Spatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning

Conference: CVPR 2026 arXiv: 2510.27606 Code: GitHub Area: Image Generation Keywords: Spatial Understanding, Self-Supervised Learning, Reinforcement Learning (RLVR), Large Vision-Language Models, Depth Perception

TL;DR

This paper proposes Spatial-SSRL, a self-supervised reinforcement learning paradigm that automatically constructs five pretext tasks (patch reordering, flip recognition, cropped patch inpainting, depth ordering, and relative 3D position prediction) from standard RGB/RGB-D images. By optimizing LVLMs with GRPO, the method achieves average improvements of 3.89%–4.63% across seven spatial benchmarks without any human annotation or external tools.

Background & Motivation

Background: Large Vision-Language Models (LVLMs) are approaching saturation on tasks such as VQA and image captioning, yet their spatial understanding capabilities remain far below human-level performance. Existing approaches fall into two categories: data-driven SFT (constructing spatial QA pairs for fine-tuning) and RLVR (reinforcement learning with verifiable rewards).

Limitations of Prior Work: SFT methods rely on costly human annotation or GPT-4-generated QA pairs and tend to overfit to dataset-specific patterns. Tool-augmented approaches (e.g., using depth estimators or object detectors) involve complex pipelines and high computational costs. Simulation-based methods (rendering 3D scenes) suffer from domain gaps with the real world. RLVR approaches are constrained by specific environments (e.g., 3D scans), limiting data scale and coverage.

Key Challenge: Spatial understanding requires large-scale verifiable supervision signals, yet all existing methods incur prohibitive costs to obtain them—either through expensive human annotation, complex toolchains, or dependence on specialized 3D datasets.

Goal: Design a zero-annotation, tool-free, and scalable self-supervised scheme that generates verifiable spatial understanding training signals and integrates naturally with the RLVR training paradigm.

Key Insight: The inherent structural consistency within images (relative depth, geometric consistency, viewpoint invariance) itself provides deterministic and verifiable supervision signals. Pretext tasks from visual self-supervised learning (SSL) are repurposed as reward functions for RLVR rather than traditional feature pre-training objectives.

Core Idea: Classic SSL tasks (jigsaw puzzles, flip detection, etc.) are reformulated as QA prompts with deterministic verification functions for LVLMs, and GRPO is directly applied for post-training.

Method

Overall Architecture

Spatial-SSRL consists of two stages: (1) self-supervised task construction—automatically generating QA pairs for five pretext tasks from RGB/RGB-D images (the Spatial-SSRL-81k dataset, with 100% annotation accuracy); and (2) RL training—an SFT cold-start phase to familiarize the model with task formats, followed by GRPO optimization. The five tasks span 2D layout understanding (depth-free) and 3D spatial reasoning (depth-based).

Key Designs

1. Shuffled Patch Reordering:

   • Function: The image is divided into an \(M \times N\) grid, randomly shuffled, and the model is asked to predict the permutation \(\pi^{-1}\) that restores the original arrangement.
   • Mechanism: For image \(I\), a patch grid \(\mathcal{X} = \{x_{i,j}\}\) is constructed and a random permutation \(\pi\) is applied to obtain the shuffled image. The ground-truth answer is the inverse permutation \(\pi^{-1} = [\pi^{-1}(0), \pi^{-1}(1), \ldots, \pi^{-1}(M \times N - 1)]\). Optionally, one random patch is masked in white to increase difficulty and prevent the model from exploiting boundary artifacts.
   • Design Motivation: Recovering the original patch ordering inherently requires understanding global 2D layout consistency and relative positional relationships—skills that transfer directly to understanding object arrangements in real scenes.

2. Flipped Patch Recognition:

   • Function: One randomly selected patch is flipped horizontally or vertically, and the model must identify the index of the flipped patch and the flip direction.
   • Mechanism: The selected patch \(\hat{x}_t\) is transformed with equal probability via vertical flip \(x_{\text{vert}}(r,c) = x(P_H - 1 - r, c)\) or horizontal flip \(x_{\text{horz}}(r,c) = x(r, P_W - 1 - c)\), with the answer being \([t, d]\).
   • Design Motivation: Detecting subtle orientation violations requires sensitivity to local geometry, mirror symmetry, and directional cues such as text, faces, and shadows.

3. Cropped Patch Inpainting:

   • Function: A randomly cropped region is masked in black, and the model must select the correct filling patch from four candidates (including three distractors).
   • Mechanism: The distractors are carefully designed—a 90° rotated version, an interior sub-region, and an exterior expanded region—all visually similar to the correct answer, forcing the model to attend to fine-grained texture continuity and semantic consistency.
   • Design Motivation: Tests texture-context matching and fine-grained structural reasoning ability.

4. Regional Depth Ordering:

   • Function: Three regions with clearly separated depths are selected, labeled with numeric indices, presented in shuffled order, and the model is asked to sort them from nearest to farthest.
   • Mechanism: Three regions satisfying the following constraints are selected from depth map \(D\): intra-region depth range \(r(R_i) < r_{\max} = 0.15\) (intra-region consistency), and inter-region separation \(d(R_i, R_{i+1}) > d_{\min} = 0.05\) (inter-region discriminability).
   • Design Motivation: Sorting requires integrating depth cues, perspective understanding, and ordinal reasoning—foundational capabilities for 3D scene understanding.

5. Relative 3D Position Prediction:

   • Function: Given the orientation of an object, the model predicts the relative position of another point in the object's coordinate frame (combinations of left/right/front/back).
   • Mechanism: A 2D rigid-body transformation maps camera-frame coordinates \((x_2, z_2)\) to the object frame \((\tilde{x}_2, \tilde{z}_2)\), and direction labels \((\tilde{p}_x, \tilde{p}_z)\) are determined by thresholding. Object orientation \(\theta\) is sampled uniformly from the four cardinal directions.
   • Design Motivation: Requires mental rotation, egocentric coordinate transformation, and depth integration—the highest-order task among the five for spatial understanding.

Loss & Training

• Cold-Start SFT: Fine-tuning for 5 epochs on approximately 3,600 samples (lr=\(1 \times 10^{-5}\)) to familiarize the model with task formats.
• GRPO Optimization: KL regularization weight 0.01, 5 rollouts per sample, temperature 1.0, batch size 128, lr=\(1 \times 10^{-6}\), 360 steps.
• Reward Design: \(r = 0.9 \cdot r_{\text{acc}} + 0.1 \cdot r_{\text{fmt}}\), with accuracy weighted substantially higher than format.
• Think tags are used to guide chain-of-thought output.

Key Experimental Results

Main Results (7 Spatial Understanding Benchmarks)

Model	Spatial457	3DSRBench	SpatialEval	QSpatial+	What'sUp	ViewSpatial	VSI-Bench	Avg
Qwen2.5-VL-3B	33.70	50.30	54.65	33.66	85.85	35.38	27.84	45.91
Spatial-SSRL-3B	46.07	51.72	59.59	39.60	86.71	36.62	33.49	50.54
Δ	+12.37	+1.42	+4.94	+5.95	+0.86	+1.24	+5.65	+4.63
Qwen2.5-VL-7B	44.67	53.39	62.37	46.53	86.95	36.83	38.08	52.69
Spatial-SSRL-7B	53.34	56.53	64.03	54.46	90.61	37.81	39.29	56.58
Δ	+8.67	+3.14	+1.66	+7.93	+3.66	+0.98	+1.21	+3.89

Improvements are observed across all seven benchmarks, with the largest gain of +12.37% on Spatial457.

Reasoning Capability Verification

Configuration	Avg Accuracy	Notes
Qwen2.5-VL-7B (w/o reasoning)	52.69	Baseline
Qwen2.5-VL-7B (w/ reasoning)	49.58	Reasoning hurts! (−3.11)
Spatial-SSRL-7B (w/ reasoning)	56.58	Reasoning chain is effective (+3.89)

Enabling chain-of-thought reasoning in the baseline model actually degrades performance (What'sUp: 86.95→70.61), indicating that the base model lacks effective spatial reasoning ability. Spatial-SSRL successfully trains the model to generate effective reasoning chains through RL.

Key Findings

• 3D reasoning benchmarks benefit most (Spatial457 +12.37%, QSpatial+ +7.93%), validating the contribution of depth-based tasks.
• The finding that enabling CoT reasoning in the base model degrades performance is significant—it demonstrates that spatial reasoning ability requires dedicated training rather than simple prompt engineering.
• On Qwen3-VL-4B, spatial performance improves by +1.29% while general VQA also improves by +1.18%, confirming that the method does not harm general capabilities.
• The Spatial-SSRL-81k dataset achieves 100% annotation accuracy because all answers are derived from deterministic transformations—a standard that noisy-detector-based methods cannot match.
• The cold-start SFT is necessary: direct RL training results in a success rate below 5% for generating correctly formatted outputs.

Highlights & Insights

• The combination of SSL and RLVR is the primary contribution. SSL pretext tasks naturally provide deterministic and verifiable answers, perfectly aligning with the verifiable reward requirement of RLVR. This insight may inspire a substantial body of follow-up work extending other SSL tasks to LVLM post-training.
• The complementary design of five tasks covers the full hierarchy from 2D layout to 3D spatial relations: patch reordering (global layout) → flip recognition (local orientation) → inpainting (texture consistency) → depth ordering (3D depth) → 3D position (egocentric coordinate transformation).
• The ingenuity of distractor design: in the inpainting task, distractors consisting of rotated versions, interior sub-regions, and exterior expanded regions prevent the model from exploiting low-level features.

Limitations & Future Work

• The two depth-dependent tasks require RGB-D data, limiting the range of applicable data sources (although depth can be estimated monocularly, this introduces noise).
• The relative weighting of the five tasks is not carefully tuned; the current setup uses equal mixing ratios.
• Evaluation is conducted only on Qwen2.5-VL and Qwen3-VL; generalization to other LVLM architectures (e.g., LLaVA, InternVL) remains unknown.
• While the 81k data scale is already smaller than typical SFT approaches, there is further room to improve RL training efficiency.
• Additional SSL tasks could be explored—such as color channel permutation, frequency-domain transformation prediction, and multi-view consistency verification.

Related Work & Insights

• vs. SpatialLadder/SpaceR: These methods depend on 3D scan datasets and complex pipelines, whereas Spatial-SSRL requires only standard RGB/RGB-D images with a minimal pipeline. In terms of performance, Spatial-SSRL-7B (56.58) surpasses SpaceR-7B (54.54).
• vs. Jigsaw-R1/Visual Jigsaw: These methods share a similar SSL+RL philosophy but cover only the jigsaw task. Spatial-SSRL designs five complementary tasks, achieving more comprehensive spatial understanding improvements.
• vs. SSL4RL: Focuses exclusively on 2D tasks, whereas Spatial-SSRL covers both 2D and 3D; the depth-based tasks contribute substantial performance gains.

Rating

• Novelty: ⭐⭐⭐⭐⭐ The paradigm of using SSL pretext tasks as RLVR rewards is pioneering, and the five-task design is comprehensive.
• Experimental Thoroughness: ⭐⭐⭐⭐ Seven benchmarks, three base models, and general capability verification are provided, though ablation studies could be more detailed.
• Writing Quality: ⭐⭐⭐⭐⭐ Method motivation is clearly articulated, mathematical formalization of task designs is rigorous, and figures are excellent.
• Value: ⭐⭐⭐⭐⭐ Zero annotation, scalable, and naturally compatible with RLVR; opens a new direction for improving spatial understanding in LVLMs.

Related Papers

• [CVPR 2026] Enhancing Spatial Understanding in Image Generation via Reward Modeling
• [ICCV 2025] CoMPaSS: Enhancing Spatial Understanding in Text-to-Image Diffusion Models
• [CVPR 2026] Circuit Mechanisms for Spatial Relation Generation in Diffusion Transformers
• [CVPR 2026] Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods
• [CVPR 2026] Just-in-Time: Training-Free Spatial Acceleration for Diffusion Transformers