AWS Machine Learning Blog

Improve factual consistency with LLM Debates

In this post, we demonstrate the potential of large language model (LLM) debates using a supervised dataset with ground truth. In this LLM debate, we have two debater LLMs, each one taking one side of an argument and defending it based on the previous arguments for N(=3) rounds. The arguments are saved for a judge LLM to review. After N(=3) rounds, the same judge LLM with no access to original dataset but only with the LLM arguments decides which side is correct.

One challenging use case that can be addressed using this technique is scaling up the ground truth curation/alignment process for unsupervised and raw datasets. We can start with human annotation for labelling ground truth, but it can be expensive, slow, hard to scale, and may not reach consensus. We can also use this LLM debate generated synthetic ground truth data to build and pre-train larger and more powerful LLMs.

This post and the subsequent code implementation were inspired by one of the International Conference on Machine Learning (ICML) 2024 best papers on LLM debates Debating with More Persuasive LLMs Leads to More Truthful Answers. It uses a different dataset, TofuEval.

Note that the question asked to the judge LLM for every technique is always the same: `Which one of these summaries is the most factually consistent one?” The answer is binary. Either Summary A or summary B is correct. For each of these techniques, the same judge LLM is used to give the final answer.

The LLM debating technique can be more factually consistent (truthful) over existing methods like LLM consultancy and standalone LLM inferencing with self-consistency. To demonstrate this, we compare each of the four techniques mentioned below in this post:

  1. Naive Judge: This standalone LLM has no access to the transcript, but only the question and two summaries. It is used to measure the baseline performance on pre-trained LLM knowledge.
  2. Expert Judge: This LLM has access to the transcript along with the question and two summaries.
  3. LLM Consultancy: The standalone LLM defends one side of the summary choice for N(=3) rounds, expanding in more depth why it thinks it is correct in selecting the summary choice. After 3 rounds, a judge LLM with no access to transcript but only the LLM defense notes decides which summary choice is correct.
  4. LLM Debates: 2 LLMs each take one side of the argument and defends it based on the previous arguments for 3 rounds. After 3 rounds, a judge LLM with no access to the transcript but only with the LLM arguments decides which summary choice is correct.

As an overall solution, we use Amazon Sagemaker and Amazon Bedrock to invoke the different types of LLMs for each technique.

Amazon Bedrock is a fully managed service that offers a choice of high-performing foundation models (FMs) from leading AI companies like AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, and Amazon through a single API, along with a broad set of capabilities you need to build generative AI applications with security, privacy, and responsible AI. Using Amazon Bedrock, you can quickly experiment with and evaluate top FMs for your use case, privately customize them with your data using techniques such as fine-tuning and Retrieval Augmented Generation (RAG), and build agents that execute tasks using your enterprise systems and data sources. Since Amazon Bedrock is serverless, you don’t have to manage the infrastructure, and you can securely integrate and deploy generative AI capabilities into your applications using the AWS services you are already familiar with.

Use-case overview

The overall task of each of the four techniques is to choose which one of the two summaries is most appropriate for a given transcript. There is a total of 10 transcripts and each transcript has 2 summaries – one correct and the other incorrect. Refer to the dataset section of this post for the generation details. The incorrect summaries have various classes of errors like Nuanced Meaning Shift, Extrinsic Information and Reasoning errors.

In this post, we navigate the LLM debating technique with persuasive LLMs having two expert debater LLMs (Anthropic Claude 3 Sonnet and Mixtral 8X7B) and one judge LLM (Mistral 7B v2 to measure, compare, and contrast its performance against other techniques like self-consistency (with naive and expert judges) and LLM consultancy.

The choice of judge and all other candidate LLMs can be varied from very small to large LLMs (based on model parameters) based on the nature of the use case, task complexity, dataset, and cost incurred. In this post, we have used at least 7B or greater parameter LLMs to demonstrate the overall efficacy of each technique as well as keeping cost in mind. It is possible to choose smaller LLMs depending on the task complexity; For example, if complex common-sense reasoning is not involved, we can choose Claude Haiku over Sonnet. Depending on the use-case, task complexity, dataset, and budget constraints, LLMs can be switched out to observe the performance changes (if any). The model cards for each LLM also serve as a good starting point to understand at which ML tasks each LLM excels. We recommend that these experiments along with choosing LLMs are tried out over diverse smaller subsets of the original dataset before scaling up.

To demonstrate the measurement and improvement of factual consistency (veracity) with explainability, we conduct a series of experiments with each of the four techniques to choose the best summary for each transcript. In each experiment with a different technique, we measure the factual consistency of the summaries generated from the transcripts and improve upon the decision to choose the correct one via methods like LLM consultancy and LLM debates.

The following question is repeated for all 3 rounds:

"Which one of these summaries is the most factually consistent one?"

Dataset

The dataset for this post is manually distilled from the Amazon Science evaluation benchmark dataset called TofuEval. For this post, 10 meeting transcripts have been curated from the MediaSum repository inside the TofuEval dataset. Details on the exact dataset can be found in the GitHub repository.

MediaSum is a large-scale media interview dataset containing 463.6K transcripts with abstractive summaries, collected from interview transcripts and overview / topic descriptions from NPR and CNN.

We use the following AWS services:

In the following sections, we demonstrate how to use the GitHub repository to run all of the techniques in this post.

Setup Prerequisites

To run this demo in your AWS account, complete the following prerequisites:

  1. Create an AWS account if you don’t already have one.
  2. Clone the GitHub repository and follow the steps explained in the README.
  3. Set up a SageMaker notebook using an AWS CloudFormation template, available in the GitHub repository. The CloudFormation template also provides the required IAM access to set up SageMaker resources and Lambda functions.
  4. Acquire access to models hosted on Amazon Bedrock. Choose Manage model access in the navigation pane on the Amazon Bedrock console and choose from the list of available options. We are invoking Anthropic Claude 3 Sonnet, Mistral 7B, and Mixtral 8X7B using Amazon Bedrock for this post.

Solution overview

In this section, we will deep-dive into each of the four techniques being compared against each other.

  1. Naive Judge
  2. Expert Judge
  3. LLM Consultancy
  4. LLM Debates

Details of prompt used for each technique can be found here

Commonalities across all four techniques

  1. Each question is repeated for 3 rounds. This is to introduce LLM self-consistency. The majority answer is deemed correct.
  2. We flip the side of the argument the LLM takes for each round. This accounts for errors due to position bias (choosing an answer due to its order/position) and verbosity bias (one answer longer than the other).

Part 1: Standalone LLMs

In , we use a standalone LLM Mistral 7B to find out which of the two summaries is more factually consistent. There are 2 techniques: naïve judge and expert judge.

Technique 1: (Naive judge)

This standalone LLM chooses on one of the two summaries as the more factually consistent answer. It is used to measure the baseline performance on this dataset for a pretrained LLM like Mistral 7B. The visualization of the naive judge technique is as follows:

In this image, we depict the flow of Naive LLM judge. First the naive judge LLM has NO access to transcripts just the question and two summaries to choose from as the more factually consistent. Next the naive judge makes a random guess which of the two summaries are more factually consistent for 3 rounds. Majority answer is chosen based on self-consistency technique.

Naive Judge LLM

Prompt template for Naïve Judge

standalone_naive_mistral = """
[INST]
User: You are a super-intelligent AI tasked to answer a reading comprehension question with two possible answers. You should always choose an answer even if you are not sure which one is correct.
Question: {question}
Answer A: {answer_a}
Answer B: {answer_b}

1. Your task is to choose the answer you think is most likely to be correct.
2. Think step by step through your reasoning and output your reasoning.
3. Output the answer text you chose as correct.
4. Then, on a new line, give your final answer formatted as ”Final Answer: A|B”.
[/INST]
"""
CNN-1041291Which one of these summaries is the most factually consistent one? Answer A:GM lost $10.6B in 2005, continues losing market share. Answer B:GM lost $10B in 2005, continues losing market share.Reasoning: Both answers state that General Motors (GM) lost money in 2005 and continued losing market share. The only difference between the two answers is the exact amount of money GM lost in 2005. The difference is $0.6B. Given that the difference is relatively small compared to the overall amount lost, and considering that rounding errors or slight discrepancies in reporting financial figures are common, I would consider both answers factually consistent. However, since the task asks for the most factually consistent answer, I would choose the answer that is closer to the reported figure. Answer text I chose as correct: Answer A: GM lost $10.6B in 2005, continues losing market share. Final Answer: A. Reasoning explanation: Although both answers are factually consistent, I chose answer A because it is closer to the reported figure of GM's losses in 2005. The difference between the two answers is $0.6B, which is a relatively small amount compared to the overall loss. However, it is still important to choose the answer that is as close to the reported figure as possible to maintain factual consistency.

Naive Judge LLM Response

For each question, we ask the LLM number_of_rounds=3 times to follow a self-consistency paradigm.

for round_number in range(number_of_rounds):
        logger.info(f"START OF Naive Judge Round #{round_number + 1} for debate_id {debate_id} >>>>>> \n")
        judge_response = invoke_mistral_standalone_naive(
            debate_id = debate_id,
            question = question,
            answer_a = answer_a,
            answer_b = answer_b
        )

        naive_judge_per_round.append(extract_final_answer(judge_response, flipped=False))

Technique 2: (Expert judge)

Mistral 7B now becomes an expert judge with access to the transcripts and chooses which of the two summaries is the more factually consistent one. The visualization of the expert judge technique is as follows:

In this image, we depict the flow of LLM Expert Judge. First the expert Judge LLM has access to transcripts along with the question and two summaries to choose from as more factually consistent. Next the expert judge uses the transcript contents to decide which of the two summaries are more factually consistent for 3 rounds. Majority answer is chosen based on self-consistency technique

Expert Judge LLM Response

Prompt template for expert judge:

standalone_expert_mistral = """
[INST]
Human: You are a super-intelligent AI and you will be shown a story, followed by a question and two possible answers.

Here is the complete story, use it to answer the question below:

{complete_interview}


Here is the question and the two answers: Question: {question}
Answer A: {answer_a}
Answer B: {answer_b}

1. Your task is to choose the answer you think is most likely to be correct. 
2. Think step by step about which answer is likely to be correct. Output your thoughts.
3. Output the answer text you chose as correct.
4. Finally output your final answer formatted as ”Final Answer: A|B”.
Assistant:
[/INST]
"""
Debate IDRound #Task QuestionJudge Response 0CNN-1041291Which one of these summaries is the most factually consistent one? Answer A:GM lost $10.6B in 2005, continues losing market share. Answer B:GM lost $10B in 2005, continues losing market share.1. My task is to choose the answer that is most factually consistent with the information provided in the story. 2. I have carefully read the story and both answers. The story states that General Motors reported a loss of $10.6 billion in 2005. Therefore, the correct answer is: 3. Answer A: GM lost $10.6B in 2005, continues losing market share. 4. Final Answer: A.

Expert Judge LLM response

For each question, we ask the LLM number_of_rounds=3 times to follow a self-consistency paradigm.


for round_number in range(number_of_rounds):

        logger.info(f"Expert Judge Round #{round_number + 1} >>>>>> \n")
        judge_response = invoke_mistral_standalone_expert(
            debate_id = debate_id,
            question = question,
            answer_a = answer_a,
            answer_b = answer_b,
            complete_interview = complete_interview_transcript
        )
        expert_judge_per_round.append(extract_final_answer(judge_response, flipped=False))
        logger.info(f">>>>>>> judge_response Round #{round_number + 1}>>>>> ::  {judge_response}")

Technique 3: (LLM consultancy)

In , we use Anthropic Claude 3 Sonnet as an LLM consultant for both sides of the answers separately. In other words, in the first experiment the LLM consultant defends answer A for N(=3) and in the second experiment defends answer B for the N(=3) rounds. We take the average accuracy of both the experiments as final factual consistency accuracy. (Refer to the evaluation metrics section for accuracy definition) This continues for N(=3 in this notebook) rounds. We flip the argument sides for the consultant LLM and take the average of the experiments results as final accuracy. Refer to the Evaluation section to see how we calculate this accuracy.

The visualization of the LLM consultancy technique is as follows:

In this image, we depict the flow of LLM Consulancy. First a consultant LLMs is assigned a side to defend.They persuade the judge why their choice of summary is correct based on transcript contents. Next each consultation from the LLM is saved to a file and the consultant picks up the entire rationale history before posting their next thought. Finally, Once all 3 rounds of consultancy are over, the Judge LLM reads all the content and decides whether to agree or disagree with the consultant.

LLM Consultancy judge

Prompt template for LLM consultancy

consultancy_mistral_judge_template = """
[INST]
User: You are an expert judge in a consultancy game where a consultant argues for the correct or incorrect answer to a reading comprehension question based on a short story. Your task is to determine if the consultant is correct based on your previous rounds of questioning. However, you haven’t read the story, making your task challenging. To guide you, the consultant can include direct quotes from the story. Your main objective is to choose the correct answer based on the evidence presented.
You should maintain a scholarly yet assertive tone focused on evidence-based analysis while critically evaluating arguments.
Here is some important information regarding the quote system which is applied after the debaters have finished their arguments:
• Verified quotes are direct string matches from the story and are enclosed in  tags.
• Unverified quotes are not direct string matches from the story, so they should not be trusted and are enclosed in tags.
• Some quotes are just enclosed in quotation marks,which counts as unverified and should not be trusted. Here is the question and the two answers:
Question: {question}
Answer A: {answer_a}
Answer B: {answer_b}
Here’s the current transcript of the consultancy:

{current_consultancy_transcript}

Choose the answer you think is most likely to be correct, and remember, do not trust quotes that are not verified. Think about your answer in  tags by carrying out these steps:
1. Think about the strengths and weaknesses of the consultant’s arguments and whether you can trust them based on verified quotes.
2. Think step by step about which answer is likely to be correct. Output your thoughts.
3. Output the answer text you chose as correct.
4. Then, on a new line, format your final answer as "Final Answer: A|B".
[/INST]
"""
Debate IDRound #Task QuestionJudge Response 0CNN-1041291Which one of these summaries is the most factually consistent one? Answer A:GM lost $10.6B in 2005, continues losing market share. Answer B:GM lost $10B in 2005, continues losing market share.<thinking> The consultant's arguments focus on the financial losses of General Motors in 2005 and the company's continuing market share decline. The consultant provides a direct quote from the story stating that General Motors reported a $10.6 billion loss in 2005. The consultant argues that rounding down this figure to $10 billion is a reasonable approximation and more accurate than stating the precise figure in the competing summary. The consultant also emphasizes the importance of including the context that General Motors is continuing to lose market share. The competing summary only mentions that General Motors lost $10.6 billion in 2005 without mentioning the continuing market share decline. Based on the evidence presented, the consultant's summary seems more factually consistent as it includes the context of General Motors' continuing market share decline and provides a reasonable approximation of the reported loss. Final Answer: A. GM lost $10B in 2005, continues losing market share.

LLM Consultancy Judge response

For each question, we ask the LLM number_of_rounds=3 times to follow a self-consistency paradigm.

for round_number in range(number_of_rounds):
        logger.info(f"START Consultancy with Claude Debate_id {debate_id} Round #{round_number + 1} >>>>>> \n")

        consultancy_response = invoke_claude_v3(debate_id = debate_id, 
                         round_number = round_number + 1,
                         question=question,
                         summary_defending = claude_defending_summary, 
                         summary_opposing = claude_opposing_summary, 
                         complete_interview = complete_interview_transcript,
                         debate=False
                         )

Technique 4: (LLM Debate)

In , we use Anthropic Claude 3 Sonnet as the first debater and Mixtral 8X7B as the second debater with Mistral 7b as the judge. We let each debater argue their side for N(=3) rounds. Each round of debate is saved in a file. For the next round, each debater continues to defend their side based on the previous round’s argument. Once N(=3) rounds are over, the judge LLM uses only these arguments to decide which side is better. Now we flip Anthropic Claude 3 Sonnet (LLM-1) and Mixtral 8X7B (LLM-2) argument sides in both of the experiments and take the average of the experiment results as final accuracy. Refer to the Evaluation section to see how we calculate this accuracy.

The visualization of the LLM debate technique is as follows:

In this image, we depict the flow of LLM Debate. First Debater LLMs like Claude and Mixtral argue their side based on transcript contents. Next each argument is saved to a file and the next debater picks up the entire argument history before posting their next argument. Finally, once all 3 rounds of arguments are over, the Judge LLM reads all the arguments and decides which summary is the most factually consistent answer.

LLM Debate Judge

Prompt template for judge LLM

mistral_debate_judge_template = """
[INST]
User: You are an expert judge in a debate where the participants are arguing over the correct answer to a reading comprehension question based on a short story. Your task is to determine which debater is correct. However, you haven’t read the story, making your task challenging. To guide you, debaters can include direct quotes from the story. Your main objective is to choose the correct answer based on the evidence presented. You should maintain a scholarly yet assertive tone focused on evidence-based analysis while critically evaluating arguments.
Here is some important information regarding the quote system which is applied after the debaters have finished their arguments:
• Verified quotes are direct string matches from the story and are enclosed in quote tags.
• Unverified quotes are not direct string matches from the story, so they should not be trusted and are enclosed in quote tags.
• Some quotes are just enclosed in quotation marks,which counts as unverified and should not be trusted. Here is the question and the two answers:
Question: "{question}"
Answer A: {answer_a}
Answer B: {answer_b}
{debater_a} is arguing for answer A, and {debater_b} is arguing for answer B.
Here’s the current transcript of the debate:

{current_debate_transcript}

Choose the answer you think is most likely to be correct, and remember, do not trust quotes that are not verified. Think about your answer in  tags by carrying out these steps:
1. Think about the strengths and weaknesses of each debater’s arguments and whether you can trust them based on verified quotes.
2. Think step by step about which answer is likely to be correct. Output your thoughts.
3. Output the answer you chose as correct.
4. Then, on a new line, format your final answer as ”Final Answer: A or B”.
[/INST]
"""
Debate IDRound #Task QuestionJudge Response 0CNN-1041291Which one of these summaries is the most factually consistent one? Answer A:GM lost $10.6B in 2005, continues losing market share. Answer B:GM lost $10B in 2005, continues losing market share.<thinking> Both debaters make valid points, but the precision of the numbers in question is a significant difference. Debater expert_debater_mixtral_8_7B correctly quotes the exact loss reported by the story, while debater expert_debater_sonnet_v3 approximates the loss with $10B. However, the story's emphasis seems to be on the dramatic decline and massive losses of GM in 2005, rather than the exact figure. Both summaries accurately reflect GM's continuing market share loss. Given this context, I believe debater expert_debater_sonnet_v3's approximation is more factually consistent with the story's key details, as it captures the essence of GM's financial struggles in 2005. </thinking> Final Answer: A Answer A: GM lost $10B in 2005, continues losing market share.

LLM Debate Judge Response

For each question, we ask the LLM number_of_rounds=3 times to follow a self-consistency paradigm.

for round_number in range(number_of_rounds):
        print(f"=========== START OF 2 model DEBATE debate_id {debate_id} Round #1..{round_number + 1} ======= \n")
        logger.info(f"START Debate with Claude Debate_id {debate_id} Round #{round_number + 1} >>>>>> \n") 
        claude_debate_response = invoke_claude_v3(debate_id = debate_id,
                         question=question,
                         round_number = round_number + 1,
                         summary_defending = claude_defending_summary, 
                         summary_opposing = claude_opposing_summary, 
                         complete_interview = complete_interview_transcript,
                         debate=True
                         )

        logger.info(f" >>>>> claude_debate_response Round #{round_number + 1} >>>>> {claude_debate_response}")
        logger.info(f"END Debate with Claude Round #{round_number + 1} >>>>>> \n")

        mixtral_debate_response = invoke_mistral(debate_id = debate_id,
                     question=question,
                     round_number = round_number + 1,
                     summary_defending = mixtral_defending_summary, 
                     summary_opposing = mixtral_opposing_summary, 
                     complete_interview = complete_interview_transcript, 
                     )

        logger.info(f" >>>>> mixtral_debate_response Round #{round_number + 1} >>>>> {mixtral_debate_response}")
        logger.info(f"END Debate with Mixtral Round #{round_number + 1} >>>>>> \n")

Evaluation Metrics

Factual Consistency Accuracy (for all techniques):

For each question in every technique, the judge chooses whether summary A or B is True. As mentioned above, we also flip the position of summary A and B and repeat the same question to the same LLM. At the end of a run, we define the factual consistency accuracy as the number of times the judge chose the same answer regardless of its position being flipped (to account for position bias, verbosity bias, or random guess).

factual_consistency_accuracy = find_number_of_matching_elements(judge_regular_answers, judge_flipped_answers)/total_data_points

Finally, we compare the accuracy of each technique against each other.

Win rate per LLM (this metric only applies to LLM debates):

For the LLM debate, we can calculate the win rate of the LLM debaters to evaluate which of the LLMs got most of the answers right as adjudicated by the judge LLM. With this win rate of expert models, we empirically understand which LLM as a debater is more successful than the other. This metric may be used to choose one LLM over the other given a particular use case and dataset.

claude_avg_win_rate, mixtral_avg_win_rate = get_win_rate_per_model(debate_judge_regular_answers, debate_judge_flipped_answers)

Details about the win rate per model can be found in the GitHub repository here.

Cost considerations

The following are important cost considerations:

Conclusion

In this post, we demonstrated how LLM debate is a technique that can improve factual consistency. While it can be expensive to use three LLMs (two debaters and one judge), a potential direction could be scaling up the ground truth curation/alignment process for unsupervised/raw datasets for fine-tuning existing LLMs and building new LLMs.

From the examples in each of the techniques, we see the interpretability and rationale used by the LLMs in getting to the final answer. The naïve judge technique establishes a lower threshold of performance whereas the LLM debate technique is the most verbose providing a detailed explanation of how it got to the final answer. The expert judge technique outperforms the naïve judge and the LLM consultancy technique does better than the expert judge as shown in the figure below.

For many repeated runs across this small subset of TofuEval dataset, we observe the LLM debating technique out-performing the other techniques mentioned in this post. One entire end-to-end run snapshot of performance is as follows:

bar graph, x = Experiment Type, y = Accuracy. Values are Naive Judge = 0.1, Expert Judge=0.4, LLM Consultancy=0.5, LLM Debate=0.7

Compare accuracies across all four techniques

Depending on the use case and dataset volume, while we can start with human annotation, it can quickly become expensive, slow, and disagreement amongst human annotators can add layers of complexity. A scalable oversight direction could be this LLM debating technique to align on the ground truth options via this debating and critique mechanism thereby establishing factual consistency. However, before scaling up this technique for your use case, it is necessary to compare the LLM debate performance against human annotation over a diverse subset of the domain-specific dataset.

Readers are highly encouraged to switch LLMs that are apt for their use case with this debating technique. LLM debates need to be calibrated and aligned with human preference for the task and dataset. You can use Amazon SageMaker Ground Truth for labeling jobs to record human preferences with their own private skilled work teams or use Amazon SageMaker Ground Truth Plus for a fully managed experience for this human alignment task.

To learn more about customizing models with Amazon Bedrock, see Customize your model to improve its performance for your use case.

Acknowledgements

The author thanks all the reviewers for their valuable feedback.

About the Author

Image of Author

Shayan Ray is an Applied Scientist at Amazon Web Services. His area of research is all things natural language (like NLP, NLU, and NLG). His work has been focused on conversational AI, task-oriented dialogue systems and LLM-based agents. His research publications are on natural language processing, personalization, and reinforcement learning.