A multi-modal, multi-agent framework in which teams of fine-tuned large-language-model expert agents jointly manage a large-cap cryptocurrency portfolio, each agent specialising in one data modality (price/on-chain factors, news sentiment, or candlestick charts). Its distinctive contribution is converting each agent's binary rise-or-fall trend classification into a continuous confidence score read from the model's next-token probabilities, then combining agents through confidence-weighted intrateam ensembling and interteam memory sharing to build an interpretable portfolio whose every decision is traceable to explicit agent reasoning.
An explainable multi-modal, multi-agent framework built on large language models for cryptocurrency portfolio management across the large-cap coins. Rather than a single opaque deep-learning predicto…