Summary (Overview)

• SpatialEvo introduces the first self-evolving framework for 3D spatial reasoning, replacing model consensus with deterministic physical feedback from a Deterministic Geometric Environment (DGE).
• The DGE formalizes 16 spatial reasoning task categories under explicit geometric validation rules, converting unannotated 3D scenes into zero-noise interactive oracles.
• A single shared-parameter policy co-evolves across Questioner and Solver roles: the Questioner generates physically valid spatial questions, while the Solver derives precise answers against DGE-verified ground truth.
• A task-adaptive scheduler endogenously concentrates training on the model's weakest categories, producing a dynamic curriculum without manual design.
• Across nine benchmarks at both 3B and 7B scales, SpatialEvo achieves the highest average scores with consistent gains on spatial reasoning and no degradation on general visual understanding.

Introduction and Theoretical Foundation

Spatial reasoning over three-dimensional scenes is a core capability for embodied intelligence, yet continuous model improvement remains bottlenecked by the cost of geometric annotation. Existing approaches address this through large-scale annotated datasets pairing multi-view images with geometric question-answer pairs, but these share a common structural weakness: the training distribution is fixed at dataset creation time. A static corpus cannot respond to where a model is weak today, cannot generate harder examples as the model grows stronger, and cannot scale without proportional investment in human annotation.

The self-evolving paradigm offers a principled answer through iterative self-play, but existing methods share a critical limitation: because ground truth cannot be read off the environment directly, the training signal must be constructed by aggregating the model's own predictions through majority voting or self-consistency. This introduces systematic bias—the pseudo-labels inherit the model's own prediction errors, and gradient updates anchored to such labels risk reinforcing rather than correcting the model's existing errors.

The authors identify a property unique to 3D spatial reasoning that circumvents this limitation: ground truth is a deterministic consequence of the underlying geometry, computable exactly from point clouds and camera poses without any model involvement. Given a dense point cloud, calibrated camera poses, and a well-formed geometric question, the correct answer can be computed exactly and programmatically. A question about absolute distance reduces to a nearest-point computation on object bounding boxes; a question about relative camera orientation reduces to an arithmetic operation on rotation matrices.

Methodology

Deterministic Geometric Environment (DGE)

The DGE constitutes the physical feedback core of SpatialEvo, serving as the Geometric Oracle within the self-evolution loop. It maps natural language questions onto underlying 3D scene assets for objective verification, replacing unreliable model judgment with programmatic geometric computation.

Task-Specific Geometric Validation Rule Sets: For each of the 16 spatial reasoning task categories, the DGE pre-defines a geometric verification rule set that decouples complex spatial intuition into executable atomic criteria. Each rule set constrains question validity along three dimensions:

• Premise consistency: All scene entities referenced must exist in the underlying assets and be uniquely localizable
• Inferential solvability: Geometric premises must be unambiguously computable
• Geometric degeneracy filtering: Physically unstable or ambiguous edge cases are discarded

Automated Verification Pipeline: The pipeline executes in three stages:

1. Entity Parsing: A lightweight LLM performs structured entity extraction from free-form question text
2. Legality Verification: Validates extracted entities against rule sets, checking premise consistency and reasoning chain validity
3. Ground-Truth Synthesis: Invokes geometric toolkit for precise numerical computation in the global coordinate frame

The DGE acceptance condition is expressed as:

$$\operatorname{Valid} (Q, t, x) = \mathbb{I} \left[ C_{\text{mode}} \wedge C_{\text{extract}} \wedge C_{\text{pool}} \wedge C_{\text{schema}} \wedge C_{\text{solver}} \right]$$

Spatial-Grounded Policy Co-Evolution

Spatial Self-Play Mechanism: A single policy model $\pi_\theta$ alternates between Questioner and Solver roles via role-conditioned prompting. Parameter sharing confers a dual advantage: gradients acquired by the solver through geometric derivation directly improve the questioner's visual-spatial perception, while the geometric intuitions developed by the questioner during boundary exploration deepen the solver's reasoning capacity.

Task-Adaptive Scheduling: A lightweight task scheduler dynamically modulates the task sampling distribution. For each task category $k$, it maintains a cumulative score $S_k$ and sample count $N_k$, estimating historical effective accuracy $\bar{a}_k$ via pseudo-observation smoothing. The sampling weight $w_k$ is negatively correlated with $\bar{a}_k$, with a minimum exploration weight $\delta$ to prevent mastered categories from being entirely excluded.

Questioner Reward Design:

$$r_Q = \alpha f_{\mathrm{fmt}} + (1 - \alpha) f_{\mathrm{valid}} \cdot f_{\mathrm{obs}}$$

where $f_{\mathrm{fmt}}$ measures format compliance, $f_{\mathrm{valid}}$ is the geometric validity score from DGE verification, and $f_{\mathrm{obs}}$ is a visual observation quality score.

Solver Reward Design:

$$r_A = \begin{cases} \alpha f_{\mathrm{fmt}} + (1 - \alpha) f_{\mathrm{acc}}, & \text{if Q is valid,} \\ \alpha f_{\mathrm{fmt}} + (1 - \alpha) f_{\mathrm{explain}}, & \text{if Q is invalid,} \end{cases}$$

where $f_{\mathrm{acc}}$ measures agreement with DGE ground truth, and $f_{\mathrm{explain}}$ measures explanation quality for invalidation reasons.

GRPO Training Procedure: Advantage values are computed independently within their respective groups:

$$\hat{A}^{(i)} = \frac{r^{(i)} - \operatorname{mean}\left(\{r^{(i)}\}\right)}{\operatorname{std}\left(\{r^{(i)}\}\right) + \epsilon}$$

Empirical Validation / Results

Experimental Setup

The DGE is constructed from training splits of ScanNet, ScanNet++, and ARKitScenes, comprising approximately 4K source scenes. SpatialEvo is applied to both Qwen2.5-VL-3B-Instruct and Qwen2.5-VL-7B-Instruct as backbone models, training entirely via online reinforcement learning using the GRPO framework without any supervised fine-tuning stage.

Main Results

SpatialEvo achieves the highest average score under both settings, reaching 51.1 (3B) and 54.7 (7B), outperforming all baselines by a consistent margin. On MMStar, SpatialEvo scores 55.2 (3B) and 62.5 (7B), remaining close to the untuned baseline while annotation-dependent methods fall noticeably below.

Ablation Studies

Replacing DGE ground truth with majority-voting pseudo-GT (w/o Physical Grounding) yields the largest degradation, with VSI-Bench plummeting to 18.8, directly validating the role of deterministic physical feedback.

Online Evolution vs. Static Learning

Despite operating under a narrower task scope, SpatialEvo surpasses the SpatialLadder RL baseline and outperforms all static dataset SFT counterparts, achieving the highest average of 46.3. The core advantage is that the training distribution of a static dataset is frozen at generation time, whereas SpatialEvo continuously aligns the distribution of training samples to the solver's current cognitive frontier through real-time interaction.

Theoretical and Practical Implications

Theoretical Implications: SpatialEvo demonstrates that 3D spatial reasoning possesses a unique property—ground truth is a deterministic consequence of the underlying geometry—that enables circumvention of the systematic bias inherent in model consensus-based self-evolution. This establishes a new paradigm for self-evolving systems where the physical world itself serves as an exact and impartial judge.

Practical Implications: The framework transforms unannotated 3D scene datasets into interactive, zero-noise ground truth judging engines, eliminating the need for costly human annotation. The task-adaptive scheduler enables automatic curriculum learning without manual design, concentrating training resources on the model's weakest categories. The approach demonstrates that spatial specialization does not degrade general visual understanding, maintaining competitive performance on benchmarks like MMStar and RealWorldQA.

Key Limitations: The framework's dependency on high-fidelity 3D assets (indoor point cloud reconstructions, calibrated camera poses) currently restricts its use to static indoor environments. Sensitivity to entity parsing quality and point cloud quality can introduce noise that deterministic geometric reasoning alone cannot fully mitigate.

Conclusion

SpatialEvo presents the first framework to introduce the self-evolving paradigm into 3D spatial reasoning. Unlike conventional self-evolving methods that rely on model voting to construct pseudo-labels, SpatialEvo exploits the unique property of spatial reasoning that visual inputs inherently carry physical information enabling programmatic computation of exact ground truth. This transforms unannotated 3D scene assets into zero-noise reward judges, replacing model consensus with deterministic physical feedback.

A single policy model co-evolves as Questioner and Solver under DGE constraints, with a task scheduler enabling adaptive curriculum self-emergence. Experiments validate significant gains across multiple spatial reasoning benchmarks. The physically grounded self-evolution paradigm explored by SpatialEvo suggests that when self-exploration is rooted in a verifiable physical environment, the continual emergence of spatial reasoning capability is driven endogenously by the model's interaction with the objective world, rather than costly human annotation. Future work may explore reducing dependence on point clouds via alternative or implicit spatial representations to improve scalability and generalization.