FinAcumen: Financial Multimodal Reasoning via Self-Evolving Experience Memory Harness
Researchers have introduced FinAcumen, a novel framework designed to enhance financial multimodal reasoning capabilities in AI agents. This system addresses the limitations of current stateless agents by incorporating a selective experience memory that distills successful strategies and cautionary rules from prior reasoning trajectories. By retrieving relevant experiences based on semantic relevance and suppressing irrelevant ones, FinAcumen aims to improve tool routing, reduce retrieval noise, and mitigate hallucination in high-stakes financial scenarios. Evaluations on four financial multimodal reasoning benchmarks show that FinAcumen significantly enhances a base 8B vision-language model, approaching the performance of leading proprietary general-purpose models. AI
IMPACT Enhances AI agent reliability in complex financial reasoning tasks by incorporating memory of past experiences.