Enhancing Multi-Agent Consensus through Third-Party LLM Integration: Analyzing Uncertainty and Mitigating Hallucinations in Large Language Models

TL;DR

A new method that integrates different LLMs to expand the knowledge boundary, reduce dependence on a single model, and promote in-depth debate among agents is proposed, surpassing traditional multi-agent baselines.

Abstract

Large Language Models (LLMs) still face challenges when dealing with complex reasoning tasks, often resulting in hallucinations, which limit the practical application of LLMs. To alleviate this issue, this paper proposes a new method that integrates different LLMs to expand the knowledge boundary, reduce dependence on a single model, and promote in-depth debate among agents. The main contributions include: 1) Introducing third-party LLMs to adjust the attention weights of agents through uncertainty estimation and confidence analysis, optimizing consensus formation in multi-agent systems; 2) Experiments on arithmetic datasets have validated the effectiveness of the method, surpassing traditional multi-agent baselines. This research provides a new perspective for large models to alleviate hallucination phenomena when dealing with complex tasks.

Figures (3)

• Figure 1: As shown in Figure 1(a), we introduce third-party LLMs to enrich the knowledge of multi-agent system, which can provide agents with different perspectives. Agents can adjust their answers based on more information and deeper thinking. The uncertainty and confidence weighting mechanism allows each agent to dynamically adjust its attention weights in the second and third rounds of debate, so that it can self-adjust according to other agents’ feedback and approach the correct answer more accurately. As shown in Figure 1(b), all agents in traditional methods depend on a single LLM, which makes the answers have consistent bias and lack of diversity and depth of thinking.
• Figure 2: The weight values of different agents represent their relative importance or trustworthiness when dealing with specific tasks. The range weights are a list that includes multiple weight items, each with a starting point and an ending point (start and end), as well as a corresponding weight value (weight). For example, when the fourth agent uses a third-party large language model to answer questions, its confidence parameter is set to 10.0, with the corresponding range being 'start': 399, 'end': 562, 'weight': 10.0. Within the range of processing context prompts, the large model will adjust the distribution of attention weights based on these parameters.Subsequently, the model outputs completion and its corresponding uncertainty value through the logits transformation mechanism. The confidence weight=1/uncertainty.For example, weight = 1/0.15331237018108368 = 6.522630879810011. In the second and third rounds of iteration, the agent's response and its confidence weight are used as reference information for other agents in subsequent rounds. Ultimately, a comprehensive response is derived after the second and third rounds of iteration.
• Figure 3: In the first round, each agent answers the question individually. In the subsequent second and third rounds, each agent provides response information and uncertainty probabilities from other agents and uses this information to refine their answers. The first three agents use the Llama3 model, modify model weight information through attention scaling, output the model's predicted results, and calculate the uncertainty measure of this response through token probabilities. The fourth agent uses the ERNIE LLM model, specifying weight coefficients manually as uncertainty probabilities. The final answer is determined by a majority vote after the third round.