MG-Thinker: Bi-Axial Self-Reflection for Multi-Image Reasoning Grounding

Heyu Huang; Chi Chen; Zonghao Guo; Yuhua Li; Maosong Sun; Ruixuan Li

cs.CVSep 29, 2026

MG-Thinker: Bi-Axial Self-Reflection for Multi-Image Reasoning Grounding

Authors: Heyu Huang, Chi Chen, Zonghao Guo, Yuhua Li, Maosong Sun, Ruixuan Li

Organizations: Huazhong University of Science and Technology Wuhan, China · Tsinghua University Beijing, China

Abstract

Reinforcement learning (RL) has recently delivered substantial gains in multimodal reasoning, opening a promising route for fine-grained visual perception. Yet for multi-image reasoning grounding (MRG), reasoning over real-world multi-image contexts toward pixel-precise localization, existing RL-based approaches overlook two characteristics intrinsic to this paradigm: a coarse-to-fine hierarchical reasoning pattern, and heterogeneously distributed task--sample difficulties. In this work, we present MG-Thinker, a post-training RL framework that advances a new MRG paradigm featuring such hierarchical reasoning, supported by a curated 25K MRG dataset with task-adaptive Chain-of-Thought (CoT) annotations that elicit multi-perspective evidence before conclusion. To remedy the heterogeneous task--sample difficulties, we further propose Bi-Axial DAPO (BiA-DAPO), which decomposes rollout advantages along an intra-group signal axis and an inter-group competence axis through two complementary mechanisms, both grounded on our defined candidate pool for stable group-level statistics. Extensive experiments show that MG-Thinker achieves state-of-the-art performance on multi-image reasoning grounding while consistently improving generalization across multi-image understanding and diverse multimodal benchmarks.

Figures & tables

Figure 1. Illustrative case showing that multi-image reasoning grounding demands a coarse-to-fine hierarchical pattern (CoT $\to$ image-id $\to$ bbox), absent in representative baselines, which MG-Thinker explicitly delivers through one-shot multi-image multi-object localization.

Figure 2. ( Left ) Task–sample difficulty heterogeneity profile of MRG vs. general tasks: per-task mean reward (bars) and within-sample rollout-std (lines) over the base model (Qwen2.5-VL) and the CoT-SFT model; S-Grd, S-Gen, M-Gen denote single-image grounding, single-image and multi-image understanding tasks, respectively. ( Right ) Per-task accuracy radar over the MIG-Bench.

Figure 3. Illustration of the multi-image grounding data pre-processing pipeline, including rule-based criteria and processing operations derived using the multimodal large language model Qwen2.5-VL-72B.

Figure 4. Overview of the proposed Bi-Axial DAPO (BiA-DAPO) for MRG. Upper: GIA scores each group along the intra-group signal axis to relieve drifting advantage-signal sparsity. Lower: CRS schedules updates over the GIA-retained groups along the inter-group competence axis to remedy competence-difficulty misalignment. Both operate on a shared candidate pool for stable group-level calculations.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7. Restructured multi-image grounding data, organized by intrinsic task heterogeneity into Visual-Heterogeneity (V-Het), Reasoning-Heterogeneity (R-Het), and Low-Heterogeneity (L-Het) groups. Each group spans both single-object and one-shot multi-object grounding, reflecting the distinct challenges of MRG.

Figure 8. CoT templates designed for the four MRG-specific task types and the standard response format.

*

denotes textual content and

\#

denotes natural numbers; in the generated CoT templates,

*

is replaced with task-specific descriptive terms.

Figure 11. Continuous-variation sensitivity of BiA-DAPO on MIG-Bench: GIA threshold $\alpha_{0}$ , candidate-pool size, and number of GIA phases.

Figure 12. Comparison of non-reasoning and various reasoning paradigms on a representative cross-image reasoning grounding task: direct prediction, CoT-prompting, MG-Thinker cold-start, and MG-Thinker fully trained with BiA-DAPO post-training.

Figure 13. Visual analysis on MIG-Bench correspondence grounding tasks; ground-truth bounding-box annotations are shown in white to distinguish them from predictions.

Figure 14. Visual analysis on MIG-Bench reasoning grounding tasks: comparison of MG-Thinker against representative multi-image baselines.

Figure 15. Visual analysis on MIG-Bench region grounding tasks under repeated distractors.

Figure 16. Visual analysis on MIG-Bench object-tracking grounding tasks; ground-truth annotations are shown in white. Numerical values follow each model’s coordinate convention (relative or absolute).

Explore similar work

Mags-RL: Wearing Multimodal LLMs a Magnifying Glass via Agentic Reinforcement Learning For Complex Scene Reasoning
May 27, 2026Xuanzhao Dong, Wenhui Zhu, Peijie Qiu +11Multimodal Large Language Models Recent Vision-Language Models
Arizona State University, Tempe, AZ, USA · Washington University in St. Louis, St. Louis, MO, USA · Clemson University, Clemson, SC, USA +3
Mimic Human Cognition, Master Multi-Image Reasoning: A Meta-Action Framework for Enhanced Visual Understanding
Jan 12, 2026Jianghao Yin, Qingbin Li, Kun Sun +10Recent Vision-Language Models Theory-Of-Mind Reasoning
East China Normal University · ByteDance
Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention
May 21, 2026Changyuan Tian, Zhicong Lu, Huaxing Liu +7Multimodal Reasoning Recent Vision-Language Models
AMAP, Alibaba Group · University of Chinese Academy of Sciences · Tsinghua University +1

May 27, 2026, cs.CVJ/K move · Enter open · S save

Mags-RL: Wearing Multimodal LLMs a Magnifying Glass via Agentic Reinforcement Learning For Complex Scene Reasoning

Xuanzhao Dong, Wenhui Zhu, Peijie Qiu, Xiwen Chen+10

Arizona State University, Tempe, AZ, USA · Washington University in St. Louis, St. Louis, MO, USA · Clemson University, Clemson, SC, USA+3

Multimodal Large Language Models Recent Vision-Language Models Scene Reasoning Icr-Rl Agentic Reinforcement Learning Generative Super-Resolution

Open paper AskPDF arXiv

Despite their popularity and success, Multimodal Large Language Models (MLLMs) often struggle to interpret images accurately, which limits their reasoning capability in complex scenarios (e.g., high object density and complex background clutter). Prior work mainly addresses this limitation by incorporating explicit visual cues like bounding boxes that require extra annotations. In addition, the resulting low-resolution crops often miss fine-grained details that MLLMs require for accurate reasoning. Therefore, we propose Mags-RL, an Agentic Reinforcement Learning (RL) framework that equips MLLMs with an external super-resolution "magnifying glass" agent for high-resolution fine-grained inspection. Specifically, the model performs two-round reasoning: in the first round, it generates an initial rationale and autonomously identifies regions of interest without relying on additional annotations; in the second round, it invokes a super-resolution agent to crop and upscale those regions, then revisits and verifies its earlier reasoning to produce the final answer. We also introduce a novel curriculum learning strategy that enables data-efficient RL training, needing as few as only 40 training samples to achieve reasonable performance. Experiments on VSR, TallyQA, and GQA subsets show its superior performance against recent strong competing methods, demonstrating high-quality reasoning with precise visual grounding. Code and weights will be released soon.

Abstract

Explore similar work

Mags-RL: Wearing Multimodal LLMs a Magnifying Glass via Agentic Reinforcement Learning For Complex Scene Reasoning

Mimic Human Cognition, Master Multi-Image Reasoning: A Meta-Action Framework for Enhanced Visual Understanding

Faithful-MR1: Faithful Multimodal Reasoning via Anchoring and Reinforcing Visual Attention