Financial AI Research Trends Seen Through ICAIF 2025

Kakao Bank's Technology Research Institute directly witnessed the latest trends in financial AI at ICAIF 2025, and its own research, "Query Generation Pipeline with Enhanced Answerability Assessment for Financial Information Retrieval," was selected for an Oral session. Of 349 total submitted papers, 113 were accepted, and only 54 of those were presented as Oral.

The conference was held at the Sheraton Towers in Singapore, and was divided into workshops and a main conference. The workshops covered topics such as XAI-FIN-2025, AI and Data Science for Digital Finance, and Rethinking Financial Time-Series, while the main conference featured Oral presentations, posters, keynotes, and panel discussions. Notably, the presence of Korean researchers was significant, with 14% of accepted papers coming from Korea.

Kakao Bank's research targeted the bottleneck in evaluating financial information retrieval systems. To reduce the cost and limitations of manually creating question-answer data, the team used an LLM to dynamically generate queries and answers, and incorporated an Answerability Assessment that judges whether a generated question can actually be answered from the document, filtering out low-quality questions. As a result, they proposed a pipeline that can more efficiently build high-quality evaluation datasets close to those made by hand.

Three trends stood out most prominently at the conference.

• RL and Agent Systems: Market simulation, trading, fraud detection, and autonomous agent-based strategies expanded.
• Investment Strategy and Market Forecasting: Decision-Focused Learning, which directly optimizes final decisions rather than prediction accuracy, emerged, and research on volatility, time series, and risk management also became more sophisticated.
• Evaluation, Ethics and Safety: Rather than how well LLMs perform in finance, how to evaluate, explain, and ensure safety became a more important topic.

In reinforcement learning and agent research, attempts to view markets as complex interactive systems stood out. There was research on interpretable market simulation, high-performance financial digital twins, and RL applied to asset-liability management and insurance pricing, along with warnings that autonomous trading agents could lead to market manipulation. It was emphasized that as agent capabilities grow, regulation, governance, and ethical verification are needed alongside them.

In the investment strategy and forecasting field, approaches that directly optimize portfolio performance rather than simple prediction stood out. Research applying Decision-Focused Learning to covariance estimation and return prediction, research calibrating stochastic local volatility models with Physics-Informed Neural Networks, and research repurposing LLMs for foreign exchange volatility prediction were introduced. Time-series modeling also evolved to be more relationship-centered, such as learning Lead-Lag relationships between assets or combining physics-based mean reversion concepts with contrastive learning for pairs trading. In risk management, approaches such as Conformal Prediction, which quantifies the uncertainty of predictions, became important.

On the evaluation and reliability side, efforts focused on measuring the limitations of financial LLMs. Key topics included multimodal benchmarks that jointly examine text, tables, and charts, hallucination evaluation on tabular data, and semantic drift analysis according to economic regime changes. In the explainability session, case-based explanations such as prototypes, criticisms, counterfactuals, and semi-factuals drew attention, and interpretable variant models such as NeuralBeta were also introduced.

The most active discussion was on Ethics and Bias in LLM-driven Finance. Research continued on bias in investment analysis, ethical judgment in market abuse situations, and verifying regulatory compliance capability, and Kakao Bank's query generation research was also presented here. The core message was that rather than simply producing good answers, creating questions that can actually be answered in the first place and checking for potential failure beforehand is the starting point of reliability in financial AI.