Technical Summary: Iterative Visual Thinking (IVT)

Problem Statement

Vision-Language Models (VLMs) currently excel at single-shot spatial grounding tasks, such as Referring Expression Comprehension (REC), where they predict bounding box coordinates directly from text descriptions. However, these models lack an inherent mechanism to observe and correct their own predictions. While humans solve spatial tasks through a closed-loop process of hypothesizing, verifying against visual evidence, and refining, VLMs operate in a "single-pass" paradigm.

The paper identifies a critical "self-correction gap": naively prompting a VLM to iterate over its own rendered predictions leads to catastrophic failure. In experiments, applying iterative visual thinking without specific training caused the accuracy at IoU ≥ 0.5 (Acc@0.5) to collapse from 79.6% to 48.7% (a 31 percentage point drop). This reveals that while VLMs possess strong grounding capabilities, they cannot natively interpret visualizations of their own errors to perform self-correction.

Methodology: Iterative Visual Thinking (IVT)

The authors propose Iterative Visual Thinking (IVT), a closed-loop framework where the model predicts a bounding box, observes the prediction rendered as a colored overlay on the original image, and uses this visual feedback to iteratively refine its output.

The training pipeline employs a two-phase recipe to bridge the self-correction gap:

Phase 1: SFT Warm-up (Supervised Fine-Tuning)

Direct Reinforcement Learning (RL) fails initially because the base model has no concept of spatial self-correction and tends to produce degenerate, identical predictions at every step. To address this, the authors introduce a data synthesis strategy:

• Student-Prediction-Based Synthesis: Instead of using artificially perturbed ground-truth boxes, the pipeline uses the student model's own realistic initial predictions (Step 0) as the starting point for error trajectories. This avoids "step-0 sandbagging," where models learn to intentionally produce poor initial predictions to mimic training data.
• Teacher Reasoning Traces: A teacher VLM generates corrective reasoning traces. The teacher observes the student's initial prediction (or the rendered overlay of previous steps) and writes a reasoning trace identifying spatial errors (e.g., "shifted," "too large") and explaining the necessary correction.
• Training: The student model is fine-tuned on these trajectories using cross-entropy loss, learning the structure of iterative visual thinking: interpreting rendered overlays, generating grounded reasoning, and outputting valid boxes.

Phase 2: GRPO Fine-tuning

Once the model understands the structure of refinement via SFT, Group Relative Policy Optimization (GRPO) is applied to stabilize the multi-step process.

• Reward Function: A simple reward is used: the final-step Intersection over Union (IoU) plus a small format bonus for valid bounding box parsing. This simplicity avoids the instability observed with complex, multi-component reward functions.
• Objective: GRPO optimizes trajectory-level outcomes using group-relative advantages, encouraging the model to maintain or improve predictions across refinement steps rather than degrading them.

Key Contributions

1. The IVT Framework: A closed-loop spatial reasoning system where VLMs predict, observe rendered predictions, and self-correct. The paper demonstrates that this capability is not native; without explicit training, iterative refinement causes severe performance degradation.
2. Self-Referential Data Synthesis: A strategy that exploits the base model's own predictions as realistic error sources for SFT training, eliminating the need for human-annotated correction trajectories.
3. Asymmetric Two-Phase Training: The paper establishes that SFT is the primary enabler of accuracy gains, teaching the model the structure of iterative thinking. GRPO acts as a stabilizer, reducing per-step performance degradation but not driving the initial accuracy leap.
4. Scalability Proof: The authors demonstrate that spatial self-correction can be instilled with a modest dataset of 2,400 samples trained on a single GPU, surpassing single-shot baselines.

Experimental Results

Experiments were conducted on a mixed benchmark of 505 test samples from RefCOCOg, Ref-Adv-S, and Ref-L4, using the Qwen3-VL-4B-Instruct model.

• The Self-Correction Gap: The base model with untrained IVT dropped Acc@0.5 from 79.6% to 48.7%.
• SFT Impact: After SFT warm-up with IVT, the model surpassed the single-shot base model on all metrics:
  • Acc@0.5: 82.0% (+2.4pp)
  • Acc@0.7: 74.1% (+3.2pp)
  • Acc@0.9: 48.3% (+2.8pp)
• GRPO Impact: Adding GRPO to SFT slightly reduced overall accuracy compared to SFT+IVT alone (Acc@0.5: 80.6% vs 82.0%) but significantly improved refinement stability.
  • The mean per-step IoU degradation dropped from 0.140 (SFT alone) to 0.029 (SFT+GRPO), a 5× reduction.
  • Without SFT, GRPO alone resulted in "degenerate stagnation" where the model simply copied its first prediction at every step.

Significance and Claims

The paper claims that spatial self-correction is a learnable capability that must be explicitly taught, as it does not emerge naturally in base VLMs even with visual feedback. The work highlights a fundamental distinction between "grounding capability" (finding an object once) and "self-correction ability" (iteratively fixing errors based on visual feedback).

The proposed predict-render-refine loop offers a new paradigm for spatial reasoning that mirrors the "act-observe-correct" cycle of embodied agents. Furthermore, the iterative process provides interpretability, as the reasoning traces and rendered overlays at each step make the model's spatial decision process transparent, contrasting with the "black-box" nature of single-shot predictions. The authors note that while the current implementation focuses on single-object grounding, the framework is representation-agnostic and extends naturally to segmentation, keypoints, and multi-object reasoning.