Скачать или смотреть A Hashgraph-Inspired Consensus Mechanism for Reliable Multi-Model Reasoning

Large language models (LLMs) and other reasoning models (RMs) often struggle with giving consistently reliable outputs, frequently producing incorrect information or "hallucinations," especially when different models are asked the same complex questions. To address this, this paper introduces a novel *consensus mechanism* inspired by *Hashgraph**, a type of distributed ledger technology. The idea is to treat each RM as a peer in a network where they can share their outputs with each other using a process similar to **gossip**. Through this exchange, models receive information about what their peers are answering, and then they use a concept called **virtual voting* to iteratively update and refine their own answers. This multi-round process encourages correct information, which is likely to be present in multiple models, to spread and be adopted, while unsupported errors or hallucinations are likely to be identified and dropped, leading the ensemble of models to converge on a single, more accurate, and reliable answer. This method aims to provide strong guarantees of consistency and fault tolerance, meaning it can help filter out mistakes even if some models are providing misleading content.

https://arxiv.org/pdf/2505.03553