chartqa

Officical repository for the paper“ChartInsights: Evaluating Multimodal Large Language Models for Low-Level Chart Question Answering”(EMNLP'24)

Vision-language AI for chart question answering using Qwen3-VL with SFT and GRPO training

VisDoTQA is a chart visual reasoning benchmark and dataset for Position, Length, Pattern, and Extract tasks.

This is the official repository for our paper 📄 “In-Depth and In-Breadth: Pre-training Multimodal Language Models Customized for Comprehensive Chart Understanding”

Reproducible framework for document-centric VLM robustness and claim-faithfulness evaluation.

Multimodal tool-use RL on ChartQA: Qwen2.5-VL learns visual tool calls (highlight/mask/box) with two-stage rollout and three generations of reward design (rule / shaped / G-RA + Dr.GRPO) on verl/EasyR1

On-policy distillation for a vision-language model (Qwen3-VL 2B student, 8B teacher, ChartQA): OPD matches full-data SFT with 10% of the questions, plus a token-level analysis of where teacher feedback lands.

Factorized candidate sharing for ChartQA and Raven-style visual reasoning.

Open multimodal evaluation harness for Xiaomi MiMo

Directional scale errors in ChartQA and interference-aware LoRA merging with MIRAGE.

Fine-tuning UniChart (vision encoder-decoder) for chart question answering on ChartQA — parameter-efficient training, relaxed-accuracy evaluation, 3x exact-match gain over base