FinSumm: Towards Analyst-Level Financial Chart Summarization with Human-Aligned Vision-Language Model
Abstract
Financial chart summarization requires numerical grounding and domain-specific reasoning, yet existing data and general-purpose VLMs provide limited support for this task. We introduce FinSumm-8B, an analyst-oriented VLM trained by supervised fine-tuning and fact-calibrated preference optimization (FCPO), which uses source-data verification to calibrate heterogeneous preference pairs. We also construct FinChartSum, a 1K-sample benchmark covering real and synthesized charts. On a balanced 200-sample cross-model evaluation, FCPO improves the base Qwen3-VL-8B LLM-Score from 0.675 to 0.762 and exceeds the 30B model (0.726), while GPT-5.4 (0.858) and Gemini-3.1-Pro (0.844) remain stronger under judge-based quality evaluation. These results support domain-specific post-training for efficient financial chart summarization while exposing remaining gaps on source-free real charts.