Vision-language models answer questions about images with near-human skill until the question asks them to rotate an object in their head, fold a paper net, or say what someone standing across the table would see. On these tasks frontier systems sit near chance, and asking them to "think step by step" often makes open models worse. This proposal tests a specific kind of self-monitoring as a fix. Before answering, the model writes a very short, checkable declaration of how it intends to solve the problem: which reference frame it is using (its own camera view, another viewer's view, or the object's own front), which geometric transformation it will apply and with what parameters, and how confident it is. Because the training problems come from a 3D program that knows the true frame and transformation, every declaration can be graded during training, so the model learns to monitor its own spatial strategy rather than to narrate.
The second half of the idea concerns evaluation. Training on procedurally generated puzzles and then testing on puzzles from the same generator is itself a form of benchmark contamination. The model can learn the generator's habits instead of the skill, and recent work shows this kind of pattern-level leakage cannot be caught by existing detectors. We therefore treat "does the gain survive a change of generator?" as the main outcome. Our claim is that frame-first monitoring narrows the gap between in-generator and out-of-generator accuracy, while plain outcome-rewarded training widens it.
Core Claim
Explicit metacognitive instruction helps spatial reasoning only when it targets the planning step and stays short and structured. Verbal reflection does not have this effect. Its benefit also has to be judged under a generator-shift protocol, because same-generator gains are indistinguishable from contamination.
Three findings motivate this. SpatialViz-Bench builds 1,180 spatial visualization items with FreeCAD and finds that textual chain-of-thought degrades open models. Vision Language Models See What You Want but not What You See shows that no model among 37 beats chance (~29%) on Level-2 perspective-taking, and that the errors come from egocentric bias: the model reports the camera's view. That is a reference-frame error, and a declaration can expose it directly. The Monitor-Generate-Verify formalism argues that autoregressive models fail by committing early to a flawed path (the "prefix dominance trap"). It proposes a pre-generation monitor phase but never puts one into practice. We turn the monitor into something concrete and gradeable.
The work also takes a constraint from On The Fragility of Benchmark Contamination Detection in Reasoning Models. That paper shows reinforcement-learning post-training pushes memorization detectors toward chance, and that contamination with reasoning traces inflates scores by 11.7% on distributionally similar unseen items with detectors at 50–58% AUROC. The same thing will happen to anyone who trains a spatial model on generator-made data. For that reason, fresh random seeds from the same generator are not enough to establish a gain.
Research Program
We will build a FreeCAD-based training generator, separate from SpatialViz-Bench's code. It will cover four families: mental rotation of block assemblies, net folding, three-view projection, and allocentric perspective queries (a rendered agent placed in a scene). Each item carries ground-truth metadata for the correct reference frame, the transformation type, and its parameters (for example "rotate 90° about the vertical axis" or "viewer at 180° from camera").
The model is trained to emit a compact declaration block of at most a few dozen tokens: frame, transformation, parameters, and a confidence $c \in [0,1]$. It then gives the answer with no free-form reasoning. Training uses group-relative policy optimization, the RL method used by SpatialThinker, with reward
$$R = \mathbb{1}[\text{answer correct}] + \lambda_d \, s_d - \lambda_c \, (c - \mathbb{1}[\text{answer correct}])^2,$$
where $s_d \in [0,1]$ is the fraction of declaration fields matching the generator's metadata, the last term is a Brier penalty that rewards calibrated confidence, and $\lambda_d, \lambda_c$ are weights tuned on a held-out validation split from our own generator. The declaration reward is lexicographically gated behind format validity, following SpatialThinker's design.
This differs from SpatialThinker in what gets supervised. SpatialThinker rewards bounding-box alignment of a perceptual scene graph, which is perception, and it evaluates on static benchmarks. We supervise the choice of strategy and frame plus calibration, and we evaluate under generator shift. It also differs from Cog-Rethinker, which adds decomposition and reflection rollouts after failures on math. Our monitoring happens before the answer, and it is graded at the level of fields rather than through free text.
At test time no verifier is needed. This matters because the dependence on domain checkers in SOFAI-LM is exactly what real photographs lack. The calibrated confidence also supports abstention: the model can decline to answer when its confidence is low.
The base model is an open ~7B vision-language model. All arms share the base model, the training items, and the RL compute budget.
Baselines
The naive reference is the base model answering directly. We report it alongside the published reference points. On SpatialViz-Bench, Gemini-2.5-pro scores 44.66% accuracy, o1 41.36%, LLaMA-4-Scout 34.24%, humans 82.46%, and chance is 25.08%, all from SpatialViz-Bench. No reported result was found on Lacuna for an open 7B vision-language model on SpatialViz-Bench, so we will measure our own base model.
The prompting baselines are zero-shot chain-of-thought, Metacognitive Prompting, and Visualization-of-Thought. Metacognitive Prompting is a five-stage verbal self-evaluation that reported 4.8–6.4% relative gains on language-understanding benchmarks. Visualization-of-Thought draws text grids after each step and reported a 23.5% gain on natural-language navigation with GPT-4. These are the strongest training-free forms of explicit metacognition. All prompting baselines run on the same base model, with prompts tuned on the same validation split.
The strongest trained competitor is outcome-only reinforcement learning: the identical GRPO run on identical generator data, rewarded only for correct answers. This is the real threat, because it isolates whether the declaration and calibration terms add anything beyond more spatial training data. A second trained competitor is a SpatialThinker-style reward. We use released SpatialThinker-7B weights if available; otherwise we re-implement its perceptual scene-graph reward on our scenes using rendered boxes. Its reported figures are a 12.1% gain over GPT-4o on 3DSRBench, 65.9% on MM-Star, and 76.3% on RoboSpatial-Home, from SpatialThinker.
Three ablations isolate the key ingredient. The first trains on declarations whose labels are shuffled across items, which tests whether gains come from correct content or merely from the format. The second places the declaration after the answer, which tests the prefix-dominance argument. The third drops the Brier term.
Beating the field means two things. Out-of-generator accuracy on SpatialViz-Bench and PerspectBench must exceed outcome-only RL by more than seed-to-seed variance (three seeds). The in-generator minus out-of-generator gap must also be smaller. A method that wins only in-generator does not count.
Why not just…?
Why not just train with outcome reward on lots of procedural spatial data? This is the most plausible alternative, and it may match us on fresh seeds from our own generator. We expect it to fail at transfer. The Fragility results show that reasoning-trained models generalize over distributionally similar items in ways that look like skill but are tied to the training distribution. The decisive comparison is the generator-shift gap. If outcome-only RL transfers as well as our method, the declaration idea adds nothing for accuracy, and we will report that plainly. In that case only a calibration or abstention benefit would remain.
Why not just prompt the model to reflect, as Metacognitive Prompting does? SpatialViz-Bench already reports that verbal reasoning hurts open models on visualization tasks, and Metacognitive Prompting itself flags overthinking and overconfident wrong answers as failure modes. We still run it, because the paradox was shown for chain-of-thought and not for structured five-stage prompts.
Why not just draw intermediate states, as Visualization-of-Thought does? Its sketchpads are 2D text grids, its gains were concentrated in GPT-4, and weaker models produced incoherent drawings. Folding and perspective tasks have no natural grid encoding. We include it on the perspective and rotation subsets where a grid rendering is possible.
Why not just call a 3D verifier at test time, SOFAI-style? On synthetic items a FreeCAD checker is effectively an oracle, so we report it as an upper bound. It does not exist for What'sUp photographs or PerspectBench scenes, and those are exactly the settings the deficiency literature cares about.
Why not just use a bigger model? Both SpatialViz-Bench and PerspectBench report essentially no gain from 7B to 72B, with identical geometric misattributions at both sizes.
Proposed Experiments
The primary test sets come from independent sources and are never seen in training. SpatialViz-Bench is a different generator with different task code (codebase). PerspectBench offers real Level-2 perspective tasks. What'sUp contains 820 controlled photographs; its best fine-tuned model reached 56% against 99% for humans. We also use 3DSRBench and MV3DSR-Bench from Aligning Cross-View Visual Geometries in LVLMs. In-generator test items use held-out parameter ranges, such as unseen rotation angles and larger assemblies.
The metrics are accuracy, expected calibration error, area under the risk–coverage curve for abstention, declaration-field accuracy on synthetic items, and the transfer gap. On PerspectBench we also measure the rate of egocentric answers (answering with the camera's view), which a correct frame declaration should reduce.
The first decisive check is the transfer gap comparison against outcome-only RL. The second is the shuffled-label ablation. If shuffled declarations help as much as correct ones, the mechanism is format rather than monitoring. The third tests whether declaration correctness predicts answer correctness on out-of-generator items. If it does, the monitor carries real information even without a verifier.
A contamination audit completes the study. We train one additional arm on items generated with SpatialViz-Bench's own code using fresh seeds. This measures how much same-family training inflates SpatialViz scores, which gives a scale for interpreting every other number. We then run loss-based and Min-K% membership detectors on that arm, extending the Fragility findings from math to multimodal spatial reasoning. We expect these detectors to sit near chance, which would confirm that generator-shift evaluation, not detection, is the defensible protocol.
References
- SpatialViz-Bench: A Cognitively-Grounded Benchmark for Diagnosing Spatial Visualization in MLLMs
- Vision Language Models See What You Want but not What You See
- Monitor-Generate-Verify (MGV): Formalising Metacognitive Theory for Language Model Reasoning
- On The Fragility of Benchmark Contamination Detection in Reasoning Models
- SpatialThinker: Reinforcing 3D Reasoning in Multimodal LLMs via Spatial Rewards
- Cog-Rethinker: Hierarchical Metacognitive Reinforcement Learning for LLM Reasoning
- Language Models Coupled with Metacognition Can Outperform Reasoning Models
- Metacognitive Prompting Improves Understanding in Large Language Models
- Visualization-of-Thought Elicits Spatial Reasoning in Large Language Models
- What's "up" with vision-language models? Investigating their struggle with spatial reasoning
- Aligning Cross-View Visual Geometries in LVLMs Through Human-Like Reasoning Learning