Chart-RVR: Reinforcement Learning with Verifiable Rewards for Explainable Chart Reasoning

TL;DR

Chart-RVR tackles brittle chart reasoning in LVLMs by coupling Group Relative Policy Optimization with verifiable rewards. It introduces surrogate tasks (chart-type prediction and chart-table reconstruction) and a process-conformity reward to guide reasoning, yielding better robustness and faithfulness of explanations. The 3B Chart-RVR models achieve state-of-the-art results on six chart benchmarks, with pronounced gains on out-of-distribution data and improved rationale quality. The approach demonstrates strong, architecture-agnostic generalization and provides interpretable, auditable chain-of-thought traces for chart reasoning.

Abstract

The capabilities of Large Vision-Language Models (LVLMs) have reached state-of-the-art on many visual reasoning tasks, including chart reasoning, yet they still falter on out-of-distribution (OOD) data, and degrade further when asked to produce their chain-of-thought (CoT) rationales, limiting explainability. We present Chart-RVR, a general framework that fine-tunes LVLMs to be more robust and explainable for chart reasoning by coupling Group Relative Policy Optimization (GRPO) with automatically verifiable rewards. Our framework comprises of three rewards that maximize: (i) correct chart-type classification, (ii) faithful chart table reconstruction, and (iii) process conformity. Applied to 3-billion-parameter LVLMs, Chart-RVR consistently outperforms standard supervised fine-tuning (SFT) on both in-distribution and out-of-distribution datasets, closing the OOD performance gap while improving rationale fidelity. The resulting models, the Chart-RVR-3B series, achieve state-of-the-art results on six chart-reasoning benchmarks spanning in-domain and OOD settings, surpassing all existing models of comparable size. Beyond accuracy, Chart-RVR yields more interpretable CoT rationales, strengthening trust and reliability - showcasing the power of verifiable rewards with GRPO for training reliable, interpretable chart-reasoning models.

Figures (11)

• Figure 1: Chain-of-thought rationales on EvoCharts (OOD). We demonstrate CoT rationales for Structured prompting on base model, SFT model, and Chart-RVR. We highlight the mistake in a particular reasoning step in red font. See Appendix for additional examples.
• Figure 2: System prompt template used across Structured CoT, SFT, GRPO, and Chart-RVR.
• Figure 3: System prompt template used across Structured CoT, SFT, GRPO, and Chart-RVR.
• Figure 4: Chart-RVR Reward maximization during training on 3 separate CoT datasets on Qwen2.5VL-3B.
• Figure 5: Example failure cases from the EvoChart dataset (OOD) where Chart-RVR outperforms CoT and SFT on 2 challenging examples.