Improving Factuality and Reasoning in Language Models through Multiagent Debate

Du, Yilun; Li, Shuang; Torralba, Antonio; Tenenbaum, Joshua B.; Mordatch, Igor

doi:10.48550/arxiv.2305.14325

PreprintCited by 175Topic ModelingNatural Language Processing TechniquesMultimodal Machine Learning Applications

Improving Factuality and Reasoning in Language Models through Multiagent Debate

Open access

Du, Yilun, Li, Shuang, Torralba, Antonio, Tenenbaum, Joshua B., Mordatch, Igor·2023·arXiv

DOI 10.48550/arxiv.2305.14325arXiv 2305.14325

Open this paper in CoChat — free account Open PDF

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities in language generation, understanding, and few-shot learning in recent years. An extensive body of work has explored how their performance may be further improved through the tools of prompting, ranging from verification, self-consistency, or intermediate scratchpads. In this paper, we present a complementary approach to improve language responses where multiple language model instances propose and debate their individual responses and reasoning processes over multiple rounds to arrive at a common final answer. Our findings indicate that this approach significantly enhances mathematical and strategic reasoning across a number of tasks. We also demonstrate that our approach improves the factual validity of generated content, reducing fallacious answers and hallucinations that contemporary models are prone to. Our approach may be directly applied to existing black-box models and uses identical procedure and prompts for all tasks we investigate. Overall, our findings suggest that such "society of minds" approach has the potential to significantly advance the capabilities of LLMs and pave the way for further breakthroughs in language generation and understanding.

AI snapshot

Written from the paper’s full text.

TL;DR

Multiagent debate, where multiple LLM instances propose answers and critique each other over rounds, improves reasoning and factual accuracy over single-agent, reflection, and majority-vote baselines across six tasks. The approach uses black-box models and identical prompts, though it is more computationally expensive and can converge confidently on incorrect answers.

Key findings

Debate improved reasoning: arithmetic accuracy was 81.8 ± 2.3 vs 67.0 ± 4.7 for single agent, 72.1 ± 4.5 for reflection, and 69.0 ± 4.6 for majority voting; GSM8K was 85.0 ± 3.5 vs 77.0 ± 4.2, 75.0 ± 4.3, and 81.0 ± 3.9; chess pawn score was 122.9 ± 7.6 vs 91.4 ± 10.6, 102.1 ± 11.9, and 102.2 ± 6.2.
Debate improved factual validity: biographies accuracy was 73.8 ± 2.3 vs 66.0 ± 2.2 for single agent and 68.3 ± 2.9 for reflection; MMLU was 71.1 ± 4.6 vs 63.9 ± 4.8 and 57.7 ± 5.0; chess move validity was 45.2 ± 2.9 vs 29.3 ± 2.6 and 38.8 ± 2.9.
On 20 GSM8K problems debated between ChatGPT and Bard, Bard solved 11, ChatGPT solved 14, and joint multi-agent debate solved 17.
Performance increased with more agents and more debate rounds; additional rounds above four gave similar final performance to 4 rounds. Longer debate prompts led to slower convergence but better final consensus, and different persona prompts improved MMLU from 71.1 to 74.2.

Methods

Multiple language model instances independently answer a query, then each reads and critiques the other agents’ responses and updates its own answer over multiple rounds; the same procedure and prompts are used across tasks, requiring only black-box model access.
Evaluated on six tasks: arithmetic (100 generated tasks), GSM8K (100 problems), chess move prediction (300 games), biographies (524 well-known computer scientists), MMLU (100 questions), and chess move validity (100 tasks), using gpt-3.5-turbo-0301; main debate runs used 3 agents and 2 rounds.
Compared against single-agent, reflection, and majority-voting baselines; reported accuracy, Stockfish pawn score, and factual validity, with ablations on number of agents, rounds, prompt length, summarization, and different models/personas.

Limitations

Stated: multiagent debate is more computationally expensive because it requires multiple generations and debate rounds; longer debates sometimes exceed models’ processing, with models focusing mainly on recent generations; debates converge but not necessarily to correct answers, and models can confidently affirm incorrect answers because they do not correctly express uncertainty.
Stated: the biography evaluation metric may miss incorrect generated facts not captured in the ground-truth bullets.
Evident from the text: main evaluations often use 100 examples per task and 3 agents with 2 rounds due to computational expense, and the cross-model comparison uses only 20 GSM8K problems.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

DataCite · 10.48550/arxiv.2305.14325

Cites

45 of 45 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 4 cited sources checked · 0 supported · 0 flagged

Findingown data
Multiagent debate improves factual accuracy.
“Table 2: Multiagent Debate Improves Factual Accuracy Multi-agent debate improves the factual accuracy.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
- Rests on an assumption: Evaluating factual accuracy by checking agreement between generated bullets and ground truth bullet points is a valid measure of factual accuracy.
Findingreasoning
Debate results are less likely to include false facts that models are internally uncertain of.
“Debate results are also less likely to include false facts that models are internally uncertain of.”
Quote in the paperSubstantial claim
Findingown data
Additional debate rounds above four led to a similar final performance to 4 rounds of debate.
“However, we found that additional debate rounds above four led to a similar final performance to 4 rounds of debate.”
Quote in the paperNot a substantial claim
Findingown data
On arithmetic, performance monotonically increases with the increased number of agents.
“On arithmetic, performance monotonically increases with the increased number of agents.”
Quote in the paperNot a substantial claim
Findingreasoning
The multiagent debate process is more computationally expensive than other prompting techniques, requiring multiple model instances and rounds.
“In comparison to other prompting techniques, our multiagent debate procedure is more computationally expensive, as it requires both multiple language generations, and an underlying debate procedure.”
Quote in the paperNot a substantial claim

Findingown data
After debating and examining other model instances' responses, the population almost always converges on a single and more accurate common answer.
“After debating and examining the responses of other model instances, we find that the population almost always converges on a single and more accurate common answer.”
Quote in the paperSubstantial claim
Findingcitation
Directly asking each agent about their confidence of the answer led to high confidence assessments on each answer.
“However, directly asking each agent about their confidence [10] of the answer led to high confidence assessments on each answer.”
Quote in the paperNot a substantial claim0 of 1 sources checked
- cites [10]
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Findingnone given
These facts are often inconsistent across different language model instances.
“Moreover, these facts often inconsistent across different language model instances.”
Quote in the paperNot a substantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
Findingown data
Existing language models had a tendency to hallucinate many facts on the biography task in preliminary testing.
“In preliminary testing, we found that existing language models had a tendency to hallucinate many facts on this task.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
- Rests on an assumption: Evaluating factual accuracy by checking agreement between generated bullets and ground truth bullet points is a valid measure of factual accuracy.
Findingown data
On facts that the language model was confident in, it was very difficult to convince an agent to change their opinion.
“Interestingly, we found that on facts that the language model was confident in (i.e.many instances of the same model all gave the same answer), it was very difficult to convince an agent to change their opinion, suggesting that "ease of persuasion" may be a method to assess factual confidence.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
Findingown data
Different language agents tended to give different answers when the underlying language model was uncertain about the question.
“We found that different language agents tended to give different answers when the underlying language model was uncertain about the question.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
Findingown data
Reflection, where a language model is asked to critique its early generation, generally gives a modest boost in performance.
“Simultaneously, we also see that reflection, where a language model is asked to critique its early generation, generally gives a modest boost in performance.”
Quote in the paperNot a substantial claim
Findingown data
Language model agents are able to converge on a single shared answer after multiple rounds of debate.
“Empirically, however, we find that language models are able to converge on a single shared answer after multiple rounds of debate (Figure 4).”
Quote in the paperSubstantial claim
Findingown data
The duration of debates can be controlled by changing how much a language model trusts its own outputs over those generated by other models through different prompts.
“We found that we could control the duration of debates by how changing how much a language model trusts its own outputs over those generated by other models through different prompts.”
Quote in the paperNot a substantial claim
Findingown data
Approaches based on reflection led to poor performance in the factuality setting.
“We found that approaches based on reflection led to poor performance in the factuality setting.”
Quote in the paperSubstantial claim
Findingown data
Debate does not just act to amplify one correct answer in a model quorum; there are many cases where all models initially make incorrect predictions but then arrive at the correct answer as debate progresses.
“Lastly, we find that debate does not just act to amplify one correct answer in a model quorum -we find many cases where all the models initially make incorrect predictions, but then arrive at the correct answer as debate progresses (Figure 4,11).”
Quote in the paperSubstantial claim
Findingreasoning
Multiagent debate gives a substantial boost in reasoning across each of the tasks.
“Multiagent debate, which may be seen as a combination of both reflection and multiagent generation, gives a substantial boost in reasoning across each of the tasks.”
Quote in the paperSubstantial claim
Findingown data
Multiagent debate significantly enhances mathematical and strategic reasoning across a number of tasks.
“Our findings indicate that this approach significantly enhances mathematical and strategic reasoning across a number of tasks.”
Quote in the paperSubstantial claim
Findingown data
There are cases in which all models initially give an incorrect response, yet the result of debate still obtains the correct answer as agents critique each others' reasoning.
“Interestingly, we find cases in which all models initially give an incorrect response, yet the result of debate still obtains the correct answer as agents critique each others' reasoning.”
Quote in the paperSubstantial claim
Findingown data
On the arithmetic task, performance monotonically increases with debate length.
“We find that on the arithmetic task, the performance also monotonically increases with debate length.”
Quote in the paperNot a substantial claim
Findingown data
Multiagent debate outperforms single model baselines such as zero-shot chain of thought and reflection on a variety of six reasoning, factuality, and question-answering tasks.
“We find that our debate approach outperforms single model baselines such as zero-shot chain of thought [11] and reflection [26,18] on a variety of six reasoning, factuality, and question-answering tasks.”
Quote in the paperSubstantial claim0 of 3 sources checked
- cites [11]
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [26]
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- cites [18]
  Couldn’t be matched to a paper in Crossref, so it can’t be checked.
Findingown data
Contemporary language models have an especially high tendency to hallucinate factually incorrect biographies, often misrepresenting relevant institutions and dates.
“We find that contemporary language models have an especially high tendency to hallucinate factually incorrect biographies, often misrepresenting the relevant institutions and dates.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
- Rests on an assumption: Evaluating factual accuracy by checking agreement between generated bullets and ground truth bullet points is a valid measure of factual accuracy.
Findingreasoning
Despite answers being incorrect, language models would confidently affirm that their answer is correct and consistent with all other agent responses.
“Despite answers being incorrect, language models would confidently affirm that their answer is correct and consistent with all other agent responses.”
Quote in the paperNot a substantial claim
Findingown data
Summarization improves the performance of debate.
“We find this improves the performance of debate, suggesting that summarization is another tool that can further improve multiagent debate.”
Quote in the paperNot a substantial claim
Findingown data
Improved performance on MMLU from 71.1 to 74.2 with different agents.
“We found that improved performance on MMLU from 71.1 to 74.2 with different agents, suggesting further gains can be obtained with different initialization prompts.”
Quote in the paperNot a substantial claim
Findingown data
When different language agents were asked to communicate with each other, each agent would quickly change their opinion to a consensus answer which was more accurate.
“However, when these different language agents were asked to communicate with each other, each agent would quickly change their opinion to a consensus answer which was more accurate.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
Findingown data
Utilizing multiple different agents to generate solutions improves performance over using a single language model agent.
“In each task, we observe that utilizing multiple different agents to generate solutions improves performance over using a single language model agent to generate a solution.”
Quote in the paperSubstantial claim
Findingown data
Individual model instances propose a diverse range of answers despite being the same model class.
“Given an initial query, we find that individual model instances propose a diverse range of answers despite being the same model class”
Quote in the paperSubstantial claim
Findingown data
Performance of debate increases with use of Chain of Thought prompting.
“Performance of debate increases with use of Chain of Thought prompting.”
Quote in the paperNot a substantial claim
Findingown data
Summarizing responses from other agents before giving them as context improves performance in multi-agent debate.
“This operation improves performance.”
Quote in the paperNot a substantial claim
Findingown data
Multi-agent debate improves the reasoning abilities of language models.
“Multi-agent debate improves the reasoning abilities of language models.”
- Also stated in: § Agent 1: · plain
  “Multiagent Debate Improves Reasoning”
Quote in the paperSubstantial claim
Findingnone given
For facts that a language model is uncertain about, different language agents generate different facts.
“For facts that a language model is uncertain about, different language agents generate different facts.”
Quote in the paperSubstantial claim
Findingown data
Generated facts in biographies could contain incorrect information that was not captured in the ground truth bullet and thus could not be validated through the evaluation metric.
“However, we found that generated facts could contain incorrect information that was not captured in the ground truth bullet and thus could not be validated through this metric.”
Quote in the paperSubstantial claim
Findingown data
Multi-agent debate improves factual accuracy.
“Multiagent Debate Improves Reasoning and Factual Accuracy.”
Quote in the paperSubstantial claim
Findingown data
The use of debate prompts that encourage agents to adapt more to the opinions of other agents improves consensus.
“The use of debate prompts that encourage agents to adapt more to the opinions of other agents improves consensus.”
Quote in the paperNot a substantial claim
Interpretationreasoning
By asking models to come to a consensus across their answers, inconsistent facts may be either removed or corrected.
“By asking models to come to a consensus across their answers, such inconsistent facts may be either removed or corrected.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
Interpretationown data
Summarization is another tool that can further improve multiagent debate.
“suggesting that summarization is another tool that can further improve multiagent debate.”
Quote in the paperNot a substantial claim
Interpretationown data
Ease of persuasion may be a method to assess factual confidence.
“suggesting that "ease of persuasion" may be a method to assess factual confidence.”
Quote in the paperNot a substantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
Interpretationreasoning
Language models do not correctly express their uncertainty when generating responses.
“LMs do not correctly express their uncertainty when generating responses”
Quote in the paperSubstantial claim
Interpretationreasoning
The method can be used with common public models serving interfaces.
“This allows our method to be used with common public models serving interfaces.”
Quote in the paperNot a substantial claim
Interpretationreasoning
The resulting quorum of models can hold and maintain multiple chains of reasoning and possible answers simultaneously before proposing the final answer.
“The resulting quorum of models can hold and maintain multiple chains of reasoning and possible answers simultaneously before proposing the final answer.”
Quote in the paperSubstantial claim
Interpretationreasoning
The debate process induces models to construct answers that are consistent with both their internal critic and sensible in light of the responses of other agents.
“This process induces models to construct answers that are consistent with both their internal critic as well as sensible in light of the responses of other agents.”
Quote in the paperNot a substantial claim

May combine two claims. 23
Classified as probably not a substantial claim. 9

May combine two claims.
Findingown data
In a set of 20 GSM8K math problems, multi-agent debate improves the performance of both chatGPT and Bard agents.
“In this set, we find that multi-agent debate improves the performance of both agents, with Bard solving 11 problems, chatGPT solving 14 problems, and joint multi-agent debate solving 17 problems.”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
Using both multiple model agents and multiple rounds of debate are important to achieve the best performance.
“Using both multiple model agents and multiple rounds of debate are important to achieve the best performance.”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
As debate progresses, individual model instances tend to disagree on uncertain facts and omit them from the answer.
“This is because as the debate progresses, individual model instances tend to disagree on uncertain facts and omit them from the answer (Figure 7).”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
Multiagent generation is beneficial with and without zero-shot chain-of-thought prompting on GSM8K.
“In both settings, multiagent generation is beneficial.”
Quote in the paperNot a substantial claim
May combine two claims.
Findingnone given
The multiagent debate procedure is simple and effective across a wide set of different reasoning and validity language modeling tasks.
“We find that the approach is simple and effective across a wide set of different reasoning and validity language modeling tasks.”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
Multiagent debate improved and settled on bullets that were more consistent across agents.
“We found that multiagent debate improved and settled on bullets that were more consistent across agents.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
May combine two claims.
Findingown data
Further gains are obtained with both more agents and rounds of debate.
“although we found further gains with both more agents and rounds of debate (Figure 10).”
Quote in the paperNot a substantial claim
May combine two claims.
Findingown data
While debates typically converged into single final answers, these answers were not necessarily correct.
“Finally, we found that while debates typically converged into single final answers, these answers were not necessarily correct.”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingown data
While both agents initially provide incorrect answers to the problem, chatGPT is able to utilize the incorrect response given by Bard to generate the final correct answer.
“While both agents initially provide incorrect answers to the problem, chatGPT is able to utilize the incorrect response given by Bard to generate the final correct answer.”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingcitation
Language model agents were relatively agreeable.
“Overall, we observed that language model agents were relatively "agreeable"”
Quote in the paperNot a substantial claim
May combine two claims.
Findingown data
Prompts that encouraged models to be more stubborn based on their own solutions led to longer debates and better final solutions.
“In general, we found that prompts that encouraged models to be more "stubborn' based on their own solutions led to longer debates and better final solutions.”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
Multiagent debate improves the factual validity of generated content, reducing fallacious answers and hallucinations.
“We also demonstrate that our approach improves the factual validity of generated content, reducing fallacious answers and hallucinations that contemporary models are prone to.”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
As debates became longer in duration, current language models sometimes struggled to fully process the entire debate input, and typically only focused on the most recent generations.
“Further, we observed that as debates became longer in duration, current language models sometimes struggled to fully process the entire debate input, and typically only focused on the most recent generations.”
Quote in the paperNot a substantial claim
May combine two claims.
Findingown data
Debates using longer prompts lead to slower convergence to correct answers, but also lead to a better final consensus on the correct answer.
“We find that debates using longer prompts lead to slower convergence to correct answers, but also lead to a better final consensus on the correct answer.”
Quote in the paperNot a substantial claim
May combine two claims.
Findingown data
Debate gives the best performance in the factuality setting and significantly outperforms each baseline.
“In contrast, debate gives the best performance in this setting also, and significantly outperforms each baseline.”
Quote in the paperSubstantial claim
- Rests on an assumption: The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
- Rests on an assumption: Evaluating factual accuracy by checking agreement between generated bullets and ground truth bullet points is a valid measure of factual accuracy.
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given
When there are many agents in a debate, summarizing responses from other agents before giving them as context reduces context length.
“When there are many agents in a debate, responses from other agents may be first summarized and then given as context, reducing context length.”
Quote in the paperNot a substantial claim
May combine two claims.
Findingnone given
Debate causes agents to converge to one fact that is more accurate, but not necessarily always factually correct.
“Debate causes agents to converge to one fact that is more accurate, but not necessarily always factually correct.”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
On the Arithmetic benchmark, multi-agent debate achieves 81.8% ± 2.3 accuracy, compared to 67.0 ± 4.7 for single agent, 72.1 ± 4.5 for single agent reflection, and 69.0 ± 4.6 for multi-agent majority.
“| Model | Arithmetic (%) ↑ Grade School Math (%) ↑ Chess (∆PS) ↑ | | | | --- | --- | --- | --- | | Single Agent | 67.0 ± 4.7 | 77.0 ± 4.2 | 91.4 ± 10.6 | | Single Agent (Reflection) | 72.1 ± 4.5 | 75.0 ± 4.3 | 102.1 ± 11.9 | | Multi-Agent (Majority) | 69.0 ± 4.6 | 81.0 ± 3.9 | 102.2 ± 6.2 | | Multi-Agent (Debate) | 81.8 ± 2.3 | 85.0 ± 3.5 | 122.9 ± 7.6 |”
Quote in the paperSubstantial claim
May combine two claims.
Findingnone given
Generated biographies after debate are more accurate, but many facts remain incorrect.
“While we found that generated biographies after debate to be more accurate, many facts remain incorrect.”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
On the Grade School Math benchmark, multi-agent debate achieves 85.0% ± 3.5 accuracy, compared to 77.0 ± 4.2 for single agent, 75.0 ± 4.3 for single agent reflection, and 81.0 ± 3.9 for multi-agent majority.
“| Model | Arithmetic (%) ↑ Grade School Math (%) ↑ Chess (∆PS) ↑ | | | | --- | --- | --- | --- | | Single Agent | 67.0 ± 4.7 | 77.0 ± 4.2 | 91.4 ± 10.6 | | Single Agent (Reflection) | 72.1 ± 4.5 | 75.0 ± 4.3 | 102.1 ± 11.9 | | Multi-Agent (Majority) | 69.0 ± 4.6 | 81.0 ± 3.9 | 102.2 ± 6.2 | | Multi-Agent (Debate) | 81.8 ± 2.3 | 85.0 ± 3.5 | 122.9 ± 7.6 |”
Quote in the paperSubstantial claim
May combine two claims.
Findingown data
In general, debate improved the performance of final generated answers, though sometimes answers would converge to the incorrect value.
“In general, we found that debate improved the performance of final generated answers, though sometimes answers would converge to the incorrect value.”
Quote in the paperSubstantial claim
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingnone given
Using a short debate prompt induces faster consensus between agents.
“Using a short debate prompt induces faster consensus between agents”
- Also stated in: § Agent 1: · plain
  “Using a short debate prompt induces faster consensus between agents”
Quote in the paperNot a substantial claim
May combine two claims.
Findingown data
On the Chess benchmark, multi-agent debate achieves a normalized score of 122.9 ± 7.6, compared to 91.4 ± 10.6 for single agent, 102.1 ± 11.9 for single agent reflection, and 102.2 ± 6.2 for multi-agent majority.
“| Model | Arithmetic (%) ↑ Grade School Math (%) ↑ Chess (∆PS) ↑ | | | | --- | --- | --- | --- | | Single Agent | 67.0 ± 4.7 | 77.0 ± 4.2 | 91.4 ± 10.6 | | Single Agent (Reflection) | 72.1 ± 4.5 | 75.0 ± 4.3 | 102.1 ± 11.9 | | Multi-Agent (Majority) | 69.0 ± 4.6 | 81.0 ± 3.9 | 102.2 ± 6.2 | | Multi-Agent (Debate) | 81.8 ± 2.3 | 85.0 ± 3.5 | 122.9 ± 7.6 |”
Quote in the paperSubstantial claim
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Interpretationnone given
The method is orthogonal to other model generation improvements such as retrieval or prompt engineering, and can be combined with zero-shot chain of thought.
“The method is also orthogonal to other model generation improvements such as retrieval or prompt engineering (in fact, we combine our debate method with zero-shot chain of thought).”
Quote in the paperNot a substantial claim
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductown data
The main evaluation used three agents with two rounds of debate due to computational expense.
“Due to computational expense, we evaluate our approach across benchmarks mainly using three agents with two rounds of debates”
Quote in the paperNot a substantial claim
Classified as probably not a substantial claim.
Study design and conductown data
The experiments used the identical starting prompt and language model across all evaluations.
“As the focus of our experiments is to verify the effectiveness of multiagent agent debate, we run both baselines and our approach, using the identical starting prompt and language model across all evaluations.”
Quote in the paperSubstantial claim
May combine two claims.
Interpretationown data
The purpose of the debate is not just to amplify a correct answer; all models can initially be wrong but arrive at the correct answer through the debate process.
“Thus, the purpose of our debate isn't just to amplify a correct answer -all models can initially be wrong but arrive at the correct answer through the debate process.”
Quote in the paperSubstantial claim
Probably not a substantial claim: may not pass S3 (load-bearing).
Interpretationown data
Further gains can be obtained with different initialization prompts.
“suggesting further gains can be obtained with different initialization prompts.”
Quote in the paperNot a substantial claim
May combine two claims.
Study design and conductreasoning
The multiagent debate approach uses the same methodology and prompt templates for all tasks and requires only black-box access to language model generations, so it may be directly applied to existing black-box models.
“Our approach may be directly applied to existing black-box models and uses identical procedure and prompts for all tasks we investigate.”
- Also stated in: § 1 Introduction · plain
  “We use the same methodology and prompt templates for all our tasks and require only black-box access to language model generations -no model-internal information such as likelihoods or gradients is needed.”
Quote in the paperSubstantial claim
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone given
The experiments use the same prompts for all agents.
“In our experiments we use the same prompts for all agents.”
Quote in the paperNot a substantial claim

30 of 30

The paper relies on this without support.
The constructed ground truth bullet point biographies of 524 well-known computer scientists are accurate.
“We constructed ground truth bullet point biographies of 524 well-known computer scientists.”
The paper relies on this without support.
The chatGPT-based language model used in the experiments is representative of contemporary language models.
“We use chatGPT-based language model [21] in all our experiments except those in Figure 11 where we compare multiple language models.”
The paper relies on this without support.
Evaluating factual accuracy by checking agreement between generated bullets and ground truth bullet points is a valid measure of factual accuracy.
“We then asked language models to generate bullet point biographies for each person, and evaluated the accuracy at which each ground truth bullet point agreed with generated bullets.”

3 of 3

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

In the crowd-sourcing debating mechanism, the capability of each agent is enhanced by understanding and incorporating the other agents' opinions.
“In this method, the capability of each agent is enhanced by understanding and incorporating the other agents' opinions.”
From A survey on large language model based autonomous agents (2024) · cites this paper as [91]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

Arithmetic example does not use six two-digit numbers
Minor
The task is described as using six two-digit numbers, but its example includes the one-digit numbers 0 and 3.
Triangle area is calculated incorrectly
Minor
The stated multiplication gives 6, not 64.
Reported 64 · recomputed 6
Prompts are not identical across tasks
Minor
The paper says it uses identical prompts for all tasks, but the appendix gives different starting prompts for arithmetic and GSM8K.

3 of 3

Chain-Of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q. V., & Zhou, D. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1800 · Cited by 4,505 · cited 1× · see citing works
“To improve reasoning, approaches have relied on prompting techniques such as scratchpads [20], verification [3], chain-of-thought demonstrations [30, 11, 25], and intermediate self-reflection [26, 18] and finetuning [13, 24, 31].”
Method
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022) · arXiv · doi:10.48550/arxiv.2201.11903 · Cited by 4,263 · see citing works
Training language models to follow instructions with human feedback
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022) · arXiv · doi:10.48550/arxiv.2203.02155 · Cited by 4,208 · see citing works
Training Language Models to Follow Instructions with Human Feedback
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-2011 · Cited by 3,117 · cited 1× · see citing works
“Overall, we observed that language model agents were relatively "agreeable", perhaps as a result of instruction tuning or reinforcement learning based on human feedback (Ouyang et al., 2022).”
The Society of Mind
Minsky, M. (1987) · Personalist Forum · doi:10.5840/persforum19873112 · Cited by 2,415 · cited 2× · see citing works
“Instead, we propose a complementary approach inspired by The Society of Mind (Minsky, 1988) and multiagent settings, where multiple language model instances (or agents) individually propose and jointly debate their responses and reasoning processes to arrive at a common answer.”
Influential
Large language models are zero-shot reasoners
Kojima, T., Gu, S. (., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1613 · Cited by 2,013 · cited 3× · see citing works
“In Figure 6, we illustrate the performance of multi-agent debate with and without zero-shot chain-ofthought prompting [11] on GSM8K.”
BackgroundMethod
Large Language Models are Zero-Shot Reasoners
Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022) · arXiv · doi:10.48550/arxiv.2205.11916 · Cited by 2,013 · see citing works
Flamingo: a Visual Language Model for Few-Shot Learning
Alayrac, J.-B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al. (2022) · arXiv · doi:10.48550/arxiv.2204.14198 · Cited by 1,555 · see citing works
Reflexion: Language Agents with Verbal Reinforcement Learning
Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., & Yao, S. (2023) · arXiv · doi:10.48550/arxiv.2303.11366 · Cited by 1,087 · see citing works
Deep reinforcement learning from human preferences
Christiano, P., Leike, J., Brown, T. B., Martic, M., Legg, S., & Amodei, D. (2017) · arXiv · doi:10.48550/arxiv.1706.03741 · Cited by 1,058 · cited 7× · see citing works
“State the destination square in the form (X), where X follows the regex [a-h][1-8], for example (e5).”
Influential
Competition-level code generation with AlphaCode
Li, Y., Choi, D., Chung, J., Kushman, N., Schrittwieser, J., Leblond, R., Eccles, T., Keeling, J., Gimeno, F., Dal Lago, A., et al. (2022) · Science · doi:10.1126/science.abq1158 · Cited by 1,043 · see citing works
Flamingo: a visual language model for few-shot learning
Alayrac, J.-B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-1723 · Cited by 966 · cited 1× · see citing works
“Our work is also related to existing works that focus on text generation by combining different models (Du et al., 2020; Liu et al., 2022b; Zeng et al., 2022; Alayrac et al., 2022; Du et al., 2023).”
Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm
Reynolds, L., & McDonell, K. (2021) · Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems · doi:10.1145/3411763.3451760 · Cited by 952 · cited 2× · see citing works
“…such as scratchpads (Nye et al., 2021), verification (Cobbe et al., 2021), chain-of-thought demonstrations (Wei et al., 2022; Kojima et al., 2022; Reynolds & McDonell, 2021), and intermediate self-reflection (Shinn et al., 2023; Madaan et al., 2023) and finetuning (Lewkowycz et al., 2022; Rajani…”
Self-Consistency Improves Chain of Thought Reasoning in Language Models
Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., & Zhou, D. (2022) · arXiv · doi:10.48550/arxiv.2203.11171 · Cited by 829 · cited 1× · see citing works
“Prior work also has explored how to take the majority vote across different models [15, 3, 29, 28] while in this work, we use the power of a language model to combine different answers.”
Method
LaMDA: Language Models for Dialog Applications
Thoppilan, R., De Freitas, D., Hall, J., Shazeer, N., Kulshreshtha, A., Cheng, H.-T., Jin, A., Bos, T., Baker, L., Du, Y., et al. (2022) · arXiv · doi:10.48550/arxiv.2201.08239 · Cited by 778 · cited 1× · see citing works
“Prior work also has explored how to take the majority vote across different models (Li et al., 2022b; Cobbe et al., 2021; Wang et al., 2022; Thoppilan et al., 2022; Lewkowycz et al., 2022) while in this work, we use rounds of debate between language model to combine answers.”
Self-citation
Training Verifiers to Solve Math Word Problems
Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., et al. (2021) · arXiv · doi:10.48550/arxiv.2110.14168 · Cited by 575 · cited 3× · see citing works
“Using the GSM8K dataset (Cobbe et al., 2021), the models must correctly solve grade school mathematical reasoning tasks.”
REALM: Retrieval-Augmented Language Model Pre-Training
Guu, K., Lee, K., Tung, Z., Pasupat, P., & Chang, M.-W. (2020) · arXiv · doi:10.48550/arxiv.2002.08909 · Cited by 574 · cited 3× · see citing works
“State your final answer in a newline with a 2 letter response following the regex [a-h][1-8].”
BackgroundMethod
Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models
Srivastava, A., Rastogi, A., Rao, A., Shoeb, A. A. M., Abid, A., Fisch, A., Brown, A. R., Santoro, A., Gupta, A., Garriga-Alonso, A., et al. (2022) · arXiv · doi:10.48550/arxiv.2206.04615 · Cited by 569 · cited 2× · see citing works
“To evaluate chess validity, we consider the BIG-Bench Chess-State Tracking Benchmark (Srivastava et al., 2022), where we used the hardest reported task in the benchmark synthetic_short .”
Self-citation
Explain Yourself! Leveraging Language Models for Commonsense Reasoning
Rajani, N. F., McCann, B., Xiong, C., & Socher, R. (2019) · Proceedings of the conference - Association for Computational Linguistics. Meeting · doi:10.18653/v1/p19-1487 · Cited by 421 · cited 1× · see citing works
“To improve reasoning, approaches have relied on prompting techniques such as scratchpads [20], verification [3], chain-of-thought demonstrations [30, 11, 25], and intermediate self-reflection [26, 18] and finetuning [13, 24, 31].”
Method
Fine-Tuning Language Models from Human Preferences
Ziegler, D. M., Stiennon, N., Wu, J., Brown, T. B., Radford, A., Amodei, D., Christiano, P., & Irving, G. (2019) · arXiv · doi:10.48550/arxiv.1909.08593 · Cited by 413 · cited 4× · see citing works
“Our results indicate that RLHF significantly improves consensus in multiagent debate.”
Influential
Solving Quantitative Reasoning Problems with Language Models
Lewkowycz, A., Andreassen, A., Dohan, D., Dyer, E., Michalewski, H., Ramasesh, V., Slone, A., Anil, C., Schlag, I., Gutman-Solo, T., et al. (2022) · arXiv · doi:10.48550/arxiv.2206.14858 · Cited by 358 · cited 3× · see citing works
“If the butter costs $4.00 a pound, how much will it cost him to purchase 6 pounds of butter? sider generating responses from multiple instances of model and performing majority voting across responses (Wang et al., 2022; Lewkowycz et al., 2022).”
Influential
Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Shi, S., & Tu, Z. (2023) · arXiv · doi:10.48550/arxiv.2305.19118 · Cited by 320 · cited 4× · see citing works
“In Table A6, we find that our approach substantially outperforms Liang et al. (2023), as it more effectively encourages language models to focus on reasoning.”
Influential
Camel: Communicative agents for" mind" exploration of large language model society
Li, G., Hammoud, H., Itani, H., Khizbullin, D., & Ghanem, B. (2023) · Advances in Neural Information Processing Systems 36 · doi:10.52202/075280-2264 · Cited by 314 · cited 2× · see citing works
“Largely concurrent to the work in this paper, a body of recent work has explored how multiagent interaction can be used with large language models (Zhang et al., 2023; Li et al., 2024; Hong et al., 2023).”
Measuring Mathematical Problem Solving With the MATH Dataset
Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., & Steinhardt, J. (2021) · arXiv · doi:10.48550/arxiv.2103.03874 · Cited by 306 · cited 1× · see citing works
“Our approach also improves the performance of a GPT-4 model on the MMLU and MATH dataset (Hendrycks et al., 2021).”
Compositional Visual Generation with Composable Diffusion Models
Liu, N., Li, S., Du, Y., Torralba, A., & Tenenbaum, J. B. (2022) · arXiv · doi:10.48550/arxiv.2206.01714 · Cited by 295 · cited 1× · see citing works
“Our work is also related to existing works that focus on text generation by combining different models [4, 17, 32, 1, 5].”
BackgroundSelf-citation
Measuring Massive Multitask Language Understanding
Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., & Steinhardt, J. (2020) · arXiv · doi:10.48550/arxiv.2009.03300 · Cited by 291 · cited 3× · see citing works
“We utilize the existing MMLU dataset [8] to benchmark the accuracy of responses.”
BackgroundMethod
Reduce, Reuse, Recycle: Compositional Generation with Energy-Based Diffusion Models and MCMCPreprint
Du, Y., Durkan, C., Strudel, R., et al. (2023) · International Conference on Machine Learning · doi:10.48550/arxiv.2302.11552 · Cited by 274 · cited 1× · see citing works
“Our work is also related to existing works that focus on text generation by combining different models (Du et al., 2020; Liu et al., 2022b; Zeng et al., 2022; Alayrac et al., 2022; Du et al., 2023).”
Self-citation
Self-Refine: Iterative Refinement with Self-Feedback
Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., et al. (2023) · arXiv · doi:10.48550/arxiv.2303.17651 · Cited by 219 · cited 2× · see citing works
“We find that our debate approach outperforms single model baselines such as zero-shot chain of thought (Kojima et al., 2022) and reflection (Shinn et al., 2023; Madaan et al., 2023) on a variety of six reasoning, factuality, and question-answering tasks.”
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
Hong, S., Zhuge, M., Chen, J., Zheng, X., Cheng, Y., Zhang, C., Wang, J., Wang, Z., Yau, S. K. S., Lin, Z., et al. (2023) · arXiv · doi:10.48550/arxiv.2308.00352 · Cited by 214 · cited 2× · see citing works
“Largely concurrent to the work in this paper, a body of recent work has explored how multiagent interaction can be used with large language models (Zhang et al., 2023; Li et al., 2024; Hong et al., 2023).”
Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language
Zeng, A., Attarian, M., Ichter, B., Choromanski, K., Wong, A., Welker, S., Tombari, F., Purohit, A., Ryoo, M., Sindhwani, V., et al. (2022) · arXiv · doi:10.48550/arxiv.2204.00598 · Cited by 205 · cited 2× · see citing works
“Most similar to our work, (Li et al., 2022a; Zeng et al., 2022) pro-pose to combine multiple different large pretrained models together for multimodal reasoning.”
Language Models (Mostly) Know What They Know
Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., et al. (2022) · arXiv · doi:10.48550/arxiv.2207.05221 · Cited by 187 · cited 2× · see citing works
“…have relied on training techniques such as RLHF (Ziegler et al., 2019; Liu et al., 2022a; Christiano et al., 2017), pruning truthful datasets (Lee et al., 2022), external knowledge retrieval (Guu et al., 2020) and training-free methods based off likelihood estimation (Kadavath et al., 2022).”
STaR: Bootstrapping Reasoning With Reasoning
Zelikman, E., Wu, Y., Mu, J., & Goodman, N. D. (2022) · arXiv · doi:10.48550/arxiv.2203.14465 · Cited by 182 · cited 1× · see citing works
“To improve reasoning, approaches have relied on prompting techniques such as scratchpads [20], verification [3], chain-of-thought demonstrations [30, 11, 25], and intermediate self-reflection [26, 18] and finetuning [13, 24, 31].”
Method
Untitled work
(2022) · doi:10.52202/068431-1126 · Cited by 182 · see citing works
Show Your Work: Scratchpads for Intermediate Computation with Language Models
Nye, M., Andreassen, A. J., Gur-Ari, G., Michalewski, H., Austin, J., Bieber, D., Dohan, D., Lewkowycz, A., Bosma, M., Luan, D., et al. (2021) · arXiv · doi:10.48550/arxiv.2112.00114 · Cited by 144 · cited 1× · see citing works
“To improve reasoning, approaches have relied on techniques such as scratchpads (Nye et al., 2021), verification (Cobbe et al., 2021), chain-of-thought demonstrations (Wei et al., 2022; Kojima et al., 2022; Reynolds & McDonell, 2021), and intermediate self-reflection (Shinn et al., 2023; Madaan et…”
Scalable Multi-Robot Collaboration with Large Language Models: Centralized or Decentralized Systems?
Chen, Y., Arkin, J., Zhang, Y., Roy, N., & Fan, C. (2024) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra57147.2024.10610676 · Cited by 111 · cited 2× · see citing works
“Largely concurrent to the work in this paper, a body of recent work has explored how multiagent interaction can be used with large language models (Zhang et al., 2023; Li et al., 2024; Hong et al., 2023).”
AI safety via debate
Irving, G., Christiano, P., & Amodei, D. (2018) · arXiv · doi:10.48550/arxiv.1805.00899 · Cited by 71 · cited 2× · see citing works
“Most similar to our work, (Irving et al., 2018) also proposes a debate procedure to verify the accuracy and safety of powerful AI agents.”
Factuality Enhanced Language Models for Open-Ended Text Generation
Lee, N., Ping, W., Xu, P., Patwary, M., Fung, P., Shoeybi, M., & Catanzaro, B. (2022) · arXiv · doi:10.48550/arxiv.2206.04624 · Cited by 68 · cited 1× · see citing works
“To improve factuality, approaches have relied on training techniques such as RLHF [33, 16, 2], pruning truthful datasets [12], external knowledge retrieval [7] and training-free methods based off likelihood estimation [10].”
Method
Untitled work
(2022) · doi:10.52202/068431-2506 · Cited by 68 · see citing works
ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate
Chan, C.-M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S., Fu, J., & Liu, Z. (2023) · arXiv · doi:10.48550/arxiv.2308.07201 · Cited by 55 · cited 3× · see citing works
“In contrast, our work and work in (Liang et al., 2023; Chan et al., 2023) focus on using debate at inference time to improve the performance of language models.”
Building Cooperative Embodied Agents Modularly with Large Language Models
Zhang, H., Du, W., Shan, J., Zhou, Q., Du, Y., Tenenbaum, J. B., Shu, T., & Gan, C. (2023) · arXiv · doi:10.48550/arxiv.2307.02485 · Cited by 52 · see citing works
Self-citation
Two-Turn Debate Doesn't Help Humans Answer Hard Reading Comprehension QuestionsPreprint
Parrish, A., Trivedi, H., Nangia, N., et al. (2022) · arXiv.org · doi:10.48550/arxiv.2210.10860 · Cited by 18 · cited 3× · see citing works
“In Parrish et al. (2022) and Michael et al. (2023), debate between two unreliable human experts is used to facilitate a less-informed human judge to answer questions about a passage.”
Instruction-Following Agents with Multimodal TransformerPreprint
Liu, H., Lee, L., Lee, K., & Abbeel, P. (2022) · arXiv.org · doi:10.48550/arxiv.2210.13431 · Cited by 17 · cited 4× · see citing works
“Our results indicate that RLHF significantly improves consensus in multiagent debate.”
Influential
Debating with More Persuasive LLMs Leads to More Truthful Answers
Khan, A., Hughes, J., Valentine, D., Ruis, L., Sachan, K., Radhakrishnan, A., Grefenstette, E., Bowman, S. R., Rocktäschel, T., & Perez, E. (2024) · arXiv · doi:10.48550/arxiv.2402.06782 · Cited by 13 · cited 1× · see citing works
“Most similar to our work, recent works have explored how debate can be applied to language models (Parrish et al., 2022; Michael et al., 2023; Khan et al., 2024).”
Composing Ensembles of Pre-trained Models via Iterative ConsensusPreprint
Li, S., Du, Y., Tenenbaum, J. B., Torralba, A. B., & Mordatch, I. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2210.11522 · Cited by 7 · cited 1× · see citing works
“Most similar to our work, [14, 32] propose to combine multiple different large pretrained models together for multimodal reasoning.”
BackgroundSelf-citation
Et al
BMJ Publishing Group Ltd, & Royal College of Psychiatrists and British Psychological Society (2008) · Evidence-Based Mental Health · doi:10.1136/ebmh.11.4.102 · cited 2×
“Prior work also has explored how to take the majority vote across different models [15, 3, 29, 28] while in this work, we use the power of a language model to combine different answers.”

Flagged first, then by citations

3 of the 48 references couldn’t be matched to a known paper.

Research works printed but not linked: 1

Y. Du, S. Li, and I. Mordatch. Compositional visual generation with energy based models. In Advances in Neural Information Processing Systems, 2020. 9

Sorted by citations

A survey on large language model based autonomous agents
Wang, L., Ma, C., Feng, X., Zhang, Z., Yang, H., Zhang, J., Chen, Z., Tang, J., Chen, X., Lin, Y., et al. (2024) · Frontiers of Computer Science · doi:10.1007/s11704-024-40231-1 · Cited by 1,941 · see citing works
The rise and potential of large language model based agents: a survey
Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al. (2025) · Science China Information Sciences · doi:10.1007/s11432-024-4222-0 · Cited by 970 · see citing works
“For example, Du et al. [92] introduced the concept of debate, endowing agents with responses from fellow peers.”
Influential
ChatGPT: The End of Online Exam Integrity?
Susnjak, T. (2022) · arXiv · doi:10.48550/arxiv.2212.09292 · Cited by 370 · see citing works
“Even if the possibility of LLM hallucination is deemed to be likely in certain assessments, recent studies demonstrate that with careful strategies and prompting techniques, hallucination can to a large degree be mitigated, though not entirely eliminated [22,50–53].”
Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Shi, S., & Tu, Z. (2024) · Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2024.emnlp-main.992 · Cited by 274 · see citing works
The Rise and Potential of Large Language Model Based Agents: A Survey
Xi, Z., Chen, W., Guo, X., He, W., Ding, Y., Hong, B., Zhang, M., Wang, J., Jin, S., Zhou, E., et al. (2023) · arXiv · doi:10.48550/arxiv.2309.07864 · Cited by 972 · see citing works
A review of large language models and autonomous agents in chemistry
Ramos, M. C., Collison, C. J., & White, A. D. (2025) · Chemical Science · doi:10.1039/d4sc03921a · Cited by 274 · see citing works
A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., et al. (2023) · arXiv · doi:10.48550/arxiv.2311.05232 · Cited by 2,516 · see citing works
“This feedback typically comes from external knowledge sources through retrieval processes (He et al., 2023; Gou et al., 2023), interactive debates with other LLMs (Du et al., 2023; Cohen et al., 2023), or guidance from external evaluation metrics (Lei et al., 2023; Khalifa et al., 2023).”
MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework
Hong, S., Zhuge, M., Chen, J., Zheng, X., Cheng, Y., Zhang, C., Wang, J., Wang, Z., Yau, S. K. S., Lin, Z., et al. (2023) · arXiv · doi:10.48550/arxiv.2308.00352 · Cited by 214 · see citing works
RoCo: Dialectic Multi-Robot Collaboration with Large Language Models
Mandi, Z., Jain, S., & Song, S. (2024) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra57147.2024.10610855 · Cited by 206 · see citing works
“Recent work also shows a dialog or debate style prompting can improve LLMs’ performance on human alignment [16] and a broad range of goal-oriented tasks [31], [25], [12].”
Background
From text to insight: large language models for chemical data extraction
Schilling-Wilhelmi, M., Ríos-García, M., Shabih, S., Gil, M. V., Miret, S., Koch, C. T., Marquez, J. A., & Jablonka, K. M. (2024) · Chemical Society Reviews · doi:10.1039/d4cs00913d · Cited by 155 · see citing works
Towards trustworthy LLMs: a review on debiasing and dehallucinating in large language models
Lin, Z., Guan, S., Zhang, W., Zhang, H., Li, Y., & Zhang, H. (2024) · Artificial Intelligence Review · doi:10.1007/s10462-024-10896-y · Cited by 141 · see citing works
OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation
Huang, Q., Dong, X., Zhang, P., Wang, B., He, C., Wang, J., Lin, D., Zhang, W., & Yu, N. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.01274 · Cited by 171 · see citing works
“Researches on mitigating current LLMs’ hallucination issues often focuses on several aspects, including refining the training process, us-ing larger and more diverse datasets [24], or implementing post-training evaluation [11] and correction mechanisms [33, 34].”
Large Generative AI Models for Telecom: The Next Big Thing?
Bariah, L., Zhao, Q., Zou, H., Tian, Y. T., Bader, F., & Debbah, M. A. (2024) · IEEE Communications Magazine · doi:10.1109/mcom.001.2300364 · Cited by 142 · see citing works
LILAC: Log Parsing using LLMs with Adaptive Parsing Cache
Jiang, Z., Liu, J., Chen, Z., Li, Y., Huang, J., Huo, Y., He, P., Gu, J., & Lyu, M. R.-T. (2024) · Proceedings of the ACM on software engineering. · doi:10.1145/3643733 · Cited by 148 · see citing works
“Therefore, compared to traditional parsing tools, the overhead of querying LLMs ( e.g. , inference time and network latency) is notably high [10, 58].”
Generative AI in EU law: Liability, privacy, intellectual property, and cybersecurity
Novelli, C., Casolari, F., Hacker, P., Spedicato, G., & Floridi, L. (2024) · Computer Law & Security Review · doi:10.1016/j.clsr.2024.106066 · Cited by 124 · see citing works
How large language models can reshape collective intelligence
Burton, J. W., Lopez-Lopez, E., Hechtlinger, S., Rahwan, Z., Aeschbach, S., Bakker, M. A., Becker, J. A., Berditchevskaia, A., Berger, J., Brinkmann, L. U. L., et al. (2024) · Nature Human Behaviour · doi:10.1038/s41562-024-01959-9 · Cited by 124 · see citing works
“For example, multi-agent systems in which multiple LLMs are allowed to converse with one another display enhanced mathematical and strategic reasoning capabilities 56 and can complete tasks more efficiently than any single LLM 57 .”
Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback
Casper, S., Davies, X., Shi, C., Gilbert, T. K., Scheurer, J., Rando, J., Freedman, R., Korbak, T., Lindner, D., Freire, P., et al. (2023) · arXiv · doi:10.48550/arxiv.2307.15217 · Cited by 101 · see citing works
“…et al., 2022b, Saunders et al., 2022, Ye et al., 2023, Kim et al., 2023, Akyürek et al., 2023, Madaan et al., 2023, Chen et al., 2023, Gilardi et al., 2023] with research agendas like Recursive Reward Modeling [Leike et al., 2018] and AI Safety via debate [Irving et al., 2018, Du et al., 2023].”
Generative AI in Marketing and Principles for Ethical Design and Deployment
Hermann, E., & Puntoni, S. (2024) · Journal of Public Policy & Marketing · doi:10.1177/07439156241309874 · Cited by 93 · see citing works
“…model instances generating, critiquing, and debating their own output to collaboratively reach a more accurate consensus response) that facilitate and foster collaboration between systems, content cross-checking, and eventually collective and complementary intelligences (Du et al. 2023).”
Designing Heterogeneous LLM Agents for Financial Sentiment Analysis
Xing, F. (2025) · ACM Transactions on Management Information Systems · doi:10.1145/3688399 · Cited by 88 · see citing works
Scalable Multi-Robot Collaboration with Large Language Models: Centralized or Decentralized Systems?
Chen, Y., Arkin, J., Zhang, Y., Roy, N., & Fan, C. (2024) · Proceedings - IEEE International Conference on Robotics and Automation/Proceedings · doi:10.1109/icra57147.2024.10610676 · Cited by 111 · see citing works
“Recent work shows that discussion among multiple LLM agents can improve factuality and accuracy[31], [32].”
Background
When Can LLMs <i>Actually</i> Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs
Kamoi, R., Zhang, Y., Zhang, N., Han, J., & Zhang, R. (2024) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00713 · Cited by 82 · see citing works
Automatically Correcting Large Language Models: <i>Surveying the Landscape of Diverse Automated Correction Strategies</i>
Pan, L., Saxon, M., Xu, W., Nathani, D., Wang, X., & Wang, W. Y. (2024) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00660 · Cited by 86 · see citing works
Security and Privacy on Generative Data in AIGC: A Survey
Wang, T., Zhang, Y., Qi, S., Zhao, R., Xia, Z., & Weng, J. (2024) · ACM Computing Surveys · doi:10.1145/3703626 · Cited by 107 · see citing works
“[123] prompted multiple language models to debate their viewpoints and reasoning processes over multiple rounds, and finally come up with a unitive”
Background
Editable Scene Simulation for Autonomous Driving via Collaborative LLM-Agents
Wei, Y., Wang, Z., Lu, Y., Xu, C., Liu, C., Zhao, H., Chen, S., & Wang, Y. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.01428 · Cited by 81 · see citing works
“Based on LLM, many works [3, 23, 29, 65, 76] improve the problem-solving abilities by integrating communication among multiple agents.”
Background
Cognitive Architectures for Language Agents
Sumers, T. R., Yao, S., Narasimhan, K., & Griffiths, T. L. (2023) · arXiv · doi:10.48550/arxiv.2309.02427 · Cited by 72 · see citing works
“…multiple language agents for social simulation (Park et al., 2023; Jinxin et al., 2023; Gao et al., 2023), debate (Chan et al., 2023; Liang et al., 2023b; Du et al., 2023), improved safety (Irving et al., 2018), or collabrative task solving (Qian et al., 2023; Wu et al., 2023; Hong et al., 2023).”
Background
Building machines that learn and think with people
Collins, K. M., Sucholutsky, I., Bhatt, U., Chandra, K., Wong, L., Lee, M., Zhang, C. E., Zhi‐Xuan, T., Ho, M. K., Mansinghka, V. K., et al. (2024) · Nature Human Behaviour · doi:10.1038/s41562-024-01991-9 · Cited by 93 · see citing works
Self-citation
Stance Detection with Collaborative Role-Infused LLM-Based Agents
Lan, X., Gao, C., Jin, D., & Li, Y. (2024) · Proceedings of the International AAAI Conference on Web and Social Media · doi:10.1609/icwsm.v18i1.31360 · Cited by 69 · see citing works
Cognitive Mirage: A Review of Hallucinations in Large Language Models
Ye, H., Liu, T., Zhang, A., Hua, W., & Jia, W. (2023) · arXiv · doi:10.48550/arxiv.2309.06794 · Cited by 61 · see citing works
Generative Ai in EU Law: Liability, Privacy, Intellectual Property, and Cybersecurity
Novelli, C., Casolari, F., Hacker, P., Spedicato, G., & Floridi, L. (2024) · doi:10.2139/ssrn.4821952 · Cited by 57 · see citing works
ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate
Chan, C.-M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S., Fu, J., & Liu, Z. (2023) · arXiv · doi:10.48550/arxiv.2308.07201 · Cited by 55 · see citing works
“Unlike previous work like Du et al. (2023), we do not explicitly ask the debater agents to reach a consensus at the end of the debate.”
BackgroundResultInfluential
Reasoning or Reciting? Exploring the Capabilities and Limitations of Language Models Through Counterfactual Tasks
Wu, Z., Qiu, L., Ross, A., Akyürek, E., Chen, B., Wang, B., Kim, N., Andreas, J., & Kim, Y. (2024) · Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) · doi:10.18653/v1/2024.naacl-long.102 · Cited by 58 · see citing works
D-Bot: Database Diagnosis System using Large Language Models
Zhou, X., Li, G., Sun, Z., Liu, Z., Chen, W., Wu, J., Liu, J., Feng, R., & Zeng, G. (2024) · Proceedings of the VLDB Endowment · doi:10.14778/3675034.3675043 · Cited by 75 · see citing works
Agent design pattern catalogue: A collection of architectural patterns for foundation model based agents
Liu, Y., Lo, S. K., Lu, Q., Zhu, L., Zhao, D., Xu, X., Harrer, S., & Whittle, J. (2025) · Journal of Systems and Software · doi:10.1016/j.jss.2024.112278 · Cited by 51 · see citing works
Reinforcement Learning for Generative AI: State of the Art, Opportunities and Open Research Challenges
Franceschelli, G., & Musolesi, M. (2024) · Journal of Artificial Intelligence Research · doi:10.1613/jair.1.15278 · Cited by 52 · see citing works
“Using AI itself to provide feedback is also an option; notably, AI-based feedback is also used outside the RL paradigm, e.g., to provide verbal feedback to be appended to prompts (Shinn et al., 2023) or to collaborate with other LLMs at inference time (Dong et al., 2023; Du et al., 2023).”
BackgroundMethod
Can ChatGPT Defend its Belief in Truth? Evaluating LLM Reasoning via Debate
Wang, B., Yue, X., & Sun, H. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.795 · Cited by 71 · see citing works
“Prior work shows that LLMs can improve their responses via feedback (Shinn et al., 2023; Paul et al., 2023; Madaan et al., 2023; Ganguli et al., 2023; Ma et al., 2023; Chen et al., 2023b; Peng et al., 2023; Kim et al., 2023; Du et al., 2023; Liang et al., 2023; Chen et al., 2023a; Pan et al., 2023).”
Background
Synergistic Integration of Large Language Models and Cognitive Architectures for Robust AI: An Exploratory Analysis
Romero, O. J., Zimmerman, J., Steinfeld, A. M., & Tomasic, A. (2024) · Proceedings of the AAAI Symposium Series · doi:10.1609/aaaiss.v2i1.27706 · Cited by 44 · see citing works
Towards Reliable Misinformation Mitigation: Generalization, Uncertainty, and GPT-4
Pelrine, K., Imouza, A., Thibault, C., Reksoprodjo, M., Gupta, C., Christoph, J. N., Godbout, J., & Reihaneh Rabbany, K. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.395 · Cited by 39 · see citing works
Large Language Models Cannot Self-Correct Reasoning Yet
Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., & Zhou, D. (2023) · arXiv · doi:10.48550/arxiv.2310.01798 · Cited by 36 · see citing works
“Transitioning to another facet of self-correction, we investigate the potential of multi-agent debate (Du et al., 2023; Liang et al., 2023) as a means to improve reasoning.”
BackgroundInfluential
ResearchAgent: Iterative Research Idea Generation over Scientific Literature with Large Language Models
Baek, J., Jauhar, S. K., Cucerzan, S., & Hwang, S. J. (2025) · Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) · doi:10.18653/v1/2025.naacl-long.342 · Cited by 46 · see citing works
ReviewerGPT? An Exploratory Study on Using Large Language Models for Paper Reviewing
Liu, R., & Shah, N. B. (2023) · arXiv · doi:10.48550/arxiv.2306.00622 · Cited by 23 · see citing works
Mitigating reasoning hallucination through Multi-agent Collaborative Filtering
Shi, J., Zhao, J., Wu, X., Xu, R., Jiang, Y.-H., & He, L. (2024) · Expert Systems with Applications · doi:10.1016/j.eswa.2024.125723 · Cited by 30 · see citing works
Large Language Models for Telecom: The Next Big Thing?Preprint
Bariah, L., Zhao, Q., Zou, H., Tian, Y. T., Bader, F., & Debbah, M. A. (2023) · doi:10.36227/techrxiv.23536440.v1 · Cited by 36 · see citing works
Multi-Agent Collaboration Mechanisms: A Survey of LLMs
Tran, K.-T., Dao, D., Nguyen, M.-D., Pham, Q.-V., O'Sullivan, B., & Nguyen, H. D. (2025) · arXiv · doi:10.48550/arxiv.2501.06322 · Cited by 26 · see citing works
“Researchers have found taking multiple LLM instances to debate for a fixed number of rounds can boost their factuality and reasoning capabilities [41, 77, 140].”
Mixture-of-Agents Enhances Large Language Model Capabilities
Wang, J., Wang, J., Athiwaratkun, B., Zhang, C., & Zou, J. (2024) · arXiv · doi:10.48550/arxiv.2406.04692 · Cited by 25 · see citing works
“Du et al. (2023) establishes a mechanism for symmetric discussions among agents.”
Background
Automating Evaluation of AI Text Generation in Healthcare with a Large Language Model (LLM)-as-a-Judge
Croxford, E., Gao, Y., First, E., Pellegrino, N., Schnier, M., Caskey, J., Oguss, M., Wills, G., Chen, G., Dligach, D., et al. (2025) · doi:10.1101/2025.04.22.25326219 · Cited by 25 · see citing works
(Perhaps) Beyond Human Translation: Harnessing Multi-Agent Collaboration for Translating Ultra-Long Literary Texts
Wu, M., Xu, J., Yuan, Y., Haffari, G., Wan, L., Luo, W., & Zhang, K. (2025) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl.a.25 · Cited by 31 · see citing works
R-Tuning: Instructing Large Language Models to Say ‘I Don’t Know’
Zhang, H., Diao, S., Lin, Y., Fung, Y., Lian, Q., Wang, X., Chen, Y., Ji, H., & Zhang, T. (2024) · Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) · doi:10.18653/v1/2024.naacl-long.394 · Cited by 28 · see citing works
How AI agents will change cancer research and oncology
Lee, Y., Ferber, D., Rood, J., Regev, A., & Kather, J. N. (2024) · Nature Cancer · doi:10.1038/s43018-024-00861-7 · Cited by 23 · see citing works
Biomedical Natural Language Processing in the Era of Large Language Models
Usuyama, N., Wong, C., Zhang, S., Naumann, T., & Poon, H. (2025) · Annual Review of Biomedical Data Science · doi:10.1146/annurev-biodatasci-103123-095406 · Cited by 21 · see citing works
DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models
Chuang, Y.-S., Xie, Y., Luo, H., Kim, Y., Glass, J., & He, P. (2023) · arXiv · doi:10.48550/arxiv.2309.03883 · Cited by 29 · see citing works
“Recent strategies involve inference-time self-consistency checks (Manakul et al., 2023), multi-agent debating (Du et al., 2023; Liang et al., 2023), and inference-time intervention using human labels (Li et al., 2023).”
Background
More Agents Is All You Need
Li, J., Zhang, Q., Yu, Y., Fu, Q., & Ye, D. (2024) · arXiv · doi:10.48550/arxiv.2402.05120 · Cited by 22 · see citing works
LLM-generated competence-based e-assessment items for higher education mathematics: methodology and evaluation
Meissner, R., Pögelt, A., Ihsberner, K., Grüttmüller, M., Tornack, S., Thor, A., Pengel, N., Wollersheim, H.-W., & Hardt, W. (2024) · Frontiers in Education · doi:10.3389/feduc.2024.1427502 · Cited by 26 · see citing works
“…chain, resulting in a ﬁne-tuned Zero Shot Chain-of-Thought prompting approach (Wei et al., 2022a; Kojima et al., 2022), which incorporates a prompt template and injection of specially prepared knowledge (Martino et al., 2023), augmented by a simpliﬁed Multiple-Agent approach (Du et al., 2024).”
Exchange-of-Thought: Enhancing Large Language Model Capabilities through Cross-Model Communication
Yin, Z., Sun, Q., Chang, C., Guo, Q., Dai, J., Huang, X., & Qiu, X. (2023) · Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/2023.emnlp-main.936 · Cited by 26 · see citing works
Improving Multi-Agent Debate with Sparse Communication Topology
Li, Y., Du, Y., Zhang, J., Hou, L., Grabowski, P., Li, Y., & Ie, E. (2024) · Findings of the Association for Computational Linguistics: EMNLP 2024 · doi:10.18653/v1/2024.findings-emnlp.427 · Cited by 20 · see citing works
Self-citation
Dual Process Theory for Large Language Models: An overview of using Psychology to address hallucination and reliability issues
Bellini-Leite, S. C. (2023) · Adaptive Behavior · doi:10.1177/10597123231206604 · Cited by 22 · see citing works
“Self-Consistency differs from society of mind approaches (Du et al., 2023; Zhuge et al., 2023), that have been recently proposed, which involves the utilization of multiple language model instances to collaboratively generate a uni ﬁ ed answer to a given query.”
Towards Responsible Generative AI: A Reference Architecture for Designing Foundation Model Based Agents
Lu, Q., Zhu, L., Xu, X., Xing, Z., Harrer, S., & Whittle, J. (2024) · IEEE International Conference on Software Architecture Companion (ICSA-C) · doi:10.1109/icsa-c63560.2024.00028 · Cited by 26 · see citing works
“Debate-based cooperation integrates concepts from game theory, where one agent can receive feedback from other agents [52], [53] and adjust the thoughts and behaviours until consensus is reached [54], [55].”
Background
AgentCoord: Visually exploring coordination strategy for LLM-based multi-agent collaboration
Pan, B., Lu, J., Wang, K., Li, Z., Wen, Z., Feng, Y., Zhu, M., & Chen, W. (2025) · Computers & Graphics · doi:10.1016/j.cag.2025.104338 · Cited by 19 · see citing works
A Survey of Reasoning with Foundation ModelsPreprint
Sun, J., Zheng, C., Xie, E., Liu, Z., Chu, R., Liu, J., Xu, J., Ding, M., Li, H., Geng, M., et al. (2023) · doi:10.31219/osf.io/ac4sp · Cited by 51 · see citing works
Why Do Multi-Agent LLM Systems Fail?
Cemri, M., Pan, M. Z., Yang, S., Agrawal, L. A., Chopra, B., Tiwari, R., Keutzer, K., Parameswaran, A., Klein, D., Ramchandran, K., et al. (2025) · arXiv · doi:10.48550/arxiv.2503.13657 · Cited by 15 · see citing works
“Our empirical analysis reveals 41% to 86.7% failure rate on 7 state-of-the-art (SOTA) open-source MAS detailed in Figure 5 (Appendix C).”
Influential
On the Calibration of Large Language Models and Alignment
Zhu, C., Xu, B., Wang, Q., Zhang, Y., & Mao, Z. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.654 · Cited by 27 · see citing works
Voting or Consensus? Decision-Making in Multi-Agent Debate
Kaesberg, L. B., Becker, J., Wahle, J. P., Ruas, T., & Gipp, B. (2025) · Findings of the Association for Computational Linguistics: ACL 2025 · doi:10.18653/v1/2025.findings-acl.606 · Cited by 14 · see citing works
“In the following, we explain the multi-agent environment, decision protocols, response generators, and datasets used in our experiments.”
MultiAgent Collaboration Attack: Investigating Adversarial Attacks in Large Language Model Collaborations via Debate
Amayuelas, A., Yang, X., Antoniades, A., Hua, W., Pan, L., & Wang, W. Y. (2024) · Findings of the Association for Computational Linguistics: EMNLP 2024 · doi:10.18653/v1/2024.findings-emnlp.407 · Cited by 15 · see citing works
AI agents in chemical research: GVIM – an intelligent research assistant system
Ma, K. (2025) · Digital Discovery · doi:10.1039/d4dd00398e · Cited by 14 · see citing works
AI-powered Automatic Item Generation for Psychological Tests: A Conceptual Framework for an LLM-based Multi-Agent AIG System
Lee, P., Son, M.-K., & Jia, Z. (2025) · Journal of business and psychology · doi:10.1007/s10869-025-10067-y · Cited by 18 · see citing works
“Such interfaces have been successfully applied across a range of domains, including software development (Qian et al., 2023), games (Park et al., 2023), robotics (Zhang et al., 2023), economics (Li et al., 2023), political science (Xiao et al., 2023), and public health (Williams et al., 2023).”
Sustainable Digitalization of Business with Multi-Agent RAG and LLM
Arslan, M., Munawar, S., & Cruz, C. (2024) · Procedia Computer Science · doi:10.1016/j.procs.2024.09.337 · Cited by 16 · see citing works
“Multi-Agent Debate is explored [31, 32] as an effective method to stimulate divergent thinking and enhance factuality and reasoning in LLMs. MetaGPT [33] introduces a specialized LLM application for automatic software development.”
Debating with More Persuasive LLMs Leads to More Truthful Answers
Khan, A., Hughes, J., Valentine, D., Ruis, L., Sachan, K., Radhakrishnan, A., Grefenstette, E., Bowman, S. R., Rocktäschel, T., & Perez, E. (2024) · arXiv · doi:10.48550/arxiv.2402.06782 · Cited by 13 · see citing works
Studies on the Use of Large Language Models for the Automation of Business Processes in Enterprise Resource Planning Systems
Schnepf, J., Engin, T., Anderer, S., & Scheuermann, B. (2024) · Lecture notes in computer science · doi:10.1007/978-3-031-70239-6_2 · Cited by 10 · see citing works
Zero-shot Knowledge Graph Question Generation via Multi-agent LLMs and Small Models Synthesis
Zhao, R., Tang, J., Zeng, W., Chen, Z., & Zhao, X. H. (2024) · ACM International Conference on Information and Knowledge Management (CIKM) · doi:10.1145/3627673.3679805 · Cited by 10 · see citing works
Stock Market Forecasting: From Traditional Predictive Models to Large Language Models
Darwish, M., Hassanien, E. E., & Eissa, A. H. B. (2025) · Computational Economics · doi:10.1007/s10614-025-11024-w · Cited by 27 · see citing works
“…and 5 , some research studies show that performance can be enhanced through fine-tun - ning LLMs with financial tailored information using LoRA (Zhang & Yang, 2023) or hybrid tuning methodology (Raffel et al., 2020 ) or building specialized financial LLMs (Li et al., 2023 ; Le Scao et al., 2022 ).”
Influential
Evil Geniuses: Delving into the Safety of LLM-based Agents
Tian, Y., Yang, X., Zhang, J., Dong, Y., & Su, H. (2023) · arXiv · doi:10.48550/arxiv.2311.11855 · Cited by 10 · see citing works
Generative AI Agents for Knowledge Work Augmentation in Finance
Ganesh, S., Ardon, L., Borrajo, D., Garg, D., Madhushani, U., Narayanan, A. L., Canonaco, G., & Veloso, M. (2024) · Annual Review of Control Robotics and Autonomous Systems · doi:10.1146/annurev-control-022823-033813 · Cited by 10 · see citing works
Leveraging Large Language Models for Suicide Detection on Social Media with Limited Labels
Nguyen, V., & Pham, C. (2024) · IEEE International Conference on Big Data (BigData) · doi:10.1109/bigdata62323.2024.10825313 · Cited by 14 · see citing works
“[22] organized multiple LLMs in a debate setting, where each LLM independently generates an initial response, and then communicates to reach a consensus on a final answer.”
A brain-inspired agentic architecture to improve planning with LLMs
Webb, T. W., Mondal, S. S., & Momennejad, I. (2025) · Nature Communications · doi:10.1038/s41467-025-63804-5 · Cited by 16 · see citing works
“…a version of GPT-4 that was also given access to detailed explanations of those solutions, sometimes referred to as chain-of-thought 36 (GPT-4 CoT); Tree-of-Thought (ToT) 37 , a method that combines LLMs with tree search, similar to the search performed by MAP; and Multi-Agent Debate (MAD) 38 .”
MARG: Multi-Agent Review Generation for Scientific Papers
D'Arcy, M., Hope, T., Birnbaum, L., & Downey, D. (2024) · arXiv · doi:10.48550/arxiv.2401.04259 · Cited by 9 · see citing works
Advancements and future directions of artificial intelligence in tumor imaging: A comprehensive review of techniques and applications
Xue, J., Yu, J., Mao, Q., & Gu, X. (2025) · EngMedicine · doi:10.1016/j.engmed.2025.100086 · Cited by 10 · see citing works
Auto-scaling LLM-based multi-agent systems through dynamic integration of agents
Perera, R., Basnayake, A., & Wickramasinghe, M. (2025) · Frontiers in Artificial Intelligence · doi:10.3389/frai.2025.1638227 · Cited by 7 · see citing works
“Multi-agent debate is an eﬀective way to improve reasoning, adaptability, and factual accuracy in language models (Du et al., 2023; Talebirad and Nadiri, 2023).”
Automated Design of Agentic Systems
Hu, S., Lu, C., & Clune, J. (2024) · arXiv · doi:10.48550/arxiv.2408.08435 · Cited by 10 · see citing works
“…(Madaan et al., 2024; Shinn et al., 2023), which allows iterative self-reflection to correct mistakes made in previous attempts; (4) LLM-Debate (Du et al., 2023), which enables different LLMs to debate with each other, leveraging diverse perspectives to find better answers; (5)…”
Influential
Can Large Language Models Really Improve by Self-critiquing Their Own Plans?
Valmeekam, K., Marquez, M., & Kambhampati, S. (2023) · arXiv · doi:10.48550/arxiv.2310.08118 · Cited by 8 · see citing works
Large Language Models for Telecom: The Next Big Thing?Preprint
Bariah, L., Zhao, Q., Zou, H., Tian, Y. T., Bader, F., & Debbah, M. A. (2023) · doi:10.36227/techrxiv.23536440 · Cited by 7 · see citing works
EcoAssistant: Using LLM Assistant More Affordably and Accurately
Zhang, J., Krishna, R., Awadallah, A. H., & Wang, C. (2023) · arXiv · doi:10.48550/arxiv.2310.03046 · Cited by 8 · see citing works
“, 2023) or to improve factuality and reasoning (Du et al., 2023).”
Background
Scaling Large Language Model-based Multi-Agent Collaboration
Qian, C., Xie, Z., Wang, Y., Liu, W., Zhu, K., Xia, H., Dang, Y., Du, Z., Chen, W., Yang, C., et al. (2024) · arXiv · doi:10.48550/arxiv.2406.07155 · Cited by 8 · see citing works
“In contrast, agent coordination, based on the implicit knowledge of pretrained LLMs, leverages the understanding and refinement of textual information through linguistic interactions, often circumventing the extensive scaling needed by neuronal coordination.”
TradingAgents: Multi-Agents LLM Financial Trading Framework
Xiao, Y., Sun, E., Luo, D., & Wang, W. (2024) · arXiv · doi:10.48550/arxiv.2412.20138 · Cited by 6 · see citing works
“These concise, focused discussions have been shown to promote deeper reasoning and integrate diverse perspectives, enabling more balanced decisions in complex, long-horizon scenarios—a method particularly relevant to the intricate environment of trading (Du et al., 2023).”
Navigating Bias: Using LLMs to Analyze Discrimination in Entrepreneurial Game Dialogues
Marsden, N., Rothe‐Wulf, A., Tang, Y., Herling, C., & Reichert, T. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-92578-8_14 · Cited by 6 · see citing works
An LLM-based multi-agent system for remote sensing analysis
Sun, Z., Zhou, Y., & Yang, J. (2026) · Big Earth Data · doi:10.1080/20964471.2025.2600178 · Cited by 8 · see citing works
“Since multi-agent systems have been shown to improve scalability and reasoning effectiveness (Du et al., 2023; Wu et al., 2023), we adopt a multi-agent structure, setting it apart from existing tools that rely on single-agent architectures.”
Influential
Prompt Engineering Through the Lens of Optimal Control
Luo, Y., Tang, Y., Shen, C., Zhou, Z., & Dong, B. (2023) · Journal of Machine Learning · doi:10.4208/jml.231023 · Cited by 9 · see citing works
“Tipping the Balance”: Human Intervention in Large Language Model Multi‐Agent Debate
Triem, H., & Ding, Y. (2024) · Proceedings of the Association for Information Science and Technology · doi:10.1002/pra2.1034 · Cited by 7 · see citing works
“Recently, multiagent work has centered around two broad structures: “collaboration” and “debate,” both of which allow for improved complexity in discussion (Du et al., 2023; H. Li et al., 2023; K. Xiong et al., 2023; Liang et al., 2023).”
Collaborative garment design through group chatting with generative industrial large models
Harish, A. R., Yuan, Z., Li, M., Yang, H., & Huang, G. Q. (2025) · Advanced Engineering Informatics · doi:10.1016/j.aei.2025.103366 · Cited by 6 · see citing works
Learning from Mistakes via Cooperative Study Assistant for Large Language Models
Wang, D., & Li, L. (2023) · Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.18653/v1/2023.emnlp-main.659 · Cited by 10 · see citing works
“Du et al. (2023) uses responses from multiple agents and the debate between these agents as the feedback.”
Background
Large language models for biomolecular analysis: From methods to applications
Feng, R., Zhang, C., & Zhang, Y. (2024) · TrAC Trends in Analytical Chemistry · doi:10.1016/j.trac.2024.117540 · Cited by 7 · see citing works
Social Learning in Neural Agent-Based Models
Douven, I. (2024) · Philosophy of Science · doi:10.1017/psa.2024.33 · Cited by 6 · see citing works
“Recent work has shown how LLMs can be made to communicate in that one LLM ’ s output serves as the prompt for one or more other LLMs, and so on, recursively (Du et al. 2023).”
On scalable oversight with weak LLMs judging strong LLMs
Kenton, Z., Siegel, N. Y., Kramár, J., Brown-Cohen, J., Albanie, S., Bulian, J., Agarwal, R., Lindner, D., Tang, Y., Goodman, N. D., et al. (2024) · arXiv · doi:10.48550/arxiv.2407.04622 · Cited by 6 · see citing works
A Self-Evolving Framework for Multi-Agent Medical Consultation Based on Large Language Models
Chen, K., Qi, J., Huo, J., Tian, P., Meng, F., Yang, X., & Gao, Y. (2025) · ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp49660.2025.10889517 · Cited by 6 · see citing works
General-Purpose Models for the Chemical Sciences: LLMs and Beyond
Alampara, N., Aneesh, A., Ríos-García, M., Mirza, A., Schilling-Wilhelmi, M., Aghajani, A. A., Sun, M., Prastalo, G., & Jablonka, K. M. (2026) · Chemical Reviews · doi:10.1021/acs.chemrev.5c00583 · Cited by 12 · see citing works
Should we be going MAD? A Look at Multi-Agent Debate Strategies for LLMs
Smit, A., Duckworth, P., Grinsztajn, N., Barrett, T. D., & Pretorius, A. (2023) · arXiv · doi:10.48550/arxiv.2311.17371 · Cited by 5 · see citing works
Gödel Agent: A Self-Referential Agent Framework for Recursively Self-Improvement
Yin, X., Wang, X., Pan, L., Lin, L., Wan, X., & Wang, W. Y. (2025) · Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2025.acl-long.1354 · Cited by 5 · see citing works
LLM-SAP: Large Language Models Situational Awareness-Based Planning
Wang, L., & Zhong, H. (2024) · IEEE International Conference on Multimedia and Expo workshops · doi:10.1109/icmew63481.2024.10645429 · Cited by 6 · see citing works
“Such collaboration leverages the individual strengths of each agent [28, 29, 30].”
Background
Beyond Isolation: Multi-agent Synergy for Improving Knowledge Graph Construction
Ye, H., Gui, H., Zhang, A., Liu, T., & Jia, W. (2025) · Communications in computer and information science · doi:10.1007/978-981-96-1809-5_6 · Cited by 6 · see citing works
Learning to Judge: LLMs Designing and Applying Evaluation Rubrics
Siro, C., Aliannejadi, P., & Aliannejadi, M. (2026) · Findings of the Association for Computational Linguistics: EACL 2026 · doi:10.18653/v1/2026.findings-eacl.335 · Cited by 4 · see citing works
Can We Trust AI Agents? A Case Study of an LLM-Based Multi-Agent System for Ethical AI
de Cerqueira, J. A. S., Agbese, M., Rousi, R., Xi, N., Hamari, J., & Abrahamsson, P. (2024) · arXiv · doi:10.48550/arxiv.2411.08881 · Cited by 4 · see citing works
“The aforementioned studies run in line with a joint study by MIT and Microsoft, Du et al. [24].”
Cooperative Scheduling and Hierarchical Memory Model for Multi-Agent Systems
Zou, H., Li, R., Sun, T., Wang, F., Li, T., & Liu, K. (2024) · 2024 IEEE International Symposium on Product Compliance Engineering - Asia (ISPCE-ASIA) · doi:10.1109/ispce-asia64773.2024.10756271 · Cited by 4 · see citing works

Authors: Du, Y., Li, S., Torralba, A., Tenenbaum, J. B., & Mordatch, I.
Affiliations: MIT Computer Science and Artificial Intelligence Laboratory (US)
Google (United States) (US)
ORCID: Li, Shuang
Torralba, Antonio
Tenenbaum, Joshua B.
Venue: arXiv
Publisher: arXiv
Published: 2023
ISSN: 2331-8422 (ISSN-L)
DOI: 10.48550/arxiv.2305.14325
arXiv: 2305.14325
OpenAlex: W4378509427
Cited by: 96
References: 48
Citations by year: 2026: 16 · 2025: 26 · 2024: 43 · 2023: 10
Type: posted content
Language: en
Open access: green
arxiv.org
Licence: cc-by
creativecommons.org/licenses/by/4.0/legalcode
Locations: arXiv (Cornell University) (submittedVersion, cc-by)
arXiv (Cornell University) (cc-by)
Topics: Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
Keywords: large language models, factuality, hallucination mitigation, mathematical reasoning, strategic reasoning, Society of Mind, self-consistency

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Improving Factuality and Reasoning in Language Models through Multiagent Debate

Abstract

AI snapshot

TL;DR

Key findings

Methods

Limitations

Integrity signals

Claims the paper makes

Claims other papers make about this one

Self-check

Cites

Cited by

Related

Record

About these checks