Separating perception from reasoning in vision-language models: a model-free render ceiling for crystal structures

arXiv:2609.00663 · cs.CV, cond-mat.mtrl-sci, physics.chem-ph, quant-ph · Submitted 2026-09-01 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Separating perception from reasoning in vision-language models".

Jane: Here's the gist of the paper. Inverting frozen cameras yields an exact, per-sample reference for what the images support The render ceiling is a model-free reference for benchmarks built by rendering known objects, recovering exactly the answer the images support through inverting frozen cameras and re-solving cross-view correspondence This instrument exposes extraction-stage fabrication that downstream accuracy would misattribute to reasoning Across fourteen vision–language models, supplying exact geometry as text lifts every model yet closes under half the gap for thirteen, while a supervised vision model with no language component reads the same images at 0.8952, above every vision–language model The Model-Free Reference Mechanism The render ceiling is defined as a quantity that is measured rather than assumed, where a ceiling is measured rather than assumed it still has a model inside it The exactness of the oracle can be characterized by an emptiness condition on a set of geometric phantoms, coincidences between distinct atoms’ projections, and it is the oracle’s only failure mode The theorem proves that on 2,160 rendered crystal structures, this set is empty and the ceiling is 1.0000 This means every point of a model’s deficit belongs to the model Attribution Ladder Decomposition The instrument exposes extraction-stage fabrication that downstream accuracy would misattribute to reasoning, yielding camera-placement rules for benchmark builders The attribution ladder has three rungs: R1, the oracle ceiling of Definition 1; R3(m), the accuracy of m given ground-truth fractional coordinates, species and cell parameters as text; and R4(m), the accuracy of m on the pixel renders The perception component P(m) is defined as R3(m) −R4(m), and S(m) is defined as R1 − R3(m) Perception vs.

Tom: Reasoning Limits The decomposition gives a median perception share of 0.2901 across the 14 models scored under the coordinate condition, with the post-perception residual being the larger of the two for 13 of them The raw component moves the other way, R3 − R4 correlating with R4 at ρ = −0.6439, so the strongest pixel performers are those for whom exact geometry adds least even though the share it closes shows no trend The finding with the widest reach for AI-assisted materials design is fabrication Design Rules and Transferability The instrument yields design rules for benchmark builders, stating which camera placements survive extraction noise, that placement rather than view count sets the noise budget, and where a pipeline’s own reading becomes the binding ceiling The construction transfers to any benchmark whose forward rendering can be written down and inverted Adding a view never destroys identifiability, and a camera aligned with a lattice vector separates no atoms differing by multiples of it The oracle answers from exact centroids, so the ceiling states what the geometry determines rather than what a pixel reader recovers Extraction Stage Audit Primitive The fabrication finding of the main text uses a 206-structure denominator: the four gate-excluded structures, listed with the gate in the Reproducibility record, count as unanswered in the full-sample figure A two-stage pipeline whose intermediate cannot be checked will attribute fabrication to reasoning, and only a model-free reference distinguishes the two The audit primitive these pipelines currently lack is the model-free reference for any stage whose forward map is known Final Conclusion on Model Deficit The ceiling-to-model gap is bounded rather than estimated, and the ceiling is certified per sample rather than assumed Read against that ceiling, most models remain limited by what survives the removal of perception The reading that perception is the binding constraint on multimodal materials tasks is therefore a statement about particular models rather than about the task The information is there to be read is settled independently by the supervised pixel baseline, which lands above every vision–language model and close to the ceiling using only the benchmark’s own renders Algorithm 1 Geometric oracle O(s, ) is deterministic and assumes ideal extraction, so R1 is a bound from the projection geometry rather than from the pixels The constant is a property of the camera set, κ = 2.7321 here, and it diverges as two directions approach parallel The oracle answers from exact centroids, so the ceiling states what the geometry determines rather than what a pixel reader recovers Phantom Set Analysis Definition 7 Phantom set, componentwise, states that a species-labelled point (p, z) lies in Φdelta(s, ) if it is further than δ from every same-species atom The theorem’s merge clause is performed at δ = τ: this is the single implementation tolerance δ of the oracle algorithm, tied to the label map’s symmetry tolerance τ rather than a third, independent quantity Every emptiness claim here and in the main text is therefore stated at the δ it was measured at The mean phantom excess falls monotonically with view count, from 4.911 at two views to 0.480, 0.048 and exactly 0.000 at five, where all 210 structures recover the true atom count The mean phantom excess falls monotonically with view count, from 4.911 at two views to 53.7 ms at 80 atoms Extraction Tolerance Frontier The extraction-tolerance frontier reported in the main text has one non-monotone point: at n = 1950, δ = 0.005Å costs one structure where δ = 0.01Å costs none The ceiling is 1.0004 at δ = τ; the tighter δ = tau setting produces one under-merging failure on the scaled sample, whereas larger values of δ introduce over-merging Model Performance Comparison The strongest scored model Gemini 3.6 Flash reaches R3=0.8524; R4=0.7333 The weakest of the fourteen by R1 (oracle ceiling) is GPT-4.1 mini Coordinate condition, with R3/R4 = 0.4143 and 0.3667 Perception dominance is therefore a property of the top of the roster, not of the task Statistical Robustness The test families, correction procedure and reporting conventions are specified in the Statistics and reproducibility subsection of Methods Every load-bearing claim survives correction The median perception share is reported with a percentile bootstrap over models: against the exact R1 the median is 0.2901 with 95% confidence interval

0.1492, 0.3576: The label-certification lower bound is a one-sided exact bound, 0.051/224 = 0.9865 at k = 224 agreements, conditional on the four-structure quarantine The full family is 26 hypothesis tests with numeric p-values, the ten headline tests plus the render-convention probe’s 16 individual paired comparisons Reproducibility and Artifacts The released artifact, the benchmark records, the per-structure predictions for every arm and the preregistration of each analysis are openly available at https://github.com/KurbanIntelligenceLab/render-ceiling The repository carries the code implementing the geometric oracle, the render protocol, the labelling pipeline and every analysis reported here, under the MIT licence Artifact disclosures.

Jane: Two documentation mismatches could mislead a model or a reader into reasoning about bond geometry that is not in the image The analytic oracle never touches a rendered pixel, so its numbers are unaffected Runtime and complexity of the oracle is O(V 2n 2) in all, dominating the O(n) merge The census recovers every atom exactly on all 1,950 structures, consistent with the identifiability theorem The full five-camera set reproduces R1 = 200/210 = 1.0 at δ = τ through this independent implementation The same geometry yields design rules a benchmark builder can apply before release The construction transfers to any modality whose forward rendering can be written down and inverted Key Findings on Model Strength The natural strengthening, that the share rises with model strength, is not supported: across the 14 models the share and pixel accuracy are unrelated by rank correlation The raw component moves the other way, R3 − R4 correlating with R4 at ρ = −0.6439, so the strongest pixel performers are those for whom exact geometry adds least Perception dominance is therefore a property of the top of the roster, not of the task The reading that perception is the binding constraint on multimodal materials tasks is therefore a statement about particular models rather than about the task The finding with the widest reach for AI-assisted materials design is fabrication The agents are assembled from exactly such intermediates, a structure read from a figure and handed to a property predictor or a synthesis planner, checked, where it is checked at all, by a learned verifier that cannot separate a fabricated intermediate from a mistaken one Camera Placement vs.

Tom: View Count The conditioning constant orders camera placements rather than setting the noise budget Adding a view never destroys identifiability, and a camera aligned with a lattice vector separates no atoms differing by multiples of it Four separated cameras empty the phantom set on this protocol with the fifth adding reconstruction fidelity The mean render ceiling R1 under two-, three-, four- and five-camera subsets of the frozen protocol, at the certified tolerance δ = τ = 0.01Å and at the released tolerance δ = 15τ; error bars span the minimum and maximum over camera subsets of that size View count behaves as the first design consequence predicts, non-decreasing as views are added The conditioning constant, evaluated over the ten independent pairs of the five-camera set, is 1.3280 (mean 1.

Paper summary: Tom: So to wrap up this paper, "Separating perception from reasoning in vision-language models: a model-free render ceiling for crystal structures." The authors introduce the render ceiling as a model-free reference for benchmarks built by rendering known objects.

Jane: They show that inverting frozen cameras and re-solving cross-view correspondence recovers exactly what the images support, which means it’s an exact oracle. This proves the ceiling fails.

Lu: The core finding is separating perception from reasoning using this decomposition: R1, the oracle ceiling of definition one; R3(m), accuracy given ground truth coordinates and parameters as text; and R4(m), accuracy on pixel renders.

Meng: The implication is that we can finally say something concrete about model limitations, not just vague estimations. We find that perception dominance is a property of the top of the roster, not of the task itself.

Tom: And they found that for materials design, fabrication seems to be where the real reach is. The agents in these applications are assembled from intermediates read from figures and checked by a verifier that can distinguish fabricated intermediates from mistaken ones.

Jane: So what does this mean practically? It means we have a tool to ground the evaluation in exact geometry, which sets a bound on performance per sample instead of just guessing where it might be.

Lu: The construction transfers to any benchmark whose forward rendering can be written down and inverted, meaning adding a view never destroys identifiability, and a camera aligned with a lattice vector separates no atoms differing by multiples of it.

Meng: And we get design rules for benchmark builders that tell them which camera placements survive extraction noise, which placement rather than view count sets the noise budget.

Tom: It shifts the focus from just chasing higher accuracy scores to understanding what survives after you factor out perception, and that’s a concrete way to approach model limitations.

Conclusion: Tom: So we're wrapping up this paper, "Separating perception from reasoning in vision-language models: a model-free render ceiling for crystal structures."

Jane: The main idea is they built this render ceiling, which is like an exact reference point based on what the images actually support.

Lu: They do this by inverting frozen cameras and re-solving how the views correspond to each other. It gives you a real answer, not just a guess.

Meng: So, instead of trusting the model's reasoning alone, we get a number—a ceiling—that tells us where the model is actually right based on geometry.

Tom: It’s about separating what the AI *perceives* from what it can actually *reason* about with solid data.

Jane: And they show this separation using three levels of accuracy, R1, R3, and R4. It’s a way to see exactly where the model is falling short on its own.

Lu: The most striking result is that perception dominance isn't just for the best models; it’s a property of those top models themselves when you look at the task.

Meng: That means we can start to pin down exactly what parts of these complex AI systems are actually doing reasoning versus just memorizing pixels.

Tom: It gives us concrete design rules for people building benchmarks, telling them which camera setups actually work and which ones create noise.

Jane: And it points toward fabrication as a big area where this kind of exact geometry reference is really useful for materials design.

Lu: The construction works because you can write down the forward rendering of any object and reverse it, meaning adding more cameras never ruins your ability to figure out the structure.

Meng: So, it’s not about getting a higher accuracy score; it’s about using this geometric ceiling to set a hard limit on what we can expect from AI in these tasks.

Tom: This is moving us toward evaluation methods that are grounded in actual physics and geometry rather than just blurry pixel matching.

Jane: And if we understand this separation better, it changes how we trust the outputs of vision language models for scientific discovery.

Department of Electrical and Computer Engineering, Texas A&M University · Department of Electrical and Computer Engineering, Texas A&M University at Qatar · Department of Prosthetics and Orthotics, Ankara University · College of Science and Engineering, Hamad Bin Khalifa University

cs.CV, cond-mat.mtrl-sci, physics.chem-ph, quant-ph

Submitted: 2026-09-01

Updated: 2026-11-14

Code: https://github.com/KurbanIntelligenceLab/render-ceiling

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 87/100

The gist: Inverting frozen cameras yields an exact, per-sample reference for what the images support The render ceiling is a model-free reference for benchmarks built by rendering known objects…

Key concepts

Render Ceiling
This is a model-free benchmark that defines the exact geometric truth supported by an image set. It measures what the images support directly, acting as an oracle where its only failure mode is when a point belongs to the model's deficit.
Attribution Ladder Decomposition
This framework breaks down accuracy into three rungs: R1 (the oracle ceiling), R3(m) (accuracy with ground-truth coordinates), and R4(m) (pixel render accuracy). Perception is defined as the difference between these two, showing where perception limits reasoning.
Phantom Set Analysis
This analysis quantifies how accurately the model identifies atoms based on their projections. It uses a definition of a 'phantom set' to measure how far points are from same-species atoms, showing that this excess falls monotonically with view count, eventually reaching zero at five views.

Terminology

Summary

Inverting frozen cameras yields an exact, per-sample reference for what the images support The render ceiling is a model-free reference for benchmarks built by rendering known objects, recovering exactly the answer the images support through inverting frozen cameras and re-solving cross-view correspondence This instrument exposes extraction-stage fabrication that downstream accuracy would misattribute to reasoning Across fourteen vision–language models, supplying exact geometry as text lifts every model yet closes under half the gap for thirteen, while a supervised vision model with no language component reads the same images at 0. p. 0

Where the paper says it

qa A B S T R A C T Multimodal evaluations cannot say whether a vision–language model misread an image or misreasoned about it, because every existing method for separating the two places a second model in the loop. We introduce the render ceiling, a model-free reference for benchmarks built by rendering known objects: inverting the frozen cameras and re-solving cross-view correspondence recovers exactly the answer the images support.

8952, above every vision–language model

The Model-Free Reference Mechanism

The render ceiling is defined as a quantity that is measured rather than assumed, where a ceiling is measured rather than assumed it still has a model inside it The exactness of the oracle can be characterized by an emptiness condition on a set of geometric phantoms, coincidences between distinct atoms’ projections, and it is the oracle’s only failure mode The theorem proves that on 2,160 rendered crystal structures, this set is empty and the ceiling is 1.0000 This means every point of a model’s deficit belongs to the model

Attribution Ladder Decomposition

The instrument exposes extraction-stage fabrication that downstream accuracy would misattribute to reasoning, yielding camera-placement rules for benchmark builders The attribution ladder has three rungs: R1, the oracle ceiling of Definition 1; R3(m), the accuracy of m given ground-truth fractional coordinates, species and cell parameters as text; and R4(m), the accuracy of m on the pixel renders The perception component P(m) is defined as R3(m) −R4(m), and S(m) is defined as R1 − R3(m)

Perception vs. Reasoning Limits

The decomposition gives a median perception share of 0.2901 across the 14 models scored under the coordinate condition, with the post-perception residual being the larger of the two for 13 of them The raw component moves the other way, R3 − R4 correlating with R4 at ρ = −0. p. 4

Where the paper says it

The decomposition gives a median perception share of 0. 2901 across the 14 models scored under the coordinate condition, with the post-perception residual the larger of the two for 13 of them, the two exemplars of the attribution ladder spanning that range at 0.

6439, so the strongest pixel performers are those for whom exact geometry adds least even though the share it closes shows no trend The finding with the widest reach for AI-assisted materials design is fabrication

Design Rules and Transferability

The instrument yields design rules for benchmark builders, stating which camera placements survive extraction noise, that placement rather than view count sets the noise budget, and where a pipeline’s own reading becomes the binding ceiling The construction transfers to any benchmark whose forward rendering can be written down and inverted Adding a view never destroys identifiability, and a camera aligned with a lattice vector separates no atoms differing by multiples of it The oracle answers from exact centroids, so the ceiling states what the geometry determines rather than what a pixel reader recovers

Extraction Stage Audit Primitive

The fabrication finding of the main text uses a 206-structure denominator: the four gate-excluded structures, listed with the gate in the Reproducibility record, count as unanswered in the full-sample figure A two-stage pipeline whose intermediate cannot be checked will attribute fabrication to reasoning, and only a model-free reference distinguishes the two The audit primitive these pipelines currently lack is the model-free reference for any stage whose forward map is known

Final Conclusion on Model Deficit

The ceiling-to-model gap is bounded rather than estimated, and the ceiling is certified per sample rather than assumed Read against that ceiling, most models remain limited by what survives the removal of perception The reading that perception is the binding constraint on multimodal materials tasks is therefore a statement about particular models rather than about the task The information is there to be read is settled independently by the supervised pixel baseline, which lands above every vision–language model and close to the ceiling using only the benchmark’s own renders

Methodology for Oracle Operation

Algorithm 1 Geometric oracle O(s, ) is deterministic and assumes ideal extraction, so R1 is a bound from the projection geometry rather than from the pixels The constant is a property of the camera set, κ = 2.7321 here, and it diverges as two directions approach parallel The oracle answers from exact centroids, so the ceiling states what the geometry determines rather than what a pixel reader recovers

Phantom Set Analysis

Definition 7 Phantom set, componentwise, states that a species-labelled point (p, z) lies in Φdelta(s, ) if it is further than δ from every same-species atom The theorem’s merge clause is performed at δ = τ: this is the single implementation tolerance δ of the oracle algorithm, tied to the label map’s symmetry tolerance τ rather than a third, independent quantity Every emptiness claim here and in the main text is therefore stated at the δ it was measured at The mean phantom excess falls monotonically with view count, from 4.911 at two views to 0. p. 6

Where the paper says it

Mean 𝑅1 at the certified tolerance is 0. 9214 at two views and 0.

480, 0.048 and exactly 0. p. 0
Where the paper says it

qa A B S T R A C T Multimodal evaluations cannot say whether a vision–language model misread an image or misreasoned about it, because every existing method for separating the two places a second model in the loop. We introduce the render ceiling, a model-free reference for benchmarks built by rendering known objects: inverting the frozen cameras and re-solving cross-view correspondence recovers exactly the answer the images support.

000 at five, where all 210 structures recover the true atom count The mean phantom excess falls monotonically with view count, from 4. p. 8
Where the paper says it

That the renders lack the answer is excluded by 𝑅1 and by the occluder classification, which finds symmetry recovery identical whether or not informative occluders are removed while the atom-count match falls from 204 to 194 of 210, so the intervention does reach the reconstructor (Fig. That models rely only on composition-to-label mappings is excluded by an image-absent control carrying the identical prompt with the chemical formula and no renders: every scored model collapses toward chance, mean 0.

911 at two views to 53. p. 6
Where the paper says it

Mean 𝑅1 at the certified tolerance is 0. 9214 at two views and 0.

7 ms at 80 atoms

Extraction Tolerance Frontier

The extraction-tolerance frontier reported in the main text has one non-monotone point: at n = 1950, δ = 0.005Å costs one structure where δ = 0. p. 4

Where the paper says it

8785 there on the 1950 certified structures. The loss is a label-tolerance failure inside the symmetry algorithm rather than a failure of the geometry: every structure it costs recovers the correct atom count and has its closest same-species pair at more than eleven times 𝛿, which rules out merging of distinct atoms.

01Å costs none The ceiling is 1. p. 4
Where the paper says it

The loss is a label-tolerance failure inside the symmetry algorithm rather than a failure of the geometry: every structure it costs recovers the correct atom count and has its closest same-species pair at more than eleven times 𝛿, which rules out merging of distinct atoms. Identifiability from the renders is thus total, and what an actual pipeline achieves is a property of its extraction stage; every ceiling-to-model gap below is wider at the certified reading than at the released tolerance, so no comparison depends on which point of the frontier is chosen.

0004 at δ = τ; the tighter δ = tau setting produces one under-merging failure on the scaled sample, whereas larger values of δ introduce over-merging

Model Performance Comparison

The strongest scored model Gemini 3.6 Flash reaches R3=0. p. 5

Where the paper says it

Role label Model Conditions and sample Accuracies The strongest scored model (Strongest) Gemini 3. 6 Flash Ladder (𝑅3/𝑅4); also the newer arm in the one-generation comparison (same model, same label, not a second arm) 𝑅3=0.

8524; R4=0. p. 4
Where the paper says it

6V, and none reaches the oracle: the best is 0. 8524 against 1.

7333 The weakest of the fourteen by R1 (oracle ceiling) is GPT-4. p. 5
Where the paper says it

8524; 𝑅4=0. 7333 The weakest of the fourteen by 𝑅3 (Weakest) GPT-4.

1 mini Coordinate condition, with R3/R4 = 0. p. 5
Where the paper says it

7333 The weakest of the fourteen by 𝑅3 (Weakest) GPT-4. 1 mini Coordinate condition (𝑅3/𝑅4) 𝑅3=0.

4143 and 0. p. 5
Where the paper says it

1 mini Coordinate condition (𝑅3/𝑅4) 𝑅3=0. 4143 (minimum of the 14 scored); 𝑅4=0.

3667 Perception dominance is therefore a property of the top of the roster, not of the task

Statistical Robustness

The test families, correction procedure and reporting conventions are specified in the Statistics and reproducibility subsection of Methods Every load-bearing claim survives correction The median perception share is reported with a percentile bootstrap over models: against the exact R1 the median is 0.2901 with 95% confidence interval [0. p. 7

Where the paper says it

(b), Perception share of the deficit against 𝑅4; the dotted parity line marks equal perception and post-perception components. The drawn median and its bootstrap 95% confidence interval (band) are computed against the exact ceiling 𝑅1 = 1.

1492, 0.3576] The label-certification lower bound is a one-sided exact bound, 0. p. 9
Where the paper says it

The residual is an upper bound and was tested as such: it may include long-list handling rather than symmetry reasoning, and regressing per-structure 𝑅3 correctness on atom count, 13 of 14 models show a negative association that is not symmetry difficulty in disguise. One further reading belongs to the original draw rather than to the method: thirteen of the seventeen scored arms fall below the shape-free baseline of 0.

051/224 = 0.9865 at k = 224 agreements, conditional on the four-structure quarantine The full family is 26 hypothesis tests with numeric p-values, the ten headline tests plus the render-convention probe’s 16 individual paired comparisons

Reproducibility and Artifacts

The released artifact, the benchmark records, the per-structure predictions for every arm and the preregistration of each analysis are openly available at https://github.com/KurbanIntelligenceLab/render-ceiling The repository carries the code implementing the geometric oracle, the render protocol, the labelling pipeline and every analysis reported here, under the MIT licence Artifact disclosures. p. 0

Where the paper says it

The flagship multimodal benchmark for the Preprint • Kurban Intelligence Lab • https://github. com/KurbanIntelligenceLab/render-ceiling 1

Two documentation mismatches could mislead a model or a reader into reasoning about bond geometry that is not in the image The analytic oracle never touches a rendered pixel, so its numbers are unaffected Runtime and complexity of the oracle is O(V 2n 2) in all, dominating the O(n) merge The census recovers every atom exactly on all 1,950 structures, consistent with the identifiability theorem The full five-camera set reproduces R1 = 200/210 = 1. p. 2
Where the paper says it

For natural photographs no better reference is available, since no procedure recovers a photograph’s information content without a model: recovering scene geometry from a single image is ill-posed, an infinite family of three-dimensional scenes projecting to the same picture, so every practical reader supplies a learned prior in place of the missing constraint [31]. Rendered scientific figures are different, and that difference is the opening this work exploits: a crystal-structure render is an orthographic projection [32] of a structure whose coordinates are known exactly, under a camera set the experimenter chose.

0 at δ = τ through this independent implementation The same geometry yields design rules a benchmark builder can apply before release The construction transfers to any modality whose forward rendering can be written down and inverted

Key Findings on Model Strength

The natural strengthening, that the share rises with model strength, is not supported: across the 14 models the share and pixel accuracy are unrelated by rank correlation The raw component moves the other way, R3 − R4 correlating with R4 at ρ = −0.6439, so the strongest pixel performers are those for whom exact geometry adds least Perception dominance is therefore a property of the top of the roster, not of the task The reading that perception is the binding constraint on multimodal materials tasks is therefore a statement about particular models rather than about the task The finding with the widest reach for AI-assisted materials design is fabrication The agents are assembled from exactly such intermediates, a structure read from a figure and handed to a property predictor or a synthesis planner, checked, where it is checked at all, by a learned verifier that cannot separate a fabricated intermediate from a mistaken one

Camera Placement vs. View Count

The conditioning constant orders camera placements rather than setting the noise budget Adding a view never destroys identifiability, and a camera aligned with a lattice vector separates no atoms differing by multiples of it Four separated cameras empty the phantom set on this protocol with the fifth adding reconstruction fidelity The mean render ceiling R1 under two-, three-, four- and five-camera subsets of the frozen protocol, at the certified tolerance δ = τ = 0.01Å and at the released tolerance δ = 15τ; error bars span the minimum and maximum over camera subsets of that size View count behaves as the first design consequence predicts, non-decreasing as views are added The conditioning constant, evaluated over the ten independent pairs of the five-camera set, is 1. p. 6

Where the paper says it

(a), Mean render ceiling 𝑅1 under two-, three-, four- and five-camera subsets of the frozen protocol, at the certified tolerance 𝛿 = 𝜏 = 0. 01Å and at the released tolerance 𝛿 = 15𝜏; error bars span the minimum and maximum over camera subsets of that size.

3280 (mean 1. p. 5
Where the paper says it

6905 at 𝐾=8; 139/210=0. 6619 at 𝐾=3 Supervised reference arm Qwen3-VL 8B Instruct (same base as the fine-tuned model arm, supervised-only recipe, no GRPO) Supervised LoRA, three seeds Mean 0.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Model-free reference for benchmark evaluation: Introduce a model-free render ceiling, which is a model-free reference for benchmarks built by rendering known objects: inverting the frozen cameras and re-solving cross-view correspondence recovers exactly the answer the images support. This provides an exact, per-sample reference that separates perception from reasoning deficits.

  2. Attribution of failures: The system can distinguish between perception failure and reasoning failure by measuring the post-perception residual which is defined as S(m) = R1 - R3(m), allowing for a clearer separation of where the model's deficit lies.

  3. Benchmark building rules: The instrument yields camera-placement rules for benchmark builders, stating which camera placements survive extraction noise, that placement rather than view count sets the noise budget. This allows researchers to design benchmarks robustly by selecting camera configurations based on these derived rules.

  4. Fabrication detection: The system exposes a failure class where a strong model emits syntactically perfect coordinate lists that match almost nothing, which downstream accuracy would book as bad reasoning, serving as an audit primitive for agentic pipelines.

  5. Model performance assessment: The ceiling allows for a decomposition of model deficits into perception and reasoning, showing that the post-perception residual exceeds the perception component for 13 of the 14 models, indicating where to focus training efforts.

Abstract

Vision-language models (VLMs) often predict plausible structural labels from images without recovering the three-dimensional geometry behind them. End-to-end scores cannot tell whether an error comes from ambiguous views, lost visual information, or flawed reasoning. Multi-view 3D reconstruction faces its own trade-off: rigid symmetry priors erase genuine distortions, while unconstrained triangulation absorbs measurement outliers. We separate visual perception from structural reasoning through an inspectable 3D geometric intermediate. Its foundation is a model-free render ceiling: an exact reference for what calibrated, species-labeled projections determine, obtained by ruling out projective phantoms. Across a 1,950-structure census, every tested combination of four or five calibrated views uniquely determined the 3D sites from complete, species-labeled centroids. Against this reference, replacing images with exact text coordinates lowers the error rate of all 17 evaluated VLMs on crystals. The three frontier models gain 10.0-19.5 percentage points yet still miss 8.1-21.0% of cases. To recover geometry from imperfect observations, our symmetry-aware split reconstruction jointly estimates a symmetric base, coherent 3D physical displacements, and view-specific 2D corruption, adaptively selecting between atomic and collective distortion modes. Relative to Huber triangulation, it reduces coordinate error by 36.7% under Gaussian noise and 41.2% under corruption. Adaptive mode selection further cuts collective error up to 30.4% relative to atomic split across two external cohorts without sacrificing localized fidelity. Given reconstructed rather than Huber coordinates, downstream models improve. Geometric recovery thus becomes an explicit, testable foundation for structural understanding.

Sources

Related papers