The field of language models is undergoing a significant shift with the emergence of end-to-end audio language models that process speech directly, preserving detailed information such as intonation and speaker identity. However, this shift also introduces new safety risks and highlights the need for responsible deployment of these models, guided by the principle of least privilege.

Recent developments have focused on establishing best practices for evaluating and training language models in financial applications, including the creation of open leaderboards and benchmarks for assessing model performance. There is also a growing interest in evaluating the capabilities of large language models in financial decision making, particularly in fund investment, using simulated live environments and forward testing methodologies.

Furthermore, the lack of benchmarks for assessing audio large language models in financial scenarios has been addressed with the introduction of new benchmarks, such as FinAudio, which evaluates the capacity of audio large language models in the financial domain.

Noteworthy papers include:

• DeepFund, which introduces a comprehensive platform for evaluating LLM-based trading strategies in a simulated live environment, providing a more accurate and fair assessment of LLMs capabilities in fund investment.
• FinAudio, which presents the first benchmark designed to evaluate the capacity of AudioLLMs in the financial domain, revealing the limitations of existing AudioLLMs and offering insights for improvement.