwise preference settings, thereby integrating and extending current alignment theories. By comparing NCA and InfoNCA, we demonstrate that the well-observed decreasing-likelihood trend of DPO/InfoNCA is caused by their focus on adjusting relative likelihood across different responses. In contrast, NCA optimizes the absolute likelihood for each response, thereby effectively preventing the chosen likelihood from decreasing. We evaluate our methods in both reward and preference settings with Mistral-8*7B and 7B models. Experiments suggest that InfoNCA/NCA surpasses various preference baselines when reward datasets are available. We also find NCA significantly outperforms DPO in complex reasoning tasks like math and coding.

Paper

References (53)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer xAZC7/10 · confidence 4/52024-07-13

Summary

The paper's main contributions are: 1. Theoretical Integration: It integrates Direct Policy Optimization (DPO) with contrastive learning theories, offering a general framework for using reward and preference data. 2. Value of Suboptimal Responses: Highlights the importance of suboptimal responses in optimizing language models, showing improved performance over other methods by fully utilizing reward data. 3. Performance Improvement: Demonstrates how NCA counters the data likelihood decline in DPO, enhancing practical performance.

Strengths

1. Utilizing the explicit reward for alignment which is quite interesting. 2. Propose a general method to address this problem. 3. It sees improvement over DPO on leaderboards. Specially, it is much better than DPO in math and coding tasks.

Weaknesses

I don't have further suggestions for this part.

Questions

How do you compare your work with the paper titled Direct Preference Optimization with an Offset(https://arxiv.org/pdf/2402.10571).

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

I don't see any issues with this part.

Reviewer WFsu7/10 · confidence 3/52024-07-15

Summary

## Summary - Typically, RLHF algorithms like DPO use preference pairs dataset. - The key question, this paper tries to answer is how to incorporate reward datasets annotated with scalar values. Previous approach usually prune the scalar reward datasets by selecting the best response and pairing with a random remaining response. However, this work shows that the extra suboptimal responses can also be beneficial for policy training. - They present two main algorithms (inspired from contrastive learning literature) here which handle the aforementioned problem - NCA Noice Contrastive Alignment - InfoNCA - They also show that NCA is able to mitigate the "decreasing likelihood trend" observed with the DPO algorithm. NCA also outperforms DPO on math and coding tasks because it prevents the likelihood of preferred responses from decreasing. The main difference between InfoNCA and NCA objective is that the NCA objective has an additional regularizer term - The "decreasing likelihood trend" can be simply described as the decrease in likelihood of the preferred response after DPO training. - This happens because DPO focuses on adjusting relative likelihoods across generations whereas NCA focuses on adjusting absolute likelhood - InfoNCA compares multiple responses and identifies the one sampled from optimal policy. NCA predicts the model source of a single response. - The authors also show that DPO is a special case of InfoNCA

Strengths

## Strengths - The loss objectives are intuitively easy to follow from equation 6 and equation 11. - The paper is well written and easy to follow, with clear and accessible writing. The literature review provides a solid contextual foundation for the work. - The experiments do a good job at substantiating the claims presented and justifying the NCA and InfoNCA algorithm. - I also like the section on empirical takeaway: when to choose NCA over DPO. - The paper overall addresses an important problem and proposes the NCA solution which helps tackle the decreasing likelihood issue. The empirical takeaway for reasoning tasks is also helpful for an ML practitioner. My overall recommendation for the paper is an accept but I would very much appreciate if the authors can answer my questions to help deepen my understanding of the work.

Weaknesses

I could not find any weakness

Questions

## Questions - I have not gone through the material in the appendix but at several places in the theorems, the authors assume a model with unlimited capacity. Why is this important and what are the implications otherwise? - It is not obvious to me how equation 11 leads to predicting the model source of a single response in Figure 3

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have adequately addressed the limitations.

Reviewer UYN17/10 · confidence 3/52024-07-17

Summary

The authors introduce a general framework for LM alignment, leveraging Noise Contrastive Estimation (NCE) to bridge the gap in handling reward datasets explicitly annotated with scalar evaluations. The framework comprises two parallel algorithms, NCA and InfoNCA, both enabling the direct extraction of an LM policy from reward data as well as preference data.

Strengths

1. The method bridge the theoretical gap between DPO and classic contrastive learning theories. InfoNCA and NCA are uniquely suited for both reward and preference data, offering a general framework that integrates preference-based algorithms. 2. The proposed method outperforms various preference methods by fully exploiting data information in reward datasets. 3. It's great to see the method not hurting UltraInteract and leads to better performance.

Weaknesses

1. The proposed method looks trivial by integrating classic contrastive learning theories. Different methods coming out at the same time. They look similar and not easy to find which is better, such as the concurrent work SimPO: Simple Preference Optimization with a Reference-Free Reward.

Questions

More training details, such as hyper-parameter search space

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

It would be better list more training details.

Reviewer EZsK7/10 · confidence 4/52024-07-19

Summary

This paper leverages Noise Contrastive Estimation (NCE) to align language models (LMs) with explicit reward data, which can handle both scalar evaluations and multiple preference data. The proposed methods, NCA and InfoNCA, extend current alignment theories by improving upon DPO and addressing the issue of decreasing likelihood observed in DPO. Experimental results demonstrate that InfoNCA and NCA outperform existing preference-based methods.

Strengths

1. The paper bridges the gap between preference-based and reward-based alignment methods by deriving InfoNCA from InfoNCE, thereby extending DPO to handle explicit rewards with a strong theoretical foundation. 2. Empirical results show significant performance improvements in various tasks, including complex reasoning tasks, by effectively leveraging both optimal and suboptimal responses from reward datasets. 3. Mitigating Likelihood Decrease: NCA effectively prevents the chosen likelihood from decreasing during training, addressing a critical issue in DPO and enhancing model performance stability.

Weaknesses

1. The motivation of this paper is weak: The description that DPO can only be used for pairwise preference data is wrong (such as Line 31 and 43). In fact, the appendix of DPO says that it can be applied to multiple preference data based on the Plackett-Luce model. 2. The main experiments lack related baselines. Since the proposed method is a reward-based method that utilizes reward data for preference ranking, Table 2 should also include PPO under the annotation type "Reward" for direct comparison since it also utilizes reward data. In addition, I suggest comparing it with PRO (https://arxiv.org/pdf/2306.17492) under the annotation type "preference" since this work directly applies DPO to the setting of multiple preference ranking. 3. The proposed method InfoNCA uses both preference data and reward data for training, resulting in more computational overhead compared with PPO and DPO. Thus, it is not surprising that better performance is achieved than with preference-based methods, although the improvement is marginal. Is it a fair comparison if the methods use different magnitudes of data? Please also see comment 1.

Questions

1. What is the amount and percentage of reward and preference data in each dataset? 2. Since InfoNCA and NCA are reward-based methods, why not compare them with PPO in terms of MT-bench, AlpcaEval, and Wins vs. PPO?

Rating

7

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

Limitations are discussed in the appendix

Authorsrebuttal2024-08-08

Additional Response

Dear reviewer, We copy the second part of our rebuttal response here to ease reading. In our responses, we have compared with more baselines and clarified the details of the data we are using based on your suggestions. We hope we can address your concerns and are glad to answer any further questions you might have. Thank you again for your valuable feedback! # Official Response to Reviewer EZsK (2/2) **Q5: suggest comparing it with PRO (https://arxiv.org/pdf/2306.17492).** **A5:** PRO is a quite related work of ours because it can also handle response data of arbitrary length. It is also highly related to the PL method as we have discussed in **A1**. Still, PRO is essentially a **ranking/preference-based** method, while (Info)NCA are **reward-based** methods. **Theory comparison:** **We have added a new section as Appendix G.2 in our paper to compare with PRO theoretically:** (cannot update the manuscript during rebuttal, though) The main difference between PRO and the PL+DPO method is that 1. PRO has different reward formulations. It formulates $r_\theta$ as the average log-probability of a sentence $\frac{1}{\|y\|}\Sigma\log \pi_\theta(y|X)$ instead of the log-probability ratio $\log \frac{\pi_\theta(y|x)}{\mu(y|x)}$ as is required by DPO. 2. Because there is no $\mu$ in reward models. PRO needs to be additionally regularized by an SFT loss in order to stay close to the pretrained LLM $\mu$. **Additional Experiments:** | Experiment| MT-bench | |----|---| | Mixtral-7B-sft | 64.5 | DPO + PL Method| 74.1 | | PRO (SFT) | 68.9 | | InfoNCA | 76.3 | Overall we find PRO (PL+SFT) method helps, but slightly underperforms the PL+DPO method. We are not confident about this conclusion because we did not faithfully replicate PRO according to its source code, but only modified our trl implementation of DPO+PL to test PRO methods. We mainly search sft_weight $\in$ [0.01, 0.05, 0.1 0.5], and keep other hyperparameters consistent with train_hh.sh of PRO repo, or the trl/DPO settings.

Authorsrebuttal2024-08-09

We appreciate the reviewer for the prompt and positive feedback and are glad we can address the reviewer's concerns.

Reviewer EyJu6/10 · confidence 4/52024-07-19

Summary

The paper addresses a weakness of DPO: it can not deal with preference data whose number of responses is larger than 2. To extend DPO, the author proposes Noise Contrastive Estimation (NCE)-based Alignment algorithms InfoNCA. The author has shown that DPO is one of the special cases of InfoNCA. To further fix the Decreased Response Likelihood issue, the author also proposes NCA with a similar derivation. Experiments demonstrate that NCA and InfoNCA outperform baselines with reward datasets for training, with NCA showing significant performance in complex reasoning tasks like math and coding. The paper offers the community a general framework for LLM alignment from a novel perspective.

Strengths

1. The paper considers the LLM alignment task from a very novel perspective of Noise Contrastive Estimation. The proposed InfoNCA and NCA methods are not only general in form (DPO can be regarded as one of the special cases), but also elegant in math (the connection of LLM alignment and InfoNCE is quite interesting). The paper brings a brand-new view to the study of LLM alignment, which can be insightful to the research community. 2. The motivation of the paper is clear, which focuses on the limitation of DPO when training with multi-response preference data. This improving direction is quite valuable for both the research and industry domains. 3. The experimental results support the effectiveness of InfoNCA and NCA, which outperform baseline methods such as DPO, IPO, and KTO. 4. The paper is clearly written, well organized, and easy to follow.

Weaknesses

1. Fundamental Assumption: My major concern is about the fundamental assumption when deriving the InfoNCA and NCA methods. The author assumes that one of the scored responses is sampled from the optimal policy, which is practically impossible because most of the preference data are sampled from non-optimal LLM policy. In some extreme situations, all the generated responses can be harmful, while our ideal optimal policy should output harmless responses, leading to a contradiction with the author's essential assumption. A possible solution to this weakness might be mixing the human written (golden) response into the preference data, then applying NCE to the prompt and the golden response, combining InfoNCA with some prior work such as APO [1] and SPIN [2]. 2. Lack of Human Evaluation: Although the author has conducted GPT-4 evaluation in the experiments, the real human evaluation results are not reported, which should have made the paper's claims more persuasive. 3. Lack an experiments of two-response alignment: Although the paper demonstrates better performance using multi-response preference data, it is interesting to compare InfoNCA, NCA, and DPO on a binary-response preference set. 4. Lack of baseline comparison: Although DPO cannot deal with multi-response preference, one can first train an RM with the multi-response preference, and next sample two responses with scores from the learned RM. Then DPO can be conducted on the RM-scored pairwise preference data. This can be a simple but important baseline that the author should take into consideration. References: [1] Adversarial Preference Optimization: Enhancing Your Alignment via RM-LLM Game, ACL 2024 [2] Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models, ICML 2024

Questions

See Weaknesses.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The author has NOT addressed the limitations of the proposed method. One of the most important limitations can be that if all the scored responses are harmful, the proposed methods can never achieve the ideal alignment optimum, which could even enhance the harmful model outputs.

Authorsrebuttal2024-08-08

Additional Response

Dear reviewer, We copy the second part of our rebuttal response here to ease reading. In our responses, we have conducted additional experiments and compared with more baselines based on your suggestions. We hope we can address your concerns and are glad to answer any further questions you might have. Thank you again for your valuable feedback! # Official Response to Reviewer EyJu (2/2) **Q5: Lack of two-response alignment: compare InfoNCA, NCA, and DPO on a binary-response preference set.** **A5:** Perhaps we did not fully understand the reviewer's comment. As a matter of fact, **the whole section 5.2 aims to compare these algorithms on binary-response preference sets.** This includes Table 3, Figure 4 (left), Figure 5, and Figure 6(left). Note that in pairwise-preference settings, InfoNCA becomes exactly the DPO algorithm, so we mainly compare NCA and DPO (InfoNCA). We copy the main results below: **UltraFeedback (binary)** | Method | MT-bench | Alpaca | |---|---|--| | InfoNCA/DPO | 73.8 | 90.7 | | NCA | 73.2 | 89.9 | **UltraInteract (binary)** |Model|Average|Reasoning|LeetCode|HumanEval|GSMPLUS|MATH|TheoremQA|SVAMP|ASDiv| |---|---|---|---|---|---|---|---|---|---| |Mixtral-7B-SFT|24.5|60.9|3.3|28.1|28.5|5.8|7.0|26.9|35.8| |$+$DPO|25.5|**61.7**|2.2|**31.7**|12.1|6.4|9.8|34.1|46.1| |$+$**NCA**|**35.7**|60.8|**3.3**|26.8|**32.3**|**11.7**|**11.0**|**65.3**|**74.3**| |Mixtral-8*7B-SFT|56.3|**75.6**|16.7|61.0|57.6|40.1|25.9|85.9|87.5| |$+$DPO|50.1|74.9|17.2|47.6|55.8|35.3|**26.9**|67.3|75.7| |$+$**NCA**|**57.9**|**75.6**|**21.1**|**62.8**|**61.5**|**41.6**|**26.9**|**86.8**|**86.9**| **Q6: The real human evaluation results are not reported, which should have made the paper's claims more persuasive.** **A6:** **We have actually reported human evaluation results in the original paper** and would like to refer the reviewer to Table 2 (win vs DPO), and Figure 7in the paper. Still, GPT-4 based evaluation results are most indicative to ourselves because datasets are annotated by GPT-4. After all, you want your training rewards and test rewards to be aligned to make sure the algorithm itself is valid given that our work is somehow more theoretical rather than product-oriented.

Authorsrebuttal2024-08-12

Looking forward to your feedback

Dear reviewer, Thank you for your valuable comments regarding our submission. We have posted our responses and hope to address your concerns about the fundamental assumption of our method (availability of the optimal policy), as well as the comparison with several baselines such as APO and SPIN. Could you please take a look at our responses and give us some further feedback so that we can continually improve our work? Thank you so much in advance! Best regards, The Authors

Reviewer EyJu2024-08-14

Thank you for your elaborate response. Since the author has expanded with more comprehensive and detailed experiments, I have decided to raise my score.

Reviewer EZsK2024-08-09

Thanks for the response. Since most of my questions and concerns are addressed, I will increase my overall recommendation.

Area Chair FQG22024-08-09

Reviewers, please respond to the author rebuttals

Dear reviewers, this is your AC! The authors have responded to your review. Could you please take a look at the rebuttal (and respond/edit your scores if need be)? (Thank you if you have already done so) Thanks!

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC