Source-linked AI summary
Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate
Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, Zhaopeng Tu
TL;DR
Self-reflection can suffer from Degeneration-of-Thought, in which LLMs remain attached to incorrect solutions and fail to generate novel thoughts. The paper proposes Multi-Agent Debate, where agents exchange arguments under a judge, and finds effectiveness on two challenging tasks, while identifying debate-design requirements and judge-fairness concerns.
Problem
Self-reflection can suffer from the Degeneration-of-Thought problem, where LLMs remain attached to incorrect answers and cannot engage in meaningful later reflection.
Method
Multi-Agent Debate uses multiple agents to exchange arguments in a “tit for tat” state while a judge manages the debate and produces a final solution.
Results
MAD is effective on Common MT and Counter-Intuitive AR, and GPT-3.5-Turbo with MAD can surpass GPT-4 on Common MT.
Takeaways & Limitations
Adaptive debate stopping and a modest “tit for tat” level are required for good MAD performance, while complex samples require more debate rounds.
Takeaways & Limitations
MAD requires more interaction time, and LLM judges may prefer outputs generated by themselves or fail to act impartially when agents use different LLMs.
Abstract
from arXiv · showhide
Modern large language models (LLMs) like ChatGPT have shown remarkable performance on general language tasks but still struggle on complex reasoning tasks, which drives the research on cognitive behaviors of LLMs to explore human-like problem-solving strategies. Along this direction, one representative strategy is self-reflection, which asks an LLM to refine the solution with the feedback generated by itself iteratively. However, our study shows that such reflection-style methods suffer from the Degeneration-of-Thought (DoT) problem: once the LLM has established confidence in its solutions, it is unable to generate novel thoughts later through reflection even if its initial stance is incorrect. To address the DoT problem, we propose a Multi-Agent Debate (MAD) framework, in which multiple agents express their arguments in the state of "tit for tat" and a judge manages the debate process to obtain a final solution. Clearly, our MAD framework encourages divergent thinking in LLMs which would be helpful for tasks that require deep levels of contemplation. Experiment results on two challenging datasets, commonsense machine translation and counter-intuitive arithmetic reasoning, demonstrate the effectiveness of our MAD framework. Extensive analyses suggest that the adaptive break of debate and the modest level of "tit for tat" state are required for MAD to obtain good performance. Moreover, we find that LLMs might not be a fair judge if different LLMs are used for agents. Code is available at https://github.com/Skytliang/Multi-Agents-Debate.
1 Introduction
The paper identifies Degeneration-of-Thought in self-reflection, where confident LLMs fail to generate novel thoughts despite incorrect initial answers, and proposes Multi-Agent Debate to encourage divergent reasoning. Across Common MT and Counter-Intuitive AR, MAD outperforms baselines, with GPT-3.5-Turbo plus MAD surpassing GPT-4 on Common MT.
- Problem: Degeneration-of-Thought occurs when an LLM cannot generate novel thoughts through self-reflection after becoming confident in an incorrect initial stance.The authors define this problem as a failure of iterative reflection to produce meaningful disagreement or alternative reasoning.
- Problem: Low disagreement across self-reflection iterations suggests that LLMs remain attached to incorrect chain-of-thought answers.The demonstration forces five rounds and measures opposition between adjacent iterations.
- Problem: Three proposed contributors to DoT are biased perception, resistance to change, and limited external feedback.The paper argues that internal reflection can miss blind spots and alternative viewpoints.
- Approach: MAD uses multiple agents that exchange arguments in a “tit for tat” state while a judge manages the debate and selects the final solution.The framework is intended to explore divergent chains of thought and provide external feedback.
- Evaluation: MAD is evaluated on Commonsense Machine Translation and Counter-Intuitive Arithmetic Reasoning, where superficial intuitions are often incorrect.These tasks require deeper contemplation than the initial surface interpretation.
- Results: GPT-3.5-Turbo with MAD surpasses GPT-4 on the Common MT dataset.The paper reports this result as evidence of MAD’s effectiveness against baseline methods.
- Analysis: Adaptive debate breaking, a modest “tit for tat” level, and an unbiased judge are important considerations for MAD performance.The analyses report adaptive stopping and moderate tit-for-tat interaction as required for improvement, while noting judge preference for its backbone model.
2 Multi-Agent Debate Framework
MAD organizes debate among affirmative and negative agents, with a judge monitoring rounds, deciding whether a solution has been obtained, and extracting a final answer when needed. Its prompts and interaction history enforce opposing arguments and support divergent reasoning.
- Framework: MAD consists of debaters that speak sequentially from the previous debate history and a judge that manages the process.The framework description includes multiple debaters, fixed speaking order, accumulated history, and judge oversight.
- Framework: Meta prompts specify the topic, number of debaters, iteration limit, and other debate requirements.These prompts establish the conditions under which the agents interact.
- Debaters: The “tit for tat” requirement creates an atmosphere of debate in which agents respond to one another’s arguments.The affirmative and negative roles are framed as devil and angel, with the angel intended to correct the devil’s mistakes.
- Debaters: The affirmative debater presents viewpoints, while the negative debater disagrees with the affirmative side and provides reasons and an answer.The prompts assign explicitly opposing roles to create debate.
- Judge: In Discriminative Mode, the judge decides whether the current iteration has produced a correct solution; if not, debate continues.A true decision ends the debate, whereas a false decision causes further interaction.
- Judge: In Extractive Mode, the judge extracts the final solution from the full debate history when no correct solution is identified within the iteration limit.This mode provides an answer after the allowed debate rounds are exhausted.
3 Experiment
Experiments evaluate MAD on commonsense machine translation and counter-intuitive arithmetic reasoning, comparing it with self-reflection and task-specific baselines. Across these tasks, MAD improves over GPT-3.5-Turbo baselines, surpasses GPT-4 on Common MT, and outperforms other GPT-3.5-Turbo methods on Counter-Intuitive AR.
- 3.1 Challenging Testbeds: The experiments cover Common MT, which tests lexical and syntactic ambiguity resolution, and Counter-Intuitive AR, which requires resistance to intuitive traps and multi-step reasoning.Common MT contains Chinese⇒English translation examples, while the Counter-Intuitive AR dataset contains 200 questions.
- 3.2 Setups: The study compares MAD with Self-Reflect and task-specific baselines including Rerank, MAPS, CoT, and Self-Consistency.Rerank and MAPS are included for Common MT; CoT and Self-Consistency are included for Counter-Intuitive AR.
- 3.3 Results on Common MT: MAD with GPT-3.5-Turbo demonstrates significant improvements over GPT-4 across automatic and human Common MT evaluations.Common MT uses COMET and BLEURT together with human translation-quality assessments on a 1-to-5 scale.
- 3.4 Results on Counter-Intuitive AR: MAD outperforms all other GPT-3.5-Turbo-based compared methods on Counter-Intuitive AR, although it does not match GPT-4.Self-Reflect provides only marginal improvement over the GPT-3.5-Turbo baseline, whereas CoT and Self-Consistency improve more.
- 3.4 Results on Counter-Intuitive AR: In a Counter-Intuitive AR example, MAD reaches the correct answer 4 after CoT and Self-Reflect incorrectly output 3.The MAD reasoning considers both rotation around circle B and rotation of circle A itself.
4 Analysis
The analysis examines how MAD mitigates Degeneration-of-Thought through diverse debate, judge behavior, debater configuration, adaptive stopping, and debate intensity. It finds that strong debaters, adaptive breaks, and moderate disagreement support performance, while heterogeneous agents and excessive debate can introduce problems.
- Mitigation of DoT: MAD introduces other agents’ views through debate to address limited external feedback and mitigate bias and rigidity associated with DoT.The analysis evaluates bias through human assessment and diversity through Self-BLEU-based comparison of translation candidates.
- Analysis of Judge: Turbo debaters consistently outperform Vicuna debaters across test sets, making debater quality more important than judge selection.Vicuna as judge underperforms Turbo as judge when Turbo debaters are used.
- Analysis of Judge: LLM-based judges may favor the debater sharing their backbone LLM when agents use different models.This preference raises fairness concerns in heterogeneous-agent debates.
- Analysis of Debaters: Increasing the number of debaters can reduce performance because longer, more complex debates exceed LLMs’ ability to retain and summarize other agents’ views.The analysis attributes this degradation to limitations in long-text modeling.
- Tit for Tat: Higher disagreement is necessary for good MAD performance, but mandatory disagreement on every point is not optimal.The reported disagreement level of 0.988 for continuous disagreement does not yield the best performance and may encourage polarization.
- Debate Iteration: MAD generally outperforms self-reflection as iterations increase, but adaptive stopping performs better than forcing debate to continue.For most examples, a good translation is generated in the first iteration; longer forced debates can harm results, while complex sentences need additional rounds.
5 Related Work
The paper builds on chain-of-thought, self-evaluation, and generative-agent research while positioning MAD as a multi-agent debate framework for addressing DoT. Its distinctive elements include an adaptive-break judge, a DoT-focused objective, and empirical use of identical backbone models.
- Chain-of-Thought Prompting: Chain-of-thought prompting generates intermediate reasoning steps to improve multi-step problem solving.Related approaches focus primarily on prompt design and decoding strategies.
- Self-Evaluation: Self-evaluation methods generate an output, evaluate it, and use feedback to refine the result.Feedback may come from the model itself or from external environments.
- Generative Agents: MAD extends generative-agent research with an adaptive-break judge, a DoT-focused goal, and agents sharing an identical backbone LLM.The paper contrasts these features with concurrent multi-agent debate studies.
6 Conclusion
The paper defines DoT as a self-reflection problem and proposes MAD to promote divergent chain-of-thoughts. Experiments and analyses support MAD’s effectiveness while identifying adaptive stopping, moderate disagreement, complex-sample iteration needs, and judge bias across models.
- Conclusion: The paper defines Degeneration-of-Thought in self-reflection and addresses it with Multi-Agent Debate to explore divergent chain-of-thoughts.MAD is evaluated on commonsense machine translation and counter-intuitive arithmetic reasoning.
- Conclusion: MAD improves performance on two challenging tasks, with GPT-3.5-Turbo using MAD surpassing GPT-4 on Common MT.The conclusion highlights adaptive breaks and a modest tit-for-tat state as requirements for good performance.
- Conclusion: The paper reports that complex samples require more debate rounds and that LLMs may be unfair judges when agents use different LLMs.Future work includes scheduling more agents appropriately and applying multi-agent intelligence to additional settings.
Limitations
The paper’s limitations include higher interaction costs, difficulties maintaining coherence in long contexts, and potential judge bias. Its evaluations use two specialized challenging task settings, including ambiguity-sensitive translation and counter-intuitive arithmetic reasoning.
- Limitations: MAD requires more time because agents exchange and refute arguments across multiple interaction rounds.This cost follows from the framework’s multi-round debate process.
- Limitations: LLM-based agents may lose coherence, relevance, and context in long scenarios, making long-text modeling a remaining challenge.The paper notes that adding debaters can increase debate complexity and information loss.
- Limitations: An LLM judge may prefer outputs generated by itself or by a debater using the same backbone model.The paper recommends either using one LLM for all roles or using distinct LLMs for judge and debaters to mitigate this bias.
- Evaluation Scope: The experiments cover Common MT and Counter-Intuitive AR, both designed to require deeper contemplation than superficial answers.Common MT tests lexical and syntactic ambiguity, while Counter-Intuitive AR tests resistance to intuitive traps and multi-step reasoning.
B Human Evaluation Details
Human evaluation randomly anonymized translations, used three professional translators, and calculated average method scores; annotators also assessed commonsense conformity and test-set bias.
- Three professional translators anonymously assessed all methods together after sentence display orders were randomly shuffled.Inter-annotator agreement was Krippendorff’s Alpha = 0.76.
- Human Score was computed by averaging the translators’ scores for each method.
- Annotators labeled each CommonMT sentence as 1 or 0 according to whether specific-word translations conformed to commonsense.The reported degree of bias covered the whole test set.
- The evaluation used the Chinese word “吃掉” as a lexical-ambiguity example, favoring “destroy” over the straightforward “eat up” translation for commonsense conformity.
C Results on math and symbolic reasoning tasks
The paper compares accuracy across math and symbolic-reasoning benchmarks and examines prompts that vary the debaters’ “tit for tat” state.
- Accuracy is compared on GSM, AddSub, and three symbolic-reasoning datasets from Big-Bench.
- The “tit for tat” state is varied through natural-language instructions appended to the debaters’ meta prompt.
E Extra Computational Cost
Inference cost is measured by generated-token counts: self-reflection costs 1.83 times baseline COT, while MAD adds 0.63 times the baseline cost.
- Inference cost is measured by the number of generated tokens.
- 1.83 times the baseline COT inference cost is incurred by self-reflection.
- 0.63 times the baseline inference cost is added by MAD.The paper deems this increase acceptable because of MAD’s reported benefit for divergent thinking without external human signals.
F.1 Commonsense Machine Translation
The examples illustrate MAD debates over commonsense translation and counter-intuitive arithmetic, with opposing arguments used to resolve disputed interpretations and calculations.
- Commonsense Machine Translation: The translation debate contrasts the literal “eat up an enemy division” with the military-contextual “eliminate an enemy division.”
- Counter-Intuitive Arithmetic Reasoning: The arithmetic debate concerns Alice’s average speed while walking uphill at 1 m/s and downhill at 3 m/s.
- Counter-Intuitive Arithmetic Reasoning: The arithmetic example initially averages the two speeds directly, while the opposing argument invokes average speed as total distance divided by total time.