{"id":"d4fd0a85-2782-4b2f-ab4d-05a87baeb179","arxiv_id":"2505.21319","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A 32^3 grid of 13-parameter RBF-weighted polynomial functions fits 3D signed distance fields as accurately as neural hash-grid and octree methods with far fewer parameters.","lead":"These researchers built a compact, non-neural way to represent 3D shapes by blending simple polynomial functions over a regular grid, with radial-basis weighting tuned during optimization. On standard signed-distance benchmarks it matched or slightly beat much bigger neural representations while using tens of times fewer numbers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mean-shift offset initialization is a load-bearing dependency: near-surface SDF error more than doubles without it, and sensitivity to surface-point quantity or noise is untested.","rationale":"The reader's weakest_assumption identifies the same risk I consider most load-bearing. Table 4's Full-3/Full-4 pair is the only ablation isolating mean-shift initialization, and the change in Near-AE (6.499 to 3.772) is one of the largest single effects in the ablation table. Because the abstract and conclusion claim SDF fitting quality, not merely low Chamfer distance on extracted meshes, this dependency should be quantified before the method is used in settings where clean surface samples are not guaranteed. The concern is addressable with a straightforward N/noise sweep and does not invalidate the representation's core idea, so rejection is not warranted. The paper's own limitations section is candid about speed and scalability, and the conditional verdict already captures the need for additional robustness evidence. I therefore recommend keeping the reader's verdict unchanged.","tokens_in":16877,"tokens_out":8201,"duration_ms":96631,"concrete_test":"Run Config-Full-4 on Armadillo and Dragon with N in Eq. 17 taking values in {0, 256, 1024, 4096, 16384, 65536}, and with surface points perturbed by isotropic Gaussian noise of standard deviation {0, 1e-4, 1e-3, 1e-2} of the bounding-box diagonal; report CD, Near-AE, and Near-IOU. If Near-AE stays within 10% of the value at N=16384 and clean points for N<=4096 and sigma<=1e-3, the concern is retired. Otherwise, the near-surface quality claim should be explicitly qualified by the requirement for dense, clean surface samples.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 4 (Config-Full-3 vs Config-Full-4) shows that enabling mean-shift initialization (Eq. 17) improves Near-AE from 6.499 to 3.772 and Near-IOU from 95.218 to 97.119, under the same 32^3 x 13 parameter budget. Thus the default model does not achieve its near-surface accuracy from the representation alone; it inherits a requirement for 16,384 clean ground-truth surface samples, a number the paper fixes 'in almost all experiments' without a sensitivity study. This matters because the central claim concerns SDF fitting quality, and near-surface accuracy is exactly the regime where the offset mechanism operates. The paper never varies N in Eq. 17, never perturbs the surface points, and does not compare against competitors that avoid this initialization. If the input is a noisy point cloud or a latent geometry without explicit surface samples, the reported near-surface advantage is not guaranteed. Chamfer distance on extracted meshes is less sensitive to this effect, but the paper also claims function approximation quality, so the missing robustness evidence is a real gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a non-neural signed distance function (SDF) representation called efunc, in which a normalized radial-basis-function weighted sum of polynomial functions is defined over a regular grid, augmented by learnable offset keys clustered near the surface. The authors derive analytic gradients, implement fused CUDA forward/backward kernels that reduce memory and time relative to automatic differentiation, and evaluate SDF fitting on nine meshes against FFN, NGLOD, Instant NGP, and DiF-Grid. The central empirical claim is that at 32^3×13 parameters (about 0.43M floats) the representation reaches Chamfer distance comparable to or better than MLP/octree/hash-grid baselines that use an order of magnitude more parameters. The paper also presents a taxonomy of function representations, ablations, and applications to frequency decomposition, normal estimation, and shape manipulation.","tokens_in":17104,"tokens_out":6311,"duration_ms":64013,"significance":"If the empirical claims hold, the compact explicit representation is a useful alternative to neural SDFs for applications needing fast analytic gradients, low memory, and interpretable grid structure. The paper's strengths are the algebraically correct gradient derivations (Eqs. 13–14 and 20–22), the complete forward/backward algorithms that are implementable from the text, and an order-of-magnitude reduction in memory compared to a naive PyTorch implementation. The proposed framework taxonomy in Section 2 is a clearly presented contribution that situates the method among classical and neural representations. The principal weakness is that the headline comparison rests on single-run Chamfer values on nine meshes and on a mean-shift initialization whose sensitivity is not examined; these gaps need to be addressed before the 'comparable or superior' claim can be considered established.","major_comments":[{"comment":"The central claim of 'comparable or superior performance' is not statistically supported. Each entry is a single run on one mesh, with no error bars, no repeated-seed results, and no significance test. The mean Chamfer distances separating Ours (32^3×13, 9.227) from NGLOD (9.273) and I-NGP (9.288) are far smaller than the per-shape variations and likely within noise. Please report means and standard deviations over at least three independent optimization runs (or across sampling seeds for the 100k-point Chamfer evaluation) and state whether the differences are significant.","section":"§4.2, Table 3"},{"comment":"The mean-shift initialization is a load-bearing component whose robustness is untested. Config-Full-4 with initialization improves Near-AE from 6.499 to 3.772 and Near-IOU from 95.218 to 97.119 relative to Config-Full-3 under the same parameter budget. The paper fixes N=16,384 ground-truth surface points and bandwidth 100 'in almost all experiments' without varying N, perturbing the surface samples, or testing noisy input. Because the method is advertised for SDF fitting in general and for future generative use, the sensitivity of near-surface accuracy to the quantity and quality of surface points should be quantified; otherwise the reported near-surface advantage is only known to hold for clean, dense mesh samples.","section":"§3.3, Eq. (17), Table 4"},{"comment":"The claim of 'significantly fewer parameters while achieving comparable approximation quality' is overstated relative to the evidence across resolutions. At 16^3×13 parameters the mean Chamfer distance is 9.659, which is worse than all four neural baselines (9.39–9.37); only the 32^3 grid is comparable or slightly better. The paper should scope the claim to the tested 32^3 resolution and, ideally, report a parameter-accuracy curve showing the resolution at which parity with each competitor is reached.","section":"§4.2, Table 3 and Conclusion"},{"comment":"The function-decomposition application is not reproducible as written because Eq. (25) uses cos(b·π·q) with q∈R^3, but the expression is written as if q were a scalar. Please specify whether the cosine is applied per coordinate (and how the resulting vector is reduced to a scalar S(q)), and reconcile the statement that each O_b uses 16^3×4 floats with the total of 20480×B parameters, which appears to include the scale parameter β.","section":"§4.4, Eq. (25)"}],"minor_comments":[{"comment":"The label 'mesh-shift initialization' appears to be a typo for 'mean-shift initialization'.","section":"Table 4"},{"comment":"Please state where the N=16,384 surface points are sampled from (e.g., uniformly on the ground-truth mesh or from a dense point cloud) and provide a rationale or sensitivity check for the fixed bandwidth 100.","section":"§3.3, Eq. (17)"},{"comment":"The experimental protocol should state the number of training iterations, the random seed for point sampling, and the hardware used, since the reported 20-minute training time is not directly comparable to the competitor timings cited in the limitation paragraph.","section":"§4.1"},{"comment":"The sentence 'Our results are getting better when the resolution is increased' is informal for a journal; please rephrase.","section":"§4.2, first paragraph"},{"comment":"In Eq. (9) the parameter vector is written as θ={k_i,β_i,v_i}∈R^{R×R×R×2}, but the keys k_i are grid points and the expression omits the keys' coordinates; please clarify which entries are trainable and which are fixed.","section":"§3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a graphics journal and the technical core appears sound. The main risk is empirical: the headline comparison is based on a small number of single-run evaluations and on a mean-shift initialization whose sensitivity is not assessed. I would encourage the editor to ask for the reported robustness experiments and, if possible, a code release, since the CUDA implementation is a stated contribution and reproducibility would strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: efunc is a legitimate, practically relevant non-neural SDF representation—globally normalized RBF interpolation of per-grid polynomial functions with offset keys—and it is engineered carefully enough to be implementable and useful. The headline claim (\"comparable or superior to octree/hash-grid methods with far fewer parameters\") is plausible but not fully proven: the evidence is nine meshes, one run each, with Chamfer differences that are small relative to likely variance.\n\nWhat is actually new: the specific O+Δ parameterization, Eq. (16), with learnable β scales, polynomial values, and offset keys initialized by mean shift. That combination is not in the cited literature (MPU/IMLS use fixed local support; M-SDF uses trilinear features). The fused CUDA forward/backward kernels are a real engineering contribution, cutting memory/time to about 1/10 of PyTorch autodiff, and the paper gives complete algorithms and correct gradient derivations. The ablation in Table 4 is honest and useful, and the limitations section is refreshing: they admit it is slower than Instant NGP (20 min vs 5 min) and cannot scale to 128³.\n\nSoft spots, in proportion:\n\n1. The empirical comparison is thin. Table 3 shows mean Chamfer distance 9.227 for 32³ Ours vs 9.273–9.390 for Instant-NGP/NGLOD/FFN/DiF-Grid. That is a roughly 0.5–1.7% difference, with no error bars, no repeated runs, no significance testing. \"Comparable or superior\" is fair as an ordering statement, but it is not established that the difference is real rather than noise. Multiple seeds and reported variance would fix this.\n\n2. The mean-shift initialization is load-bearing. Table 4, Config-Full-3 vs Full-4, shows enabling it drops Near-AE from 6.499 to 3.772 and raises Near-IOU from 95.218 to 97.119. The default setup uses N=16384 surface points that are never varied or perturbed. If the input is a noisy point cloud or a latent geometry without explicit surface samples, the near-surface accuracy benefit may not transfer. The paper does not test this. That is a real gap, but not a fatal one.\n\nThe stress-test note is on target. The paper's own limitations also confirm that the strong point is compactness and analytic gradients, not speed.\n\nWho this is for: anyone working on SDF fitting, shape representation, or generative geometry modeling. It is worth a serious referee: the method is new, the derivations check out, and the gaps are fixable. I would send it to review and ask for robustness data, not desk-reject.","headline":"A genuinely useful compact non-neural SDF representation with a clean parameterization and honest ablations, but the headline accuracy claim rests on thin single-run evidence and a load-bearing offset initialization whose sensitivity is untested.","tokens_in":17643,"tokens_out":2498,"would_cite":true,"duration_ms":27059,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D05","65D17","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A $32^3$ grid with 13 floats per cell — softmax-weighted RBF-interpolated polynomial fields plus learned surface offsets — fits signed distance functions as accurately as million-parameter neural methods.","keywords":["3D shape representations","radial basis functions","signed distance functions","function interpolation","polynomial fields","compact representation","CUDA optimization","neural representations"],"falsifier":"Fit the same nine test meshes with the $32^3\\times 13$ representation with the mean-shift initialization disabled or fed corrupted surface samples, then measure near-surface Chamfer distance and Near-AE; a sharp degradation would confirm that the accuracy claim depends on external surface knowledge rather than on the RBF-polynomial form alone. Alternatively, compare against a hash-grid baseline under an equal parameter budget on a class of shapes whose ground-truth surfaces are not directly sampled, such as latent generative priors, where the offsets could not be initialized.","tokens_in":16637,"feed_emoji":"🧊","tokens_out":14658,"duration_ms":126409,"temperature":0.7,"pith_summary":"This paper tries to show that high-quality 3D signed distance function (SDF) fitting does not need neural networks or hierarchical data structures. It proposes a compact representation: on a regular $32^3$ grid, each cell stores a polynomial field plus an inverse width, and interpolation is performed by softmax-normalized radial basis weights so every grid point contributes at every query. A second set of grid keys with learned offsets concentrates near the surface and captures detail. On nine test meshes, the $32^3\\times 13$ variant achieves the lowest average Chamfer distance among the compared methods while using roughly 0.43 million floats, versus millions for octree- and hash-grid neural baselines. The paper also reports a fused CUDA implementation that uses about a tenth of the time and memory of automatic-differentiation frameworks, making the method practical on desktop GPUs.","feed_headline":"0.43M floats rival million-parameter SDF fitting","feed_subtitle":"Regular-grid RBF polynomials plus surface offsets match or beat neural baselines with far fewer parameters.","key_machinery":"The central machinery is the normalized RBF interpolation of polynomial Value functions on a regular grid, $O(q)=\\sum_i \\mathrm{softmax}(-\\beta_i\\|q-k_i\\|^2) f(q-k_i;\\phi_i)$, where $f$ is a truncated Taylor polynomial (degree one in the default setting) and $\\beta_i$ is a learned inverse width. Each grid cell contributes a whole function rather than a scalar, and because the softmax weights have global support, every key contributes at every query, so no neighborhood search or hierarchical structure is needed. The second mechanism is the offset set $\\{k_i+\\Delta_i\\}$, initially placed near the surface by mean shift, then learned, which adds local detail without changing the grid topology. Parameters are fitted with AdamW on mean squared error, and the forward and backward passes are fused into a single CUDA kernel that keeps only the softmax denominator from the forward pass, cutting memory from $\\Theta(IJ)$ to $\\Theta(J)$.","core_discovery":"The central claim is that replacing scalar grid values with polynomial functions and interpolating them with globally supported softmax-normalized radial basis weights yields an SDF representation whose approximation quality is comparable or superior to state-of-the-art neural, octree, and hash-grid representations at a small fraction of the parameter count. The argument is carried by the combined $O{+}\\Delta$ representation of Eq. (16): one regular-grid RBF-polynomial field represents the bounding volume, and a second set of offset keys, initialized by mean shift on surface samples and updated during training, represents the near-surface region. The default configuration stores $32^3\\times 13$ floats and achieves the best average Chamfer distance in the comparison. The paper further claims that the form is analytically differentiable, supports normal estimation through the gradient formula of Eq. (13), and permits direct grid-level shape manipulation.","pith_inferences":["If the accuracy-per-parameter behavior carries to higher resolutions once kernel limits are lifted past $128^3$, this representation would be a natural latent-space prior for diffusion or autoregressive 3D generation, where per-shape codes must stay compact and differentiable.","The global-support softmax weighting is mathematically the same shape as attention, so the representation could generalize to irregular or learned key sets such as point clouds or octree leaves without changing the loss; the paper does not explore that direction.","The mean-shift initialization ties accuracy to a fixed sample of ground-truth surface points; a generative setting with latent geometry would need an amortized predictor for offsets, and the paper does not evaluate that regime.","A direct extension would swap the polynomial Value for other analytic families such as trigonometric or exponential functions, potentially trading parameter count for frequency content on open or high-frequency surfaces; the paper leaves this untested."],"forward_implications":["SDF fitting can run on a plain regular grid of about 0.43M floats, matching or beating octree- and hash-grid neural methods in average Chamfer distance on the tested meshes.","Surface normals follow from one analytic forward pass via Eq. (13), without finite differences or a separate network.","Because the grid is regular and the weights are global, two shapes can be combined by splicing halves of their grids, a manipulation the paper demonstrates directly.","The fused CUDA kernels cut training memory and time to roughly a tenth of a naive automatic-differentiation implementation, enabling training in about 16 MB at $32^3$ resolution on a desktop GPU.","The same functional form composes with Fourier-cosine weights (Eq. 25) to decompose a shape into frequency bands, and the band-limited partial sums reconstruct the surface progressively."],"supporting_citations":[{"why":"Provides the classical radial-basis-function interpolation baseline for SDF reconstruction that the paper's normalized RBF formulation extends and compares against.","marker":"Carr et al. [2001]"},{"why":"Introduces multi-level partition-of-unity implicits with quadratic polynomial pieces, the key prior function-interpolation method that the polynomial RBF Values are positioned against.","marker":"Ohtake et al. [2003]"},{"why":"Supplies the moving-least-squares function-interpolation baseline (linear functions over RBF weights) that a configuration of the paper roughly matches.","marker":"Kolluri [2008]"},{"why":"The multiresolution hash-grid neural representation that is the main parameter-heavy competitor in the Chamfer distance comparison.","marker":"Müller et al. [2022]"},{"why":"The octree-based neural SDF method used as a state-of-the-art competitor and as an example of local-support interpolation.","marker":"Takikawa et al. [2021]"},{"why":"Provides the Fourier feature network baseline, a Type-II neural SDF fitter.","marker":"Tancik et al. [2020]"},{"why":"Provides the Dictionary Fields baseline, a Type-III grid-vector-MLP competitor in the Chamfer comparison.","marker":"Chen et al. [2023]"},{"why":"M-SDF shows a recent function-interpolation formulation (trilinear grid as Value) that the paper interprets within its framework.","marker":"Yariv et al. [2024]"}],"fun_headline_variants":["0.43M floats match or beat million-parameter SDF nets","No neural nets: 0.43M floats outperform SDF baselines","RBF polynomial grids: SDF fitting with a fraction of parameters","Parameter-light RBF polynomial grids beat neural SDFs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that ground-truth surface points are available for the mean-shift initialization of the offset keys; when that initialization is removed, the paper's own ablation (Config-Full-3 versus Config-Full-4) shows near-surface absolute error roughly doubling, so noisy or missing surface samples would directly degrade accuracy.","fun_headline_variants_meta":{"raw":{"variants":["0.43M floats match or beat million-parameter SDF nets","No neural nets: 0.43M floats outperform SDF baselines","RBF polynomial grids: SDF fitting with a fraction of parameters","Parameter-light RBF polynomial grids beat neural SDFs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001922,"raw_usage":{"total_tokens":7498,"prompt_tokens":888,"completion_tokens":6610,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":6534}},"tokens_in":504,"tokens_out":6610,"duration_ms":49007,"temperature":1.0,"reasoning_tokens":6534,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:30:25.335118+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the same nine test meshes with the $32^3\\times 13$ representation with the mean-shift initialization disabled or fed corrupted surface samples, then measure near-surface Chamfer distance and Near-AE; a sharp degradation would confirm that the accuracy claim depends on external surface knowledge rather than on the RBF-polynomial form alone. Alternatively, compare against a hash-grid baseline under an equal parameter budget on a class of shapes whose ground-truth surfaces are not directly sampled, such as latent generative priors, where the offsets could not be initialized.","supporting_citations":[],"review_version":1}