Neural networks that represent 3D shapes as continuous functions usually report, for any point in space, its signed distance to the surface: negative inside, positive outside. Meshing such a network is easy because the surface is wherever the sign flips. That only works for closed, "watertight" objects. Garments, leaves, and car interiors have open sheets with no inside, so the field instead predicts an unsigned distance, which is never negative. With no sign flip to locate, the standard meshing algorithm fails, and the replacements are slow, fragile, or need a human to say whether the shape is open. This proposal asks a diagnostic question taken from the interpretability literature. A network trained only on unsigned distance must still compute something different for a point just above a sheet and a point just below it. Does it carry, in its frozen hidden activations, a linearly readable "which side am I on" signal that its output throws away?
If the answer is yes, a single linear readout trained once on a frozen, shape-conditioned unsigned-distance network would supply the missing sign. Ordinary Marching Cubes could then mesh open surfaces in seconds, with no second network and no iterative shell-shrinking. If the answer is no, that is also informative. It would be direct evidence for the claim, so far only asserted, that distance and orientation cannot coexist in one network, which is the motivation behind decoupled designs. In two sentences: we train a linear probe on hidden activations of a frozen unsigned distance network to predict local surface side, and test whether that probe's sign makes cheap, standard meshing of open surfaces work. The comparison that matters is against classifiers that see only the network's outputs and spatial gradients.
Core Claim
Our hypothesis is that the penultimate layers of a shape-conditioned unsigned distance network encode a locally consistent orientation, or "phase," that is linearly decodable with high accuracy within a thin band around the surface. We further expect this signal to be more reliable than what can be recovered from output distances and gradients. The motivating analogy comes from probing work showing that internal states hold information the output suppresses. Linear probes on intermediate language-model states detect wrong answers with AUC 0.85–0.95 even when the generated answer is wrong (LLMs Know More Than They Show). Vision encoders hold 2D coordinates almost perfectly before the language layers lose them (Diagnosing Bottlenecks in Data Visualization Understanding). CNNs encode shape at 77.9% linear accuracy while classifying by texture (Origins and Prevalence of Texture Bias).
An unsigned distance field has the same structure. The final output is symmetric across a sheet by construction, but nothing forces the hidden features to be symmetric. Gaps 1–3 in the research file note that this has never been checked in 3D coordinate networks. The same probing literature also constrains the design. Truth probes do not transfer across task families (LLMs Know More Than They Show), so we must test transfer across shape categories rather than assume a universal orientation direction.
Research Program
The object of study is a shape-conditioned unsigned distance network in the style of Neural Unsigned Distance Fields (NDF). NDF uses a point-cloud encoder and a decoder shared across shapes, which predicts unsigned distance to the nearest surface. A shared decoder matters here because one probe trained on training shapes can then be applied unchanged to unseen test shapes. We train NDF as published and freeze it. We also use a per-shape variant trained with the projection losses of LevelSetUDF, to check whether a smoother zero level set changes what is decodable.
Labels. From ground-truth meshes we sample query points in a band of width $\delta$ around the surface. Each point is labelled by the side of its closest surface triangle. Open sheets have no global inside, and ShapeNet normals are inconsistently oriented, so we avoid an absolute sign. Instead the probe $s(x) = w^\top h_\ell(x)$, where $h_\ell(x)$ is the frozen layer-$\ell$ activation at point $x$, is trained with a pairwise loss on nearby point pairs. The loss rewards equal signs for pairs on the same side of the local sheet and opposite signs for pairs on opposite sides. This makes the target invariant to the arbitrary global flip of each shape.
Layer-wise analysis. We fit probes at every layer and plot decodability against depth and distance band. This follows the depth analyses in Exploring Concept Depth and How is BERT surprised?. We also estimate the dimensionality of the orientation subspace with linear discriminant analysis, as On Entity Identification did for entities. This directly tests the tension in Gap 4: whether distance and phase occupy separable linear directions inside one network.
Use in meshing. The probe yields a signed proxy $\phi(x) = d(x)\tanh(\beta s(x))$, where $d$ is the network's unsigned output and $\beta$ is a fixed sharpness. Standard Marching Cubes runs on $\phi$.
The sign field around an open sheet must be discontinuous somewhere beyond the sheet's boundary edge, analogous to a branch cut. Crossings in that region would produce spurious faces, so we discard any crossing whose interpolated $d$ exceeds a small threshold. This is the same masking logic MeshUDF-style extractors use.
The construction resembles Metric–Phase Fields (MPF), which trains a second network $\theta(x)$ end-to-end with Eikonal, Laplacian, and alignment losses. Our phase instead comes for free from a frozen network. It also resembles Neural Surface Detection (NSD), which classifies each grid cell into one of 128 sign configurations from 32 external numbers: eight corner distances and 24 gradient components. Our classifier reads internal features instead of those external numbers.
Baselines
The naive reference point is a linear probe on output-only features: the predicted distance, its autodiff gradient, and the query coordinates passed through Fourier features. It is fit with the same pairwise loss, data, and band. A second naive control is the same probe on activations of a randomly initialized network of identical architecture. This is the standard check that the probe, rather than the trained model, is doing the work. Hidden-state probes must beat both controls by a clear margin in pairwise side accuracy within $\delta$ of the surface, or the claim fails.
The strongest competitor on the representational question is NSD (Neural Surface Detection). It uses exactly the external evidence our method claims to improve on. NSD reports a median Chamfer distance of 2.52 on ABC at $128^3$, versus 2.97 for MeshUDF and 44.6 for DCUDF. For a fair comparison we retrain NSD's classifier on the same frozen NDF fields, the same training shapes, and the same grid, with a hyperparameter search budget equal to the one used for the probe. Beating it means a lower median Chamfer distance on ABC at $128^3$ together with fewer boundary micro-cracks, measured as counts of boundary edges not present in the ground truth. Matching it with a single linear map and no gradient evaluation would count as a partial result, a speed and simplicity win rather than an accuracy win.
For end-to-end reconstruction of thin structures, the strongest method is MPF (Metric–Phase Fields). It reports a Chamfer distance of 0.672 on its synthetic thin-structure set, against 4.977 for unsigned and 5.591 for signed baselines, and an extraction time of about 0.13 minutes. We run MPF on the same inputs. We expect to approach, not beat, its accuracy, while removing its second network and loss tuning; we report training cost alongside accuracy.
For mesh integrity on garments we compare against DCUDF (Robust Zero Level-Set Extraction). DCUDF reports 0% non-manifold vertices on DeepFashion3D, compared with over 18,000 per model for MeshCAP. It sets a ceiling we must match on manifoldness, and its required user flag for open versus closed shapes is the manual step we aim to remove.
Two ablations isolate the key ingredient. The first concatenates hidden features with NSD's external features, which tells us whether the hidden features add information or merely duplicate it. The second replaces the linear probe with a two-layer MLP probe; a large gap between the two would mean the probe is computing orientation itself rather than reading it from the network.
Why not just…?
Why not just use external gradients, as NSD does? This is the real threat. The gradient of an unsigned field flips direction across the sheet and is undefined at it (NDF; LevelSetUDF). NSD itself reports micro-cracks and jagged open boundaries. Still, gradients sampled at cell corners slightly off the surface may carry all the side information the network has. The concatenation ablation decides this. If adding hidden features to NSD's inputs leaves accuracy unchanged, the internal signal is redundant, and we would reframe the contribution as a cheaper but equivalent readout.
Why not just train a sign head jointly, or adopt MPF's two fields? Joint training changes the model and requires the composite-loss balancing that MPF acknowledges is brittle. Our claim concerns frozen, already trained fields. We nevertheless train a jointly supervised sign head on NDF as an upper bound. A probe that reaches most of that bound shows the orientation was already present in the network rather than created by the extra supervision.
Why not just rely on post-hoc meshing such as DCUDF or MeshUDF? These methods take 3.6–15.2 minutes per mesh according to Metric–Phase Fields. DCUDF needs a user to declare open or closed, and it reached a median Chamfer distance of 44.6 on ABC in NSD's evaluation. We time every pipeline on the same hardware. We also test whether the probe's magnitude near the surface can predict open versus closed status per component, which would replace DCUDF's manual flag.
Proposed Experiments
We train and evaluate on ShapeNet Cars (ShapeNet), whose interiors provide nested open sheets; NDF reports Chamfer-$L_2$ 0.05 there. We also use ABC (ABC) and DeepFashion3D for open garment boundaries. Transfer is tested by training the probe on cars and evaluating it on garments and on Multi-Garment Net shapes (Multi-Garment Net); no reported result was found on Lacuna for any method on MGN. We report metrics at two levels. The probe-level metric is pairwise side accuracy as a function of distance band and layer. The mesh-level metrics are Chamfer distance, non-manifold vertex count, spurious boundary faces, and wall-clock extraction time.
The first decisive check is representational. Within 0.5% of the bounding box from the surface, a linear probe on hidden features must beat both the output-feature and random-network probes. If it does not, we stop and report that unsigned networks discard orientation internally, which would support MPF's architectural choice. The second check is practical. Probe-signed Marching Cubes must achieve a median Chamfer distance at or below NSD's 2.52 on ABC at $128^3$ under the matched protocol, keep non-manifold vertices near DCUDF's 0% on DeepFashion3D, and run at Marching Cubes speed. The third check is cross-category transfer. Given that truth probes in language models fail to transfer across tasks, a probe that survives the shift from cars to garments would be the strongest evidence that orientation is a general internal variable rather than a feature specific to one category.
References
- LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations
- Diagnosing Bottlenecks in Data Visualization Understanding by Vision-Language Models
- The Origins and Prevalence of Texture Bias in Convolutional Neural Networks
- Neural Unsigned Distance Fields for Implicit Function Learning
- Learning a More Continuous Zero Level Set in Unsigned Distance Fields through Level Set Projection
- Exploring Concept Depth: How Large Language Models Acquire Knowledge and Concept at Different Layers?
- How is BERT surprised? Layerwise detection of linguistic anomalies
- On Entity Identification in Language Models
- Metric—Phase Fields: Decoupling Distance and Sign for Thin-Structure Reconstruction from Unoriented Point Clouds
- Neural Surface Detection for Unsigned Distance Fields
- Robust Zero Level-Set Extraction from Unsigned Distance Fields Based on Double Covering
- ShapeNet: An Information-Rich 3D Model Repository
- ABC: A Big CAD Model Dataset For Geometric Deep Learning
- Multi-Garment Net: Learning to Dress 3D People from Images