MINT Enables Zero-Shot Prediction for Transaction Data

This paper presents MINT, the Multimodal Instruction Network for Transactions, a novel framework designed for universal zero-shot prediction on sequential financial transaction data. Banks typically analyze this data for tasks like fraud prevention and credit risk, often using Payments Foundation Models to generate contextual embeddings. However, these models lack the flexibility for zero-shot reasoning on new, unseen tasks. MINT addresses this limitation by integrating a pretrained transaction sequence encoder with a decoder-only Large Language Model (LLM). This connection is achieved through lightweight embedding injection, transaction-language alignment, and instruction tuning. The framework significantly improves predictive question-answering performance for both in-distribution and out-of-distribution queries. Crucially, MINT achieves these results while substantially reducing input tokens, latency, and memory consumption compared to traditional text-serialization baselines, establishing compact transaction embeddings as a superior approach for multimodal reasoning and zero-shot prediction.