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

• Existing MLLMs rely on direct prediction and face grounding drift, unstable formats, and weak cross-image alignment; single-image CoT can misroute attention [11] [14] [36] [48].Source: §1 · p. 1
• MRG requires coarse-to-fine reasoning from semantics through image routing to localization, alongside methods that address heterogeneous task–sample difficulties.Source: §1 · pp. 1–2, Figure 1 · p. 1
• MG-Thinker restructures MGrounding-630K into a 25K MRG-specific CoT dataset with task-adaptive prompts that elicit multi-image evidence before localization.Source: §1 · p. 2
• BiA-DAPO combines variance-based rollout-group selection with mean-reward competence scheduling, using a shared candidate pool for stable group-level statistics.Source: §1 · p. 2
• MG-Thinker achieves state-of-the-art multi-image reasoning grounding and consistently generalizes across multi-image understanding, single-image grounding, and diverse multimodal benchmarks.Source: §1 · p. 2

Visual Grounding

• Visual grounding has progressed from single-image referring-expression localization [28, 54] to reasoning-intensive instruction grounding and free-form multi-image grounding.Source: §2 · p. 2

Reinforcement Learning-based Reasoning

• RL-based multimodal reasoning commonly uses explicit CoT data for SFT and GRPO with rule-based rewards; VL-Rethinker [40] replays high-value samples, while AdaRFT [35] schedules difficulty.Source: §2 · p. 2

• Figure 3 illustrates preprocessing of multi-image grounding data using rule-based criteria and operations derived with Qwen2.5-VL-72B.Source: Figure 3 · p. 3

3.1 Overview

• MRG produces a reasoning trace and image-indexed grounded predictions for an image set, prompt, and optional reference.Spontaneous grounding relies on the prompt alone; referential grounding also uses a textual mention or visual exemplar.Source: §3.1 · p. 3

3.2 Data Curation and MRG-Specific CoT Activation

• The data pipeline restructures MGrounding-630K and distills task-adaptive CoT demonstrations into a 25K cold-start set for reasoning-grounding RL.It creates 480K refined samples, uses 320K for stage-one SFT, and filters CoT traces for rationale quality and IoU improvement before selecting the 25K cold-start samples.Source: §3.2 · p. 3

3.3 BiA-DAPO for Reasoning Enhancement

• BiA-DAPO pairs variance-based GIA filtering with mean-reward CRS scheduling to address drifting signal sparsity and competence mismatch using stable shared-pool statistics.GIA’s variance threshold decays during training; CRS orders retained groups from competence-mature to competence-developing, without external difficulty labels or generation-batch enlargement.Source: §3.3 · pp. 3–4

3.4 Reward Modeling and Training Objective

• The MRG reward combines format compliance with accuracy checks that validate image identity before measuring bounding-box IoU.For multi-image multi-object inputs, image-ID and IoU rewards are computed per target and summed sequentially; IoU uses high and low thresholds with linear interpolation.Source: §3.4 · p. 5

4.2 Main Experimental Results

• 76.71 average on MIG-Bench versus 74.22 for the strongest baseline; MG-Thinker also exceeds Migician by 13.2 points and UniVG-R1 by 4.1 points.The 7B model exceeds Qwen2.5-VL-72B by 30.5 points on average.Source: Table 1 · p. 6, §4.2 · p. 6
• 82.21 one-shot multi-object accuracy versus 81.88 with polling on average; MG-Thinker also exceeds polling on Common and Region.Other approaches degrade markedly in the one-shot setting.Source: Table 2 · p. 6, §4.2 · p. 7
• MG-Thinker transfers zero-shot to single-image and video reasoning-grounding benchmarks, substantially surpassing Migician and matching UniVG-R1 at 59.02% average.The authors report that ReVOS and ReasonVOS gains arise without video-specific supervision.Source: Table 3 · p. 6, §4.2 · p. 7
• Under identical training, BiA-DAPO maintains a smooth upward reward trajectory while GRPO/DAPO oscillate and intermittently regress.BiA-DAPO retains medium/high-informativeness groups, whereas GRPO/DAPO over-retain groups with zero advantage.Source: Figure 5 · p. 7, §4.2 · p. 7, Figure 6 · p. 7
• Across heterogeneity groups, UniVG-R1† matches BiA-DAPO on V-Het but lags on R-Het, while AdaRFT† stays close on R-Het but loses on V-Het.BiA-DAPO leads on both axes simultaneously.Source: Table 4 · p. 7, §4.2 · p. 7
• MG-Thinker delivers strong results across multi-image understanding suites and remains competitive on standard RefCOCO/+/g single-image grounding.The paper reports a lead on MuirBench counting and visual-grounding subtasks.Source: Table 5 · p. 7, §4.2 · p. 8

4.3 Ablation Studies

• Stage-1 SFT exceeds Qwen2/2.5-VL-7B across diverse subtasks, with over 2× average-accuracy gain on MIG-Bench; CoT-SFT and then BiA-DAPO further improve results.The stages progressively activate grounding, integrate reasoning and grounding, and enhance reasoning.Source: Table 6 · p. 8, §4.3 · p. 8
• Removing GIA hurts tasks with wide intra-group spread, while removing CRS hurts tasks with wide between-group competence spread; combining both recovers the optimum.The affected tasks include OT and Co-Re for GIA, and Robust, MV, and GG for CRS.Source: Table 6 · p. 8, §4.3 · p. 8

5 Conclusion

• MG-Thinker targets MRG’s coarse-to-fine reasoning pattern and task-sample difficulty heterogeneity with a 25K MRG-specific CoT dataset and BiA-DAPO.BiA-DAPO separates intra-group signal and inter-group competence using GIA and CRS over a shared candidate pool.Source: §5 · p. 8
• Experiments confirm state-of-the-art performance on MRG and consistent generalization to broader multimodal benchmarks.Source: §5 · p. 8
• The paper proposes the bi-axial view as a potentially useful design pattern for other hierarchical, difficulty-heterogeneous multimodal reasoning tasks.Source: §5 · p. 8

A Multi-Image Grounding Tasks Categorization

• The ten grounding subtasks are regrouped by visual variation and reasoning depth into V-Het, R-Het, and L-Het categories.These categories correspond respectively to mean-reward drift, variance drift, and relative stability in both statistics.Source: A Multi-Image Grounding Tasks… · p. 11

B CoT Prompt Templates for MRG-Specific Activation

• Task-adaptive templates guide cross-image cue interpretation through a coarse-to-fine CoT→image-id→bbox trajectory and support one-shot multi-image multi-target localization.They use four MRG-specific task types and structured reasoning and answer blocks.Source: B CoT Prompt Templates for… · p. 11, Figure 7 · p. 12

C DAPO Clipped Surrogate Objective and Group-Relative Advantage

• BiA-DAPO retains the DAPO clipped-surrogate form and reuses within-group mean reward and standard deviation as competence- and signal-axis statistics.The within-group accuracy reward is defined as the sum of image-id and IoU rewards, with a numerical guard ε=10^-6.Source: C DAPO Clipped Surrogate Objective and… · p. 12

D Empirical Dynamics of GIA and CRS

• Across training, BiA-DAPO migrates group-level reward-mean and standard-deviation density toward high competence with moderate spread; GRPO/DAPO accumulate low-competence mass.The low-competence mass appears both at high spread and near-zero spread.Source: D Empirical Dynamics of GIA and CRS · p. 12
• BiA-DAPO progressively shrinks failed per-step IoU outcomes and grows saturated outcomes, while GRPO/DAPO retain a large failed share.The failed share under GRPO/DAPO reflects groups with near-zero variance and collapsed advantage signals.Source: D Empirical Dynamics of GIA and CRS · p. 12, Figure 9 · p. 13
• The two diagnostics support GIA and CRS acting on complementary statistical axes rather than redundantly.This is consistent with the cross-task heterogeneity analysis in Tab. 4.Source: D Empirical Dynamics of GIA and CRS · pp. 12–13

E RL Post-Training Configuration and Hyperparameters

• Figure 10 depicts reward-density migration from low competence and high spread toward high competence and moderate spread during joint GIA and CRS intervention.Figures 11–13 cover hyperparameter sensitivity, reasoning-paradigm comparisons, and visual correspondence-grounding analysis.Source: Figure 10 · p. 13, Figure 11 · p. 13, Figure 12 · p. 15
• RL samples multiple responses per prompt and optimizes a DAPO-style clipped surrogate with a GRPO-style group-relative advantage estimator and explicit low-variance KL loss.The implementation uses VeRL.Source: E RL Post-Training Configuration and… · pp. 13–14
• GIA uses a four-phase, phase-decayed gate informed by four observed rollout-spread regimes, starting at α0=0.05 and halving the threshold at each phase.The regimes are zero-, low-, medium-, and high-spread groups; the gate progressively admits lower-spread groups as training stabilizes.Source: E RL Post-Training Configuration and… · pp. 14–15
• CRS uses a candidate pool three times the training batch to maintain a consistent group-reward ranking budget across updates.The pool size matches DAPO’s generation batch.Source: E RL Post-Training Configuration and… · pp. 14–15

F Hyperparameter Sensitivity Analysis

• All configurations use Poll output format: its default reads 76.36, while Table 1 reports 76.71 for MG-Thinker’s better-of-format result.The two values reflect different output-format reporting conditions.Source: F Hyperparameter Sensitivity Analysis · p. 15
• 76.36/76.15 (−0.21)/75.57 (−0.79) are the average accuracies for GIA thresholds 0.05/0.025/0.0, favoring a positive threshold.The passage reports that moderate threshold variations are well tolerated.Source: F Hyperparameter Sensitivity Analysis · p. 15
• 76.36/76.22 (−0.14)/75.44 (−0.92) are the average accuracies for candidate-pool sizes 48/32/16, with the smallest pool harming group-statistics stability.A medium-sized pool stays close to the default.Source: F Hyperparameter Sensitivity Analysis · p. 15
• 76.36/75.43 (−0.93)/75.30 (−1.06) are the average accuracies for 4/2/1 GIA phases, with coarser schedules underperforming.The default configuration is α0=0.05, candidate pool 48, and four GIA phases.Source: F Hyperparameter Sensitivity Analysis · p. 15
• Each single-axis corner-case ablation underperforms the full default by about one point on average and on most subtasks.The ablations disable the threshold gate, candidate-pool expansion, or phase decay; the table identifies these variants and their settings.Source: F Hyperparameter Sensitivity Analysis · p. 16

G Analysis of Reasoning Paradigms (Case Study)

• The fully trained MG-Thinker identifies target attributes in Image-1, verifies matching cues in Image-2, then produces accurate localization with a coherent explanation.The case asks the model to find in Image-2 an object matching a target person in Image-1.Source: G Analysis of Reasoning Paradigms… · p. 16
• Direct prediction fails to connect cross-image attributes, generic CoT may drift from discriminative cues, and cold-start localization can accompany an inconsistent rationale.In the illustrated case, direct prediction and generic CoT both yield erroneous localization.Source: G Analysis of Reasoning Paradigms… · p. 16
• The case illustrates improved consistency between reasoning traces and grounding outputs, which the passage identifies as critical for reliable multi-image correspondence and localization.Source: G Analysis of Reasoning Paradigms… · p. 16

H Performance on RefCOCO/+/g Benchmarks

• MG-Thinker remains strongly competitive on RefCOCO/+/g single-image referring-expression grounding, and the comparison table covers these benchmarks.The passage presents this as evidence that post-training preserves general grounding capability rather than over-specializing to multi-image tasks.Source: H Performance on RefCOCO/+/g Benchmarks · p. 16

I Qualitative Cases on MIG-Bench

• MG-Thinker more reliably uses semantic and logical constraints for precise, consistent localization, while Qwen2.5-VL, Migician, and UniVG-R1 show sequence, reasoning, or spatial-alignment weaknesses.Qwen2.5-VL often fails to connect queries to image sequences; Migician tends to rely on surface text cues, while UniVG-R1 reasoning may misdirect localization.Source: I Qualitative Cases on MIG-Bench · p. 16, Figure 14 · p. 17
• Under repeated distractors, MG-Thinker combines positional relations and fine-grained attributes to suppress interference and select regions more stably.The cases involve multiple candidate regions and inconspicuous repeated small objects with similar appearances.Source: I Qualitative Cases on MIG-Bench · p. 16, Figure 15 · p. 17
• In simultaneous multi-object tracking, MG-Thinker keeps outputs more coherent and boxes aligned with temporal and identity cues, although the task challenges all models.Qwen2.5-VL produces chaotic, repetitive outputs; UniVG-R1 struggles to reconcile multiple boxes with its reasoning procedure, and Migician’s accuracy degrades as localization scales.Source: I Qualitative Cases on MIG-Bench · p. 16, Figure 16 · p. 17