Research map / Vision-language and multimodal / Multimodal LLMs

SketchVL: Policy Optimization via Fine-Grained Credit Assignment for Chart Understanding and More

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SketchVL, a reasoning-on-image MLLM trained with FinePO, an RL method that redistributes GRPO trajectory advantages across drawing steps using a process reward model (FinePRM).

Problem

Trajectory-level advantage estimation in RL (e.g., GRPO) broadcasts a single reward to all tokens, so it cannot distinguish correct from incorrect reasoning steps in multi-step chart understanding.

Approach

Intra-trajectory credit assignment: a FinePRM scores each drawing action (intent plus before/after images), and the group-relative advantage is redistributed across steps via a length-weighted deviation from the response mean, with KL-based action-type regularization and clipping to conserve total advantage.

Key results

• SketchVL-3B improves over Qwen2.5VL-3B by 15.32 points on ChartQA and 3.76 points on ChartBench.
• SketchVL-7B reaches 83.96 on ChartQA vs 82.00 for Qwen2.5VL-7B and 63.50 on MathVista vs 61.40.
• Ablation: replacing FinePO with naive GRPO on SketchVL-3B drops ChartQA from 77.20 to 75.12 and PlotQA from 48.32 to 44.72.

Surprising findings

• Removing the KL action regularization yields near-identical or slightly higher scores on several benchmarks, and the paper attributes its benefit mainly to preventing action-type collapse rather than accuracy.
• The 3B model gains more from FinePO than the 7B model, which the authors hypothesize relates to reduced capacity to exploit the FinePRM.

Limitations

• No standardized benchmark exists for quantitatively evaluating the FinePRM.
• Constructing an unbiased process reward model is inherently difficult; reward hacking risk noted.
• KL action regularization can slightly constrain performance on benchmarks dominated by one action type.

Experimental setup

Datasets: EvoChart GQA ChartQA-Train OpenImages Vision-R1-RL ChartBench-Train VisualCoT-Train

Benchmarks: EvoChart-QA ChartQA ChartQA-Pro ChartBench PlotQA MathVista MMStar

Metrics: accuracy (judge-based correctness) average FinePRM process score

Baselines: VLM-R1 ChartSketcher-2B Qwen2.5VL-7B Qwen2.5VL-3B naive GRPO

Builds on: GRPO VisualPRM Set-of-Mark DeepEyes ChartSketcher

Closest papers by meaning

• Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.90)
• ChartSketcher: Reasoning with Multimodal Feedback and Reflection for Chart Understanding (NeurIPS 2025 · similarity 0.90)
• Chart-FR1: Visual Focus-Driven Fine-Grained Reasoning on Dense Charts (CVPR 2026 · similarity 0.89)
• RewardMap: Tackling Sparse Rewards in Fine-grained Visual Reasoning via Multi-Stage Reinforcement Learning (ICLR 2026 · similarity 0.88)
• ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding (CVPR 2026 · similarity 0.87)
• GoT-R1: Unleashing Reasoning Capability of Autoregressive Visual Generation with Reinforcement Learning (ICLR 2026 · similarity 0.87)