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

2026-07-15 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

SD-MAR (Synthetic Data for Multi-image Analytical Reasoning) is a new framework designed to train and evaluate Vision Language Models (VLMs) on complex analytical reasoning tasks involving multiple visual states. Addressing current VLM limitations in multi-image comparison, change detection, and multi-step visual inference, SD-MAR generates synthetic data by creating paired visual scenarios with controlled perturbations and corresponding reasoning tasks. The framework utilizes GRPO-lite with Backward Discounted Allocation (BDA), a reinforcement learning method, for fine-tuning. Experiments on Qwen2.5-VL-7B and InternVL3-8B demonstrated substantial improvements, with in-domain accuracy increasing by up to 36.95%. Notably, Qwen2.5-VL-7B outperformed GPT-4.1 on the SD-MAR benchmark. Out-of-domain generalization was maintained or enhanced, showing performance within 1% on MME, MMMU-Pro, and MathVista, and up to 4% improvement on MMBench. Furthermore, LLM-as-judge evaluations confirmed improved logical coherence and explanation quality.

Key takeaway

For Machine Learning Engineers developing Vision Language Models for multi-image analytical tasks, you should consider integrating synthetic data generation and reinforcement learning fine-tuning. Implementing the SD-MAR framework, particularly GRPO-lite with Backward Discounted Allocation, can significantly improve in-domain accuracy and maintain or enhance out-of-domain generalization, as demonstrated by Qwen2.5-VL-7B surpassing GPT-4.1 on specific benchmarks. This approach offers a clear path to building more robust and logically coherent VLMs.

Key insights

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

Principles

• VLMs need explicit multi-image comparison.
• Synthetic data bridges reasoning gaps.
• RL fine-tuning boosts analytical conclusions.

Method

SD-MAR constructs paired visual scenarios through controlled perturbations, generating reasoning tasks. It fine-tunes VLMs using GRPO-lite with Backward Discounted Allocation (BDA), removing KL regularization and crediting later reasoning steps.

In practice

• Fine-tune VLMs with GRPO-lite.
• Generate synthetic data via perturbations.
• Employ LLM-as-judge for quality.

Topics

• Vision Language Models
• Multi-image Reasoning
• Synthetic Data
• Reinforcement Learning
• GRPO-lite
• Model Fine-tuning

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 can significantly enhance Vision Language Models' multi-image analytical reasoning capabilities.
• Decompose, Look, and Reason: Reinforced Latent Reasoning for VLMs — DLR enhances VLM visual reasoning by dynamically decomposing queries and extracting premise-conditioned continuous visual latents.
• Be Consistent! Enhancing Robust Visual Reasoning in LVLMs with Consistency Constraints — LVLMs can achieve robust, consistent visual reasoning by training with logically equivalent question pairs and a dual accuracy-consistency reward.
• Structured Role-Aware Policy Optimization for Multimodal Reasoning — SRPO improves LVLM multimodal reasoning by assigning role-aware, token-level credit for perception and reasoning.
• Hint-Guided Diversified Policy Optimization for LLM Reasoning — Diversifying solution exploration and explicitly incentivizing reliability significantly improves LLM reasoning performance.
• DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models — DACA-GRPO improves diffusion LLM reinforcement learning by assigning temporal credit and reducing likelihood estimation bias.

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 Takara TLDR - Daily AI Papers.