DeCoRe: Decoding by Contrasting Retrieval Heads to Mitigate Hallucinations

I believe the paper might benefit from an aggregated score for each method and error bars or statistical tests for the main claims. This can simplify understanding the significance of the main results.

Thank you for the suggestion. Here is the aggregated score per table, and the combined average:

¹ Missing values filled with base model scores.

We calculated the aggregated scores as such:

• Faithfulness Score is an average of XSum (ROUGE-L, BERTScore-F1, factKB), MemoTrap (Macro Accuracy, Micro Accuracy), IFEval (Prompt Accuracy, Instruct Accuracy), NQ-Open EM, NQ-Swap EM.
• Factuality Score is an average of Averaged MC scores, TriviaQA EM, PopQA EM, TruthfulQA Generation (% T ∩ I), NQ-Open EM.
• CoT score: MuSiQue Closed Book without CoT EM, MuSiQue Open Book without CoT EM, MuSiQue Closed Book with CoT EM, MuSiQue Open Book with CoT EM.

From the table, we observe that the DeCoRe Entropy method achieves the highest overall average score of 48.59, indicating a consistent improvement across all evaluation metrics compared to the base model and other decoding strategies. The CAD method also demonstrates strong performance with an overall average of 48.17.

Regarding Statistical Tests, we appreciate your suggestion to include error bars or statistical tests to simplify understanding the significance of the main results. However, since our evaluations are conducted using a greedy decoding setup, thus running multiple trials would yield identical results, rendering standard statistical tests inapplicable in this context.

Nevertheless, we acknowledge the importance of statistical validation and are open to exploring alternative methods that could provide insights into the significance of our results. If you have specific statistical tests or methodologies in mind that are suitable for deterministic outputs, we would be eager to consider them in our analysis.

I believe the paper might benefit from a qualitative analysis that supports the main claims. For example, when is using conditional entropy beneficial and when does it hurt performance, and when does DeCoRe outperform the baselines?

Thank you for the suggestion! In this study, using conditional entropy to control the contrastive penalty (DeCoRe_entropy) is generally beneficial compared to using a static penalty (DeCoRe_static). In cases when DeCoRe_static outperforms DeCoRe_entropy, the difference is negligible. We look forward to further feedback on this point and aim to incorporate any further suggested analyses before the rebuttal deadline.

Regarding CAD and MemoTrap, maybe I am missing something, but perhaps it’s possible to use the task as the additional context, e.g., for the task - “Write a quote that ends in the word "talk": Don't bark if you can't”, the additional context can be “Write a quote that ends in the word "talk".

Thank you for the suggestion! We have added this experiment to the manuscript. CAD outperforms DeCoRe on MemoTrap.

Despite being second best in several tasks, as you stated, the flexibility of DeCoRe allows it to be applicable to more tasks.

Thank you for the review!

It seems that in several experiments, the improvements from DeCoRe are either marginal or it even performs significantly worse than other methods.

DeCoRe yields more accurate results than the majority of the baselines in most of the tasks. As shown in Tables 1, 2, and 3, DeCoRe achieves first place in six out of 11 tasks, second place in four, and third place in one, specifically:

• 1st place: MemoTrap, NQ-Open (Open Book), TriviaQA, PopQA, NQ-Open (Closed Book), MuSiQue
• 2nd place: XSum (factKB), NQ-Swap, TruthfulQA (MC), TruthfulQA (Gen)
• 3rd place: IFEval

In our Conclusions section (“Limitations” paragraph), we mention that, in some specific tasks, there may be baselines that achieve better results than DeCoRe. However, these baselines often offer limited improvements or yield significantly less accurate results in other tasks:

• ITI: 1st place in TruthfulQA (MC) and TruthfulQA (Gen); Significant drops in all other tasks.
• CAD: 1st place in XSum (factKB), NQ-Swap, MemoTrap; Not applicable to other tasks.

On the other hand, we show that DeCoRe can yield state-of-the-art or competitive results on a wide array of tasks.

We believe that the reason why ITI performs very well on TruthfulQA but not the others is because it is fine-tuned on TruthfulQA, which might have harmed the generalisation capability of the model. We believe that CAD yields significantly more accurate predictions on NQ-Swap and MemoTrap because it is designed to penalise the parametric knowledge of the model, thus very suitable for these tasks that require total faithfulness to the context.

Additionally, in our response to reviewer 1dQs, we also added a new experiment using multi-documents retrieval where only one document contains the answer with the Lost-in-the-Middle setup. We found that DeCoRe yields the most accurate responses, while the strongest contextual faithfulness baseline (CAD) suffers in this setup, which reinforce our previous hypothesis that CAD only performs well when the model have to be completely faithful to all given contexts:

We hope that these explanations clarify our position and look forward to any further feedback on these points.

Thank you for the detailed response! I have some follow-up questions.

• In the response, you mentioned that the manuscript was updated with the CAD results for MemoTrap, but I could not find these in the paper, am I missing something? I believe the paper could also benefit from the aggregated results.

• Regarding statistical significance, I recommend checking [1] or standard methods to calculate variance. Do you think any of these methods is applicable to your setting?

I agree that the paper has merits and I could be missing something, but my main concern is that due to the small differences between CAD, DeCoRe-Static, and DeCoRe-Entropy (<1 point on aggregate), and similarly small deltas in Tab.2 and Tab.3, some of the main claims in the paper may need to be slighly softened (specifically the improvement in accuracy of DeCoRe over baselines and advantages of entropy as a guide and especially if trends are not statistically significant).

[1] The Hitchhiker’s Guide to Testing Statistical Significance in Natural Language Processing

Thank you for the review!

Although it is good to build upon recent attention head insights for contrastive decoding, the generalizability of those identified retrieval attention seems very limited.

We agree that retrieval head insights are relatively recent. Our work represents an initial exploration in leveraging these insights in hallucination mitigation. Our experiments indicate that the approach of leveraging retrieval heads using contrastive decoding can improve contextual faithfulness across diverse tasks.

The improvement over baselines is at odds and not pronounced enough. [...] It is good to analyze the response difference to drive deeper understanding.

DeCoRe yields more accurate results than the majority of the baselines in most of the tasks. As shown in Tables 1, 2, and 3, DeCoRe achieves first place in six out of 11 tasks, second place in four, and third place in one, specifically:

• 1st place: MemoTrap, NQ-Open (Open Book), TriviaQA, PopQA, NQ-Open (Closed Book), MuSiQue
• 2nd place: XSum (factKB), NQ-Swap, TruthfulQA (MC), TruthfulQA (Gen)
• 3rd place: IFEval

In our Conclusions section (“Limitations” paragraph), we mention that, in some specific tasks, there may be baselines that achieve better results than DeCoRe. However, these baselines often offer limited improvements or yield significantly less accurate results in other tasks:

• ITI: 1st place in TruthfulQA (MC) and TruthfulQA (Gen); Significant drops in all other tasks.
• CAD: 1st place in XSum (factKB), NQ-Swap, MemoTrap; Not applicable to other tasks.

On the other hand, we show that DeCoRe can yield state-of-the-art or competitive results on a wide array of tasks.

We believe that the reason why ITI performs very well on TruthfulQA but not the others is because it is fine-tuned on TruthfulQA, which might have harmed the generalisation capability of the model. We believe that CAD yields significantly more accurate predictions on NQ-Swap and MemoTrap because it is designed to penalise the parametric knowledge of the model, thus making it very suitable for tasks that require total faithfulness to the context.

We hope that these explanations clarify our position and look forward to any further feedback on these points.

It is good to explore alternative attention selection strategies, [...] it might be interesting to see whether certain attention heads are more related to shorter vs longer context retrieval.

We have experiments with random selections in Appendix G. We referred to Appendix G.1 in the “Effect of Retrieval Head Masking on Task Performance of DeCoRe” paragraph. However, we will amend this to refer to the entire Appendix G for clarity. Regarding the idea of examining whether certain attention heads are more related to shorter vs. longer context retrieval, we agree that this is an intriguing direction, thank you for the suggestion. While it may fall outside the current scope of this paper, we believe it is a promising avenue for future work, and we plan to explore it in subsequent research.

It is good to explore some datasets with longer context (e.g., retrieving more passages) and compare the method there

We have experiments using MuSiQue which requires a longer retrieved context (i.e., multi-hop reasoning using multiple documents). Additionally, we just conducted an experiment on Lost-in-the-Middle setups which may resemble retrieving more passages as per your suggestion:

We found that DeCoRe yields the most accurate responses, while the strongest contextual faithfulness baseline (i.e., CAD) suffers in this setup, which reinforces our previous hypothesis that CAD only performs well when the model has to be completely faithful to all given contexts.

We will continue experimenting with all baselines if this addresses your concern.

Thanks the authors for providing more details and results.

There is no further question from my side.

Thank you for the review!

The model's performance improvements are inconsistent across different datasets, and it fails to outperform baseline models in many settings (e.g., XSum, IFEval). This is inconsistent with the summary conclusions presented in the abstract.

DeCoRe yields more accurate results than the majority of the baselines in most of the tasks. As shown in Tables 1, 2, and 3, DeCoRe achieves first place in six out of 11 tasks, second place in four, and third place in one, specifically:

• 1st place: MemoTrap, NQ-Open (Open Book), TriviaQA, PopQA, NQ-Open (Closed Book), MuSiQue
• 2nd place: XSum (factKB), NQ-Swap, TruthfulQA (MC), TruthfulQA (Gen)
• 3rd place: IFEval

In our Conclusions section (“Limitations” paragraph), we mention that, in some specific tasks, there may be baselines that achieve better results than DeCoRe. However, these baselines often offer limited improvements or yield significantly less accurate results in other tasks:

• ITI: 1st place in TruthfulQA (MC) and TruthfulQA (Gen); Significant drops in all other tasks.
• CAD: 1st place in XSum (factKB), NQ-Swap, MemoTrap; Not applicable to other tasks.

On the other hand, we show that DeCoRe can yield state-of-the-art or competitive results on a wide array of tasks.

We believe that the reason why ITI performs very well on TruthfulQA but not the others is because it is fine-tuned on TruthfulQA, which might have harmed the generalisation capability of the model. We believe that CAD yields significantly more accurate predictions on NQ-Swap and MemoTrap because it is designed to penalise the parametric knowledge of the model, thus making it very suitable for these tasks that require total faithfulness to the context.

Additionally, in our response to reviewer 1dQs, we also added a new experiment using multi-document retrieval where only one document contains the answer with the Lost-in-the-Middle setup. We found that DeCoRe yields the most accurate responses, while the strongest contextual faithfulness baseline (CAD) suffers in this setup, which reinforces our previous hypothesis that CAD only performs well when the model has to be completely faithful to all given contexts:

We hope that these explanations clarify our position and look forward to any further feedback on these points.

The paper only compares a limited number of baselines, primarily focusing on contrastive methods, without evaluating against many related models.

We compare DeCoRe against five strong baselines on top of the greedy decoding baseline:

• 3 Contrastive decoding methods (CAD, CD, DoLa). Despite being not applicable to some tasks, We included CAD as it is a very strong baseline.
• 2 Non-contastive decoding methods: ITI and Activation Decoding. ITI specifically requires fine-tuning, as opposed to the other selected baselines.

We believe this is sufficient; however, if the reviewer would like to specify additional baselines, we will aim to incorporate them in the revision before the rebuttal deadline.

The approach in this paper appears to be related to self-consistency and should better include a discussion on this topic, along with relevant baseline comparisons.

Thank you for this suggestion – indeed, DeCoRe can be seen as related to self-consistency as it involves combining multiple predictions; however, while self-consistency mainly relies on majority voting among generations from a single model, DeCoRe attempts to amplify the difference between a base model and a masked model. We will add a discussion on this link between self-consistency and DeCoRe in the Related Work section of the revised version of the paper. Specifically: “DeCoRe can also be viewed as an asymmetric version of self-consistency (Wang et al., 2023); rather than relying on majority voting among generations from a single model, DeCoRe amplifies the difference between a base model and its masked variant.”

We hope that this answers your concerns and we look forward to any further feedback.

Dear Reviewer,

Since the deadline for the rebuttal phase is approaching, we kindly ask you to confirm whether we have sufficiently addressed your comments or if there are any remaining concerns.

Thank you!

Authors

Thank you for the review!

it is required to provide evidence to show this decoding method can generate text with high quality with low hallucination.

Thank you for spotting this – we included several examples illustrating the fluency and coherence of the generated outputs of DeCoRe in Appendix K.2; however, this was not mentioned in the main text, which may have caused some oversight. Following your comment, we will expand more on the text quality generated by DeCoRe and refer to Appendix K at the end of Section 4 in the revised version of the paper.

It seems that in the experiments, many baselines can achieve better results than DeCoRe.

DeCoRe yields more accurate results than the majority of the baselines in most of the tasks. As shown in Tables 1, 2, and 3, DeCoRe achieves first place in six out of 11 tasks, second place in four, and third place in one, specifically:

• 1st place: MemoTrap, NQ-Open (Open Book), TriviaQA, PopQA, NQ-Open (Closed Book), MuSiQue
• 2nd place: XSum (factKB), NQ-Swap, TruthfulQA (MC), TruthfulQA (Gen)
• 3rd place: IFEval

In our Conclusions section (“Limitations” paragraph), we mention that, in some specific tasks, there may be baselines that achieve better results than DeCoRe. However, these baselines often offer limited improvements or yield significantly less accurate results in other tasks:

• ITI: 1st place in TruthfulQA (MC) and TruthfulQA (Gen); Significant drops in all other tasks.
• CAD: 1st place in XSum (factKB), NQ-Swap, MemoTrap; Not applicable to other tasks.

On the other hand, we show that DeCoRe can yield state-of-the-art or competitive results on a wide array of tasks.

We believe that the reason why ITI performs very well on TruthfulQA but not the others is because it is fine-tuned on TruthfulQA, which might have harmed the generalisation capability of the model. We believe that CAD yields significantly more accurate predictions on NQ-Swap and MemoTrap because it is designed to penalise the parametric knowledge of the model, thus very suitable for these tasks that require total faithfulness to the context.

Additionally, in our response to reviewer 1dQs, we also added a new experiment using multi-documents retrieval where only one document contains the answer with the Lost-in-the-Middle setup. We found that DeCoRe yields the most accurate responses, while the strongest contextual faithfulness baseline (CAD) suffers in this setup, which reinforce our previous hypothesis that CAD only performs well when the model have to be completely faithful to all given contexts:

We hope that these explanations clarify our position and look forward to any further feedback on these points.

The paper seems a combination of the existing findings, which applies the retrieval heads (Wu et al., 2024) to the Contrastive Decoding (Li et al., 2023) method, potentially limiting the originality of the contribution.

Indeed, we agree that DeCoRe is grounded on intuitions from the retrieval heads and contrastive decoding literature; however, our work is the first to integrate retrieval head masking and contrastive decoding for hallucination mitigation. Furthermore, we propose a simple method for automatically selecting the DeCoRe hyper-parameters. We clarify this in the revised version.

Main comments addressed in the response

All Reviewers

Inconsistent improvements across tasks: This is the main comment mentioned by all reviewers. In the revised version, we have addressed this by:

• Adding Additional Lost-in-the-Middle Experiments (Table 2): These experiments further validate DeCoRe's effectiveness in contexts with distractor documents.
• Including Statistical Analyses (Appendix H): We performed statistical significance tests to support our claims and soften our claims where necessary.
• Providing Aggregate Scores in Table 1, 2, 3, 4, and 5: We included average performance metrics to give a clearer picture of DeCoRe's overall effectiveness across tasks.

Reviewer 6H8B acknowledged our efforts, stated that we have made “significant efforts to address these limitations during the discussion period (by adding experiments, introducing statistical tests, and softening the main claims)”. All changes are currently written in blue in the revised manuscript.

In summary, DeCoRe achieves more accurate results compared to most baselines in the majority of tasks:

• 1st place in 10 out of 17 tasks (NQ-Open (Oracle), NQ-Open (Gold at 1), NQ-Open (Gold at 4th), NQ-Open (Gold at 9th), NQ-Open (Closed book), TriviaQA, PopQA, MuSiQue (without CoT, Closed Book), MuSiQue (without CoT, Open Book), MuSiQue (with CoT, Open Book));
• 2nd place in 6 out of 17 tasks (XSum, MemoTrap, NQ-Swap, TruthfulQA (MC), TruthfulQA (Gen), MuSiQue (with CoT, Closed Book));
• <2nd place in 1 out of 17 task (IFEval).

In cases where a baseline outperforms DeCoRe on specific tasks, these improvements are often marginal or come at the cost of significantly worse performance on other tasks. For instance:

• ITI: 1st place in TruthfulQA, while yielding significant drops in all other tasks;
• CAD: 1st place in XSum (factKB), NQ-Swap, MemoTrap, while yielding significant drops in Lost-in-the-Middle experiments and not applicable to factuality evaluation tasks.

DeCoRe consistently yields state-of-the-art or competitive performance across a wide variety of tasks, making it a broadly applicable and reliable solution.

Specific Reviewers

Reviewer yWX4

Reviewer yWX4 assigned a borderline rating of 5, while acknowledging several strengths of our work. We have addressed their comments as follows:

• Weakness 1 (Inconsistent Improvements): Addressed as per the actions taken for all reviewers. Additionally, Reviewers 6H8B and 1dQs have acknowledged our efforts in addressing this limitation.
• Weakness 2 (Limited Baseline Comparisons): In our response, we clarified that DeCoRe has been compared against five strong baselines in addition to the greedy decoding baseline. These include three contrastive decoding methods (CAD, CD, DoLa) and two non-contrastive methods (ITI, Activation Decoding). Reviewers 9XYH, 6H8B, and 1dQs have acknowledged the thoroughness of our experiments.
• Weakness 3 (Relation to Self-Consistency): We added a discussion in the Related Work section on the relationship between DeCoRe and self-consistency methods.

Reviewer 9XYH

Reviewer 9XYH assigned a borderline rating of 5, while acknowledging the comprehensive experiments that demonstrates the capability of DeCoRe in mitigating hallucinations. We have addressed their comments as follows:

• Weakness 1 (Qualitative Examples of Coherent Text): We added a reference to Appendix L.2, which includes examples illustrating the fluency and coherence of outputs generated by DeCoRe.
• Weakness 2 (Baselines Achieving Better Results): Addressed as per the actions taken for all reviewers. Reviewers 6H8B and 1dQs have acknowledged our efforts in addressing this limitation.
• Weakness 3 (Novelty of the Work): In our rebuttal, we clarified that our work is the first to integrate retrieval head masking and contrastive decoding for hallucination mitigation. Additionally, we introduced a simple method for automatically selecting the DeCoRe hyperparameters. Reviewers yWX4, 6H8B, and 1dQs have acknowledged the novelty of our approach.