Technical Summary: The Art of Interrogation: Consistency Amplifies Factuality in Spatial Reasoning

Problem Statement

Large Reasoning Models (LRMs) and Vision-Language Models (VLMs) exhibit significant deficiencies in spatial reasoning tasks, such as determining relative position, depth ordering, and size comparison. Current benchmarks indicate these models underperform humans by 30–40% and often fail on compositional tasks with near-random accuracy. Existing approaches attempt to bridge this gap by treating it as a knowledge deficit, relying on Supervised Fine-Tuning (SFT) with labeled data from synthetic 3D simulators or real images annotated by cascades of vision models (e.g., depth estimators, detectors). However, these methods require expensive ground-truth supervision and may not address the root cause: a lack of internal logical coherence in the model's reasoning process.

Methodology

The authors propose that pre-trained LRMs already possess latent spatial reasoning capabilities but lack alignment with fundamental geometric principles. Instead of ingesting new factual knowledge, the paper introduces a self-supervised reinforcement learning (RL) framework that enforces logical consistency under geometric and semantic transformations.

1. Consistency Verifiers

The core innovation is the consistency verifier, a reward function that operates without ground-truth labels. It evaluates whether a model's answers to a prompt $x$ and an augmented prompt $x' = T(x)$ satisfy a known relationship induced by the transformation $T$.

• Transformations: The framework applies image transformations (e.g., horizontal flips, object-preserving crops) and text transformations (e.g., swapping object order, negating relations like "left" $\to$ "right").
• Invariance vs. Equivariance:
  • Invariant: Transformations that preserve the answer (e.g., cropping, color jitter). The model's answers to $x$ and $x'$ must match.
  • Equivariant: Transformations that predictably flip the answer (e.g., horizontal flip for "left/right" questions). The model's answers must disagree.
• Reward Signal: The verifier returns 1 if the pair of answers satisfies the expected relationship (match or mismatch) and 0 otherwise. This provides a self-supervised reward signal derived entirely from the transformation logic, not external labels.

2. OT-GRPO (Optimal Transport Group Relative Policy Optimization)

Standard Group Relative Policy Optimization (GRPO) assumes scalar rewards for individual completions. Since consistency verifiers score pairs of completions, the authors introduce OT-GRPO, a minimal-matching variant:

• Adversarial Pairing: Instead of random pairing or averaging (one-to-all), the algorithm solves a linear assignment problem to find the permutation $\sigma^*$ that minimizes the total consistency score between the set of completions from the original prompt and the augmented prompt.
• Rationale: This "minimal consistency" strategy forces the model to confront its most challenging inconsistencies. If a model is truly consistent, the minimal pairing will still yield high rewards; if it is inconsistent, the adversarial pairing exposes these failures, preventing the model from gaming the system through lucky alignments.
• Optimization: The matched pairs receive a shared advantage, and the GRPO update is applied to both the original and augmented completions.

3. Training Setup

• Models: Fine-tuning of Qwen2.5-VL (3B and 7B parameters) using LoRA.
• Datasets: SUN RGB-D (indoor) and KITTI (outdoor) with Omni3D annotations.
• Tasks: Four binary spatial reasoning tasks: Orientation (left/right), Depth (closer/further), Size (volume comparison), and Relative Distance.
• Format: The model is prompted to use Chain-of-Thought (CoT) reasoning within specific tags, with an additional format reward ensuring structural adherence.

Key Results

The paper demonstrates that consistency-based training approaches the performance of ground-truth supervised training without requiring any labeled data.

1. Performance Parity: On the SUN RGB-D dataset, consistency-trained models achieved accuracy within 2.3–2.7 percentage points of models trained with ground-truth labels across all four tasks. For the 7B model, consistency training reached 76.8% accuracy compared to 79.6% for accuracy-trained models.
2. Generalization:
   • Cross-Task: Training on one spatial task (e.g., orientation) improved performance on others (e.g., depth), with consistency training transferring nearly as effectively as accuracy training.
   • Cross-Domain: Models trained on indoor scenes (SUN RGB-D) generalized well to outdoor scenes (KITTI), with consistency training showing minimal degradation compared to supervised baselines.
3. Superiority over Self-Supervised Baselines: The proposed method outperformed seven existing self-supervised baselines, including Visual Jigsaw and SSL4RL variants, by significant margins (up to 24 percentage points over some SSL4RL variants).
4. Robustness to Label Corruption: In experiments where ground-truth labels were corrupted (flipped) with increasing noise, consistency training (which uses no labels) outperformed accuracy training once label noise exceeded 10–20%. This suggests consistency training is more robust to the error accumulation inherent in automated annotation pipelines.
5. Extension to Numeric Tasks: The framework was successfully extended to numeric outputs (counting and absolute distance estimation), where consistency training again closely tracked the performance of accuracy-trained models.

Significance and Claims

The paper claims that enforcing logical coherence through consistency verifiers unlocks latent spatial reasoning abilities already present in pre-trained models. Key takeaways include:

• Label-Free Efficiency: High-quality spatial reasoning can be achieved without the cost and complexity of generating ground-truth spatial annotations.
• Mechanism of Improvement: The results support the view that RL with verifiable rewards acts as a filter, selecting and reinforcing pre-existing reasoning patterns that adhere to geometric laws, rather than creating entirely new capabilities.
• Scalability: The approach is applicable to diverse tasks and domains, suggesting a general path for improving VLMs in areas where ground-truth data is scarce or expensive to obtain.

The authors conclude that consistency training is a viable, robust alternative to supervised fine-tuning for spatial reasoning, offering a path to more reliable and generalizable embodied AI and robotics applications.