SD-MAR: Multi-image Analytical Reasoning via Synthetic Data and Reinforcement Learning

2026-07-15 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision, Natural Language Processing · Depth: Expert, quick

Summary

The SD-MAR (Synthetic Data for Multi-image Analytical Reasoning) framework addresses limitations of Vision Language Models (VLMs) in tasks requiring analytical reasoning across multiple visual states, such as multi-image comparison and change detection. Existing benchmarks often overlook the need for explicit visual comparison combined with analytical reasoning. SD-MAR tackles this by constructing paired visual scenarios through controlled perturbations and generating reasoning tasks focused on semantic change attribution and quantitative comparison. The framework employs GRPO-lite with Backward Discounted Allocation (BDA), a reinforcement learning method that removes KL regularization to enhance policy optimization and credit allocation for later reasoning steps. Experiments with Qwen2.5-VL-7B and InternVL3-8B demonstrated in-domain accuracy improvements of up to 36.95%. Notably, Qwen2.5-VL-7B surpassed GPT-4.1 on the SD-MAR benchmark. Out-of-domain generalization was preserved or improved, with performance remaining within 1% on MME, MMMU-Pro, and MathVista, and improving by up to 4% on MMBench. LLM-as-judge evaluations confirmed enhanced logical coherence and explanation quality.

Key takeaway

For Machine Learning Engineers developing Vision Language Models for complex multi-image analytical reasoning, you should consider integrating synthetic data generation and GRPO-lite fine-tuning. This approach, demonstrated by SD-MAR, significantly improves in-domain accuracy by up to 36.95% and preserves or enhances out-of-domain generalization. By focusing on controlled visual perturbations and targeted reinforcement learning, you can achieve superior logical coherence and explanation quality, potentially outperforming models like GPT-4.1 on specific reasoning benchmarks.

Key insights

Synthetic data and reinforcement learning can significantly enhance Vision Language Models' multi-image analytical reasoning capabilities.

Principles

• Analytical reasoning in VLMs benefits from explicit visual comparison.
• Controlled perturbations generate effective synthetic data for reasoning tasks.
• GRPO-lite with BDA improves policy optimization for complex reasoning.

Method

SD-MAR constructs paired visual scenarios via controlled perturbations, generating semantic change attribution and quantitative comparison tasks. It fine-unes VLMs using GRPO-lite with Backward Discounted Allocation (BDA), removing KL regularization and crediting later reasoning steps.

In practice

• Generate synthetic visual data with controlled perturbations for VLM training.
• Employ GRPO-lite with BDA for fine-tuning VLMs on reasoning tasks.
• Use LLM-as-judge for evaluating VLM logical coherence and explanation quality.

Topics

• Vision Language Models
• Multi-image Reasoning
• Synthetic Data
• Reinforcement Learning
• GRPO-lite
• VLM Benchmarking

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• SD-MAR: Multi-image Analytical Reasoning via Synthetic Data and Reinforcement Learning — Synthetic data and reinforcement learning significantly enhance Vision Language Models' multi-image analytical reasoning capabilities.
• Stop Thinking, Start Looking: Efficient Post-Training for Multimodal Document Question Answering via Reasoning-Free Alignment — Direct perception-based policies outperform reasoning-centric RL for efficient multimodal document question answering.
• Be Consistent! Enhancing Robust Visual Reasoning in LVLMs with Consistency Constraints — ConVBench and ConVLM improve LVLM visual reasoning robustness and logical consistency through a new benchmark and a GRPO-based reinforcement learning method.
• Reinforcing the Generation Order of Multimodal Masked Diffusion Models — A GRPO-trained control module optimizes generation order in multimodal diffusion models, improving alignment and understanding.
• Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning — The generation pathway of visual thoughts encodes reasoning that can be distilled for efficient, text-only multimodal models.
• Phi-4-reasoning-vision and the lessons of training a multimodal reasoning model — Efficient multimodal reasoning models can achieve high performance with less data and compute through careful design and data curation.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.