Research map / Vision-language and multimodal / Multimodal LLMs

Chart Deep Research in LVLMs via Parallel Relative Policy Optimization

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 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 PRPO, a GRPO extension that optimizes separately across reward dimensions and capability-based data partitions, and MCDR-Bench, a 3,084-sample chart deep-research benchmark built with controlled error injection.

Problem

Chart understanding models handle shallow tasks (recognition, factual QA) but not deep analytical research; post-training suffers from multi-dimensional reward interference and heterogeneous data gradient conflicts, and existing evaluations cannot measure end-to-end analytic reasoning.

Approach

Parallel relative advantage computation: GRPO advantages are computed separately per reward dimension and per capability-based data partition (with outlier samples relegated to rollout-level grouping), and combined with weighted sums; evaluation converts open-ended report generation into error identification via controlled error injection under the "error uniqueness principle".

Key results

• On MCDR-Bench, PRPO with Think raises Qwen2.5-VL-7B mean score to 76.89% versus 63.99% for GRPO Think.
• On ChartQAPro, PRPO reaches 47.69% mean versus 41.33% for the Qwen2.5-VL-7B baseline.
• Ablation on MCDR-Bench: GRPO 61.71%, Reward-PRPO 64.30%, Data-PRPO 63.55%, full PRPO 69.62% (direct answer setting).

Surprising findings

• Simple aggregation of data levels performs poorly: High Level data alone scores 18.78% and Level5 only 32.30%, below several single-level subsets.

Limitations

• Data-PRPO partition threshold and reward/data weighting hyperparameters (lambda_k, lambda_m) are described only loosely, with default mean aggregation.
• Reward-PRPO ablation is on geometry3k math data rather than chart data.
• MCDR-Bench error injection is designed and verified by humans and GPT-4o-based annotation, with limited details on scale of human review.

Experimental setup

Datasets: SophiaVL-R1-130k R1-Onevision geometry3k MCDR-Bench ChartQAPro

Benchmarks: MCDR-Bench ChartQAPro geometry3k

Metrics: accuracy (error identification score) mean score across capability splits response length

Baselines: GRPO GPT-4o GPT-4.1 Claude-3.7 Sonnet Gemini-2.5-Pro ChartReasoner Claude-3.5 Sonnet Gemini-Flash-2.0 InternVL2.5-8B LLaVA-Next-7B

Builds on: GRPO DAPO GSPO PPO DPO

Closest papers by meaning

• SketchVL: Policy Optimization via Fine-Grained Credit Assignment for Chart Understanding and More (CVPR 2026 · similarity 0.86)
• CharXiv: Charting Gaps in Realistic Chart Understanding in Multimodal LLMs (NeurIPS 2024 · similarity 0.85)
• Breaking the SFT Plateau: Multimodal Structured Reinforcement Learning for Chart-to-Code Generation (ICLR 2026 · similarity 0.85)
• Chart-FR1: Visual Focus-Driven Fine-Grained Reasoning on Dense Charts (CVPR 2026 · similarity 0.84)
• ChartNet: A Million-Scale, High-Quality Multimodal Dataset for Robust Chart Understanding (CVPR 2026 · similarity 0.84)
• EvoChart: A Benchmark and a Self-Training Approach Towards Real-World Chart Understanding (AAAI 2025 · similarity 0.84)