PaperlayerView exact revision on arXiv

Source-linked AI summary

Examining Inter-Consistency of Large Language Models Collaboration: An In-depth Analysis via Debate

Kai Xiong, Xiao Ding, Yixin Cao, Ting Liu, Bing Qin

arXiv:2305.11595v3cs.CLcs.AI

TL;DR

The paper addresses the underexplored inter-consistency of multiple LLMs collaborating toward a shared goal. It introduces FORD, a three-stage debate framework evaluated on commonsense reasoning, and finds that LLMs can reach consensus, although ability imbalances can produce domination and distraction.

  • Problem

    The paper studies whether multiple LLMs can collaborate toward a shared goal, complementing prior work focused mainly on inconsistency within a single LLM.

  • Method

    FORD conducts fair, mismatched, and roundtable debates, with alternating arguments and a judge summarizing the debate and producing the final conclusion.

  • Results

    LLMs can achieve consensus through debate, but stronger models may be distracted by weaker models in mismatched debates.

  • Takeaways & Limitations

    The findings support studying inter-consistency as a basis for developing future LLM collaboration methods.

  • Takeaways & Limitations

    The study is limited mainly to commonsense reasoning and multi-choice tasks, motivating evaluation on broader tasks and more kinds of LLMs.

Abstract

from arXiv · show

Large Language Models (LLMs) have shown impressive capabilities in various applications, but they still face various inconsistency issues. Existing works primarily focus on the inconsistency issues within a single LLM, while we complementarily explore the inter-consistency among multiple LLMs for collaboration. To examine whether LLMs can collaborate effectively to achieve a consensus for a shared goal, we focus on commonsense reasoning, and introduce a formal debate framework (FORD) to conduct a three-stage debate among LLMs with real-world scenarios alignment: fair debate, mismatched debate, and roundtable debate. Through extensive experiments on various datasets, LLMs can effectively collaborate to reach a consensus despite noticeable inter-inconsistencies, but imbalances in their abilities can lead to domination by superior LLMs. Leveraging a more advanced LLM like GPT-4 as an authoritative judge can boost collaboration performance. Our work contributes to understanding the inter-consistency among LLMs and lays the foundation for developing future collaboration methods. Codes and data are available at https://github.com/Waste-Wood/FORD

1 Introduction

The paper examines whether multiple LLMs can collaborate toward shared goals, focusing on inter-consistency rather than the single-model inconsistency studied by prior work. It introduces FORD, a three-stage debate framework, and finds that collaboration can reach consensus but is affected by ability imbalances.

  • The paper investigates whether multiple LLMs can cooperate effectively and performantly toward a shared goal.
  • LLMs can quickly shift viewpoints or remain steadfast, raising questions about compromise and consensus during collaboration.
  • FORD uses debate theory to quantitatively study inter-consistency and differences in LLMs’ understandings during collaboration.
  • FORD organizes collaboration into fair, mismatched, and roundtable debates aligned with comparable abilities, unequal abilities, and more than two LLMs.
  • LLMs can potentially reach consensus through FORD, but comparable models collaborate more effectively while superior models may dominate mismatched debates.

2 Preliminaries

The paper evaluates FORD on seven commonsense reasoning datasets using six LLMs and defines INCON to quantify inter-consistency. It compares single-model prediction with soft and hard collaboration baselines.

  • The experiments use seven multi-choice commonsense reasoning datasets spanning abductive, question-answering, causal, social, physical, and implicit reasoning tasks.
  • The study evaluates six LLMs, including four OpenAI models and two open-source 13B models, LLaMA and Vicuna.
  • INCON quantifies inter-inconsistency among multiple LLMs by comparing their predictions across dataset samples.
  • FORD’s workflow collects independent stances and arguments, debates inconsistent samples, and uses a judge to summarize the debate and produce a conclusion.
  • The baselines include single-LLM prediction, Collaboration-Soft averaging, and Collaboration-Hard prediction requiring all LLMs to be correct.

3 FORD: Formal Debate Framework

FORD is a three-step formal debate framework that elicits independent stances, alternates argument exchange, and summarizes the debate to study whether LLMs can collaborate toward consensus.

  • 3 FORD: Formal Debate Framework: The framework formalizes debate to explore differences between LLM understandings and whether mutual discussion can support consensus and performance gains.It is designed for quantitative analysis of LLM collaboration.
  • 3.1 Stance Selection & Argument Generation: FORD begins by asking each LLM to independently choose a stance and provide a supporting argument for each sample.The framework focuses subsequent collaboration on samples where the LLMs disagree.
  • 3.2 Alternate Debate Process: The alternate debate process has LLMs take turns countering arguments or seeking compromise until consensus or a round limit is reached.Stances are hidden during this stage to reduce viewpoint swings.
  • 3.2 Alternate Debate Process: Debate participants can defend their positions or compromise after considering arguments from earlier rounds.The framework illustrates this process with successive defenses followed by compromise.
  • 3.2 Alternate Debate Process: Table 2 compares FORD with collaboration baselines and single-LLM methods using dataset-level accuracy and average accuracy.Underlined values identify the best collaboration method, while bold values identify the best overall method.
  • 3.3 Debate Summarization: A judge summarizes the full debate and draws the final conclusion, using the consensus stance when agreement is reached and equal argument weights otherwise.The summary is generated from the arguments and is intended to make the debate process interpretable.
  • 3.3 Debate Summarization: FORD evaluates three scenarios: fair debates between comparable LLMs, mismatched debates between unequal LLMs, and roundtable debates with more than two LLMs.These scenarios support comprehensive investigation of collaboration under different capability arrangements.

4 Fair Debate

Fair debates show that comparable LLMs can progressively reduce inter-inconsistency and often improve task performance, although similarly capable models may provide little additional information.

  • 4.1 Initial INCON of LLMs Pairs: Different LLM types exhibit nearly 20%-30% initial INCON on almost all datasets despite sharing an underlying model.The overlapped portion contributes nearly 50% of INCON, indicating comparable but substantially different capabilities.
  • 4.1 Initial INCON of LLMs Pairs: ChatGPT-0301 does not supersede ChatGPT, indicating that newer versions gain some abilities while losing others.The authors therefore caution that updated models may not reproduce unavailable early-version results.
  • 4.2 Results of Fair Debate: FORD outperforms Col-S, Col-H, and corresponding single LLMs on almost all datasets, except for LLaMA–Vicuna on Social IQa.The reported exception is tied to the fair-debate performance comparison.
  • 4.2 Results of Fair Debate: FORD yields less improvement for ChatGPT–ChatGPT-0301 because their similar capabilities usually produce similar opinions and little new information.The same pair has higher performance floors under conservative collaboration, while ChatGPT–Davinci-003 has higher performance ceilings under FORD.
  • 4.2 Results of Fair Debate: Across fair debates, INCON decreases progressively after each round, showing that comparable LLMs can learn from differences and reach agreement.ChatGPT–Davinci-003 and ChatGPT–ChatGPT-0301 approach zero INCON, whereas LLaMA–Vicuna retain noticeable inconsistency.
  • 4.2 Results of Fair Debate: ChatGPT–ChatGPT-0301 reaches INCON coverage after 2 rounds, earlier than the other fair debates.Their similar capabilities produce similar opinions that converge more quickly.

5 Mismatched Debate

Mismatched debates can still reduce inter-inconsistency, but weaker participants disrupt collaboration and stronger LLMs impose a performance ceiling through dominance.

  • 5.1 Results of Mismatched Debate: FORD outperforms collaboration baselines and weaker LLMs but loses to stronger LLMs in mismatched debates.The resulting performance ceiling is tied to the stronger LLMs’ accuracy.
  • 5.1 Results of Mismatched Debate: INCON continues to drop despite mismatched capabilities, showing that unequal LLMs can reach consensus while weaker participants disturb collaboration.Consensus formation therefore does not guarantee desirable final performance.
  • 5.1 Results of Mismatched Debate: Dominant LLMs can still improve mismatched debates, although they may be distracted by weaker participants.The reported improvements occur for FORD involving ChatGPT–Davinci-003 and LLaMA–Vicuna counterparts.
  • 5.1 Results of Mismatched Debate: LLaMA–ChatGPT performs far below the ceiling because LLaMA cannot evaluate arguments and mainly repeats its stance, distracting ChatGPT.This behavior is presented as a specific failure mode of the mismatched pairing.
  • 5.2 The dominance of LLMs: Dominance measures the portion of initially inconsistent samples on which one LLM causes the other to compromise.The metric is intended to reflect how strongly each LLM adheres to its viewpoint.
  • 5.2 The dominance of LLMs: Stronger LLMs have an absolute dominance advantage in mismatched debates, though they become easier to disturb when less confident.This pattern is reported for GPT-4 and ChatGPT interacting with weaker models.
  • 5.2 The dominance of LLMs: LLaMA–ChatGPT shows a smaller dominance gap because LLaMA’s poor debating behavior makes ChatGPT swing.LLaMA often fails to evaluate others’ arguments and produces only a simple stance claim.

6 Roundtable Debate

The roundtable experiments test whether three LLMs can collaborate under mismatched and fair ability settings. FORD reduces inconsistency and improves performance, with stronger models helping when they are not overwhelmed by weaker participants.

  • Experimental setup: The study compares a mismatched roundtable, ChatGPT–Davinci-003–GPT-4, with a fair roundtable, ChatGPT–Davinci-003–ChatGPT-0301.The debates use e-CARE and PIQA and run for up to 9 rounds.
  • Results: FORD significantly outperforms Col-S and Col-H in both roundtable debates.In the fair R2 setting, FORD also outperforms all single LLMs.
  • Results: The INCON is significantly alleviated, indicating that more than two LLMs can collaborate toward consensus.
  • Results: FORD in R1 surpasses FORD in R2, suggesting that adopting a stronger LLM can improve debate performance despite possible distraction by weaker models.
  • Results: In R2, FORD outperforms ChatGPT–ChatGPT-0301 and performs similarly to ChatGPT–Davinci-003 because similar models provide little new information.

7 Analysis

The analysis examines whether debate order, judging, and debate examples affect FORD’s behavior. Reversing order preserves the broad advantage but reduces performance, while GPT-4 judging further improves results by weighting arguments more precisely.

  • GPT-4 as judge: GPT-4 as judge can further boost FORD performance by assigning higher weights to convincing arguments and drawing more precise conclusions.
  • Debate order: The ablation reverses the step-2 order for ChatGPT–Davinci-003 and ChatGPT–ChatGPT-0301 while keeping other settings unchanged.
  • Debate order: Reversing the ChatGPT–Davinci-003 debate order still leaves FORD above Col-S, Col-H, and the corresponding single LLMs.
  • Debate order: Reversing the debate order reduces performance but speeds INCON convergence, because starting with Davinci-003 might mislead the debate.
  • Debate example: The FORD example shows one LLM supplying an overlooked perspective, producing more convincing interpretability and a more accurate decision.

8 Related Work

Related work covers instruction-tuning and chain-of-thought methods for eliciting reasoning, alongside research on inconsistency within individual LLMs. This paper extends that focus to inter-consistency among multiple collaborating LLMs.

  • Reasoning methods: Instruction-tuning methods such as FLAN, T0, InstructGPT, Flan-PaLM, and Tk-INSTRUCT train or align LLMs for instruction following.
  • Reasoning methods: Chain-of-thought methods use few-shot, zero-shot, automatic, or complex prompting to elicit reasoning abilities.
  • Debate example: The example debate illustrates a disagreement over answers that ends with the opponent agreeing with the proponent’s interpretation.
  • Inconsistency: Prior inconsistency studies examine self-consistency, logical consistency, conceptual consistency, or consistency across prompts within a single LLM.
  • This paper: This paper instead investigates inter-consistency between two or more LLMs and whether debate enables collaboration toward consensus.

9 Conclusion

The paper studies inter-inconsistency among LLMs and introduces FORD to test collaboration through debate. Across fair, mismatched, and roundtable settings, debates can improve performance and consistency, but weaker models may distract stronger ones.

  • Contribution: FORD examines whether multiple LLMs can collaborate to reach consensus on shared tasks by addressing inter-inconsistency.
  • Framework: FORD evaluates three real-world debate scenarios: fair debate, mismatched debate, and roundtable debate.
  • Findings: The findings indicate that debates can improve LLM performance and inter-consistency, while mismatched debates can distract stronger LLMs with weaker participants.

Limitations

The study’s evaluation is limited to commonsense reasoning and should be extended to broader reasoning, generation, and real-world tasks.

  • The experiments should cover additional tasks, including mathematical reasoning and MMLU, rather than only commonsense reasoning.
  • Future work should extend beyond multiple-choice formats to natural-language generation and complex reasoning tasks.
  • Further exploration should incorporate more kinds of LLMs and real-world tasks.
  • The authors identify trading reproducibility for diverse debate behaviors as another topic for discussion.

A.1.1 Experiment Details

The experiments use standardized prompts and debate procedures across several LLM pairings, with GPT-4 also serving as a judge in in-depth analysis.

  • Prompting and settings: The study uses zero-shot prompts for ChatGPT and ChatGPT-0301, with fixed temperature 0.0 and maximum generation length 512.The remaining API hyperparameters use default values.
  • Prompting and settings: The experiments use unified or shared prompts across seven commonsense reasoning datasets, with task-specific prompt formats for StrategyQA.The datasets include αNLI, CommonsenseQA, COPA, e-CARE, Social IQa, PIQA, and StrategyQA.
  • Debate setup: Debate prompts are adapted to model type, with separate formats for chat and text-completion models and pairing-specific prompts.ChatGPT and ChatGPT-0301 share a prompt, while LLaMA and Vicuna share a debate prompt.
  • Debate setup: The mismatched-debate experiments include GPT-4, while the roundtable setting uses ChatGPT, Davinci-003, and GPT-4 as three participants.The roundtable assigns ChatGPT as user1, Davinci-003 as user2, and GPT-4 as user3.
  • Debate setup: FORD uses stance selection, alternating debate, and final agreement or voting to produce the answer.When debaters agree, their agreed answer is used; otherwise, the stances are collected and voted on.
  • Judging and analysis: GPT-4 is used as a judge to summarize debate processes and provide final conclusions in the in-depth analysis.The associated evaluation is reported for debates between ChatGPT and ChatGPT-0301.

G Ablation Study

The ablation materials examine prompt examples, debate-order effects, and dominance in roundtable debates, alongside cases illustrating how model arguments interact.

  • Ablation materials: The ablation study includes few-shot examples for Davinci-003, LLaMA, and Vicuna on αNLI and CommonsenseQA.
  • Ablation materials: The study reports an ablation for reversing debate order in ChatGPT and ChatGPT-003.
  • Ablation materials: The study also reports overall dominance in roundtable debates.
  • Debate examples: The materials include examples of debate exchanges in which models provide competing explanations and challenge one another’s plausibility judgments.
Loading 2305.11595v3…