{"id":"81619bc8-9504-4cb1-98ca-59eb4c363a3e","arxiv_id":"2608.12825","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Augmenting query-based feed-forward 3D Gaussian splatting with explicit 3D anchor centers and support radii improves rendering quality and spatially organizes the predicted Gaussians.","lead":"LocusGS adds a learnable 3D location and support radius to each Gaussian query in a feed-forward 3D scene reconstruction network, so each query gathers image evidence from and generates Gaussians in a coherent local region. It reports higher rendering quality than the TokenGS baseline at the same Gaussian budget, plus more compact token-level Gaussian groups.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Attribution confound: LocusGS adds 19.5M parameters and multi-layer anchor supervision vs TokenGS, so §4.2 gains do not isolate spatial grounding; no parameter/supervision-matched TokenGS control is tested.","rationale":"The reader's weakest assumption is exactly the attribution problem, and I agree it is the most load-bearing issue. The method itself is coherent: the anchor state, refinement equations, and losses are well-specified, and the internal ablations support the value of each component. However, the headline comparison to TokenGS is confounded by a 19.5M parameter increase and by added intermediate supervision plus anchor visibility regularization. These are not controlled in any experiment, and Table 3's spatial-compactness metric is structurally tied to Eq. 9's anchor-centered decoding. A matched control is therefore necessary before the central causal claim can be accepted. This is an addressable experimental gap rather than a demonstrated contradiction, so the reader's CONDITIONAL verdict is appropriate; I would not move to REJECT or ACCEPT without the control.","tokens_in":14562,"tokens_out":5381,"duration_ms":60080,"concrete_test":"Train a matched non-anchor control: start from TokenGS, add the same number of parameters used by LocusGS's anchor MLPs and prediction heads (e.g., per-query MLPs whose outputs are added as constant biases rather than used as anchors), and apply the identical multi-layer rendering supervision and visibility regularization schedule from Eq. 10–11 (with visibility applied to predicted Gaussian centers instead of anchors). Keep token/GS budgets fixed and use the same optimizer, schedule, and epochs on RE10K and DL3DV. If this control matches or closes the PSNR/SSIM gap in Tables 1–2, the spatial-grounding attribution fails; if the gap persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central causal claim—that explicit 3D anchor states, rather than added capacity or supervision, produce the reported gains—rests on the comparison to TokenGS in §4.2. That comparison holds token count and Gaussian budget fixed (Tables 1–2) but not model size or training objective. Table 8 shows LocusGS has 241.5M parameters versus TokenGS's 222.0M, and §3.2.7/§A.6 add two things TokenGS lacks: intermediate-layer rendering supervision (Eq. 10–11) and an extra visibility term on anchor centers. Either could explain part or all of the PSNR/SSIM gap. The component ablations in Table 4 show each anchor mechanism helps inside LocusGS, but they never augment a non-anchor TokenGS-style baseline with the same extra parameters and supervision. Table 3's dispersion metric is also partly mechanical: Eq. 9 writes Gaussian centers as anchor + radius × offset, so within-token dispersion is scaled by the learnable radius and cannot be read as independent evidence of spatial structure. The weakest point is therefore attribution: without a matched-control baseline, the abstract's claim that explicit anchor states improve rendering quality is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LocusGS, a feed-forward 3D Gaussian Splatting method that augments each learnable query token with an explicit 3D anchor state consisting of a center and a support radius. The anchor is progressively refined across decoder layers and is used in self-attention, cross-attention through an anchor-to-ray geometric bias, and Gaussian center decoding via radius-scaled offsets. Experiments on RE10K and DL3DV compare LocusGS against TokenGS under matched token and Gaussian budgets and report consistent improvements in PSNR, SSIM, and LPIPS, together with ablations and qualitative analysis of anchor layouts, cross-attention patterns, and token-level Gaussian compactness.","tokens_in":14827,"tokens_out":6591,"duration_ms":69147,"significance":"If the central claim holds, this is a useful and timely contribution: it identifies a concrete limitation of query-based feed-forward 3DGS, proposes a simple and plausible remedy, and evaluates it on large-scale benchmarks with fixed token and Gaussian budgets. Strengths include the use of external novel-view-synthesis benchmarks, component-level ablations that are mostly consistent, and a clear comparison against a well-matched query-based baseline in terms of token and Gaussian counts. The significance is moderated by an attribution confound: LocusGS differs from TokenGS not only by the anchor states but also by a non-negligible parameter increase and an additional multi-layer supervision schedule, so the reported gains are not yet cleanly attributable to spatial grounding. The token-level dispersion metric is also partly enforced by the proposed center parameterization, which weakens the claim of independent evidence for spatial organization. With matched controls, the contribution could be solid.","major_comments":[{"comment":"The paper claims in Section 4.1 that comparing under matched token and Gaussian budgets 'allows us to isolate the effect of explicit spatial grounding,' but the comparison against TokenGS does not control for model capacity or training objective. Table 8 reports 241.5M parameters for LocusGS versus 222.0M for TokenGS, and §3.2.7/§A.6 add two training signals that TokenGS does not have: intermediate-layer rendering supervision (Eq. 10–11) and an additional visibility-regularization term on anchor centers. Table 5 shows that, within LocusGS, switching from final-layer-only supervision to {6,12} supervision changes PSNR by about 0.735 dB (23.549 to 24.284), which is a substantial fraction of the 1.36 dB gap over TokenGS in the 4-view DL3DV row of Table 1. A matched control that augments TokenGS with the same extra parameters and the same supervision schedule, but without anchor states, is needed before the abstract's claim that explicit anchor states improve rendering quality can be accepted. This is load-bearing because the central contribution is causal attribution of the gains to spatial grounding.","section":"§4.1–§4.2, Tables 1–2, Table 8, Eq. (10)–(11)"},{"comment":"The token-level dispersion metric is partly mechanical. Eq. 9 sets the Gaussian center as µG_i,k = µL_i + rL_i δ_i,k, so the within-token spread of centers is explicitly scaled by the learned radius rL_i. A lower Ccentroid value for LocusGS is therefore expected from the decoding parameterization alone and is not independent evidence of a more structured learned representation. The comparison against TokenGS, which uses unconstrained centers, conflates the effect of the center parameterization with the effect of spatial grounding through attention and refinement. To support the claim, dispersion should be compared against a TokenGS variant that uses the same anchor-centered decoding but without anchor-aware attention or anchor refinement, or the dispersion should be normalized by the anchor radius. As written, Table 3's large reductions (e.g., 5.1164 to 0.1978 on RE10K) are consistent with a mechanical explanation.","section":"§4.3.3, Eq. (9), Eq. (12), Table 3"},{"comment":"The main quantitative claims are made without any measure of uncertainty. Tables 1 and 2 report point estimates only, with no standard deviations, number of seeds, or statistical tests. Given that some of the reported differences are modest in absolute terms, such as 28.41 vs. 28.89 PSNR on RE10K with 4096 tokens, the reader cannot assess whether these differences are robust. I request means and standard deviations over at least three training runs, or a paired evaluation across the same held-out scenes, for the primary TokenGS comparisons.","section":"§4.2, Tables 1–2"}],"minor_comments":[{"comment":"The phrase 'identical token and Gaussian budgets' should be qualified by noting that the models differ in parameter count and supervision schedule; otherwise the wording overstates the degree of control in the comparison.","section":"§4.1"},{"comment":"The qualitative claims about anchors concentrating around informative regions and radii adapting to local detail are supported only by selected visualizations. Reporting an aggregate statistic, such as the correlation between anchor radius and local image gradient or depth discontinuity, would strengthen this analysis.","section":"§4.3.2 and Figure 5"},{"comment":"The point-to-ray distance formula in Eq. (24) is written for a normalized direction, but the main text introduces D(µ_i^l, ℓ_j) in Eq. (3) without stating that normalization is applied first. Please add a forward reference to §A.5 at the first occurrence.","section":"§A.5, Eq. (24)"},{"comment":"The forward-time comparison shows LocusGS is about 19% slower than TokenGS (407.3 ms vs. 341.0 ms per sample), but the main paper does not discuss this compute overhead. Since the paper emphasizes a controlled comparison, the latency and memory cost should be mentioned in the main text along with the rendering-quality gains.","section":"§B.3, Table 8"},{"comment":"Reference [8] cites a CVPR 2026 paper for TokenGS; please verify the publication venue and page numbers at the time of final submission, especially given the arXiv version date.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the experiments are broadly consistent with the authors' narrative, but the central attribution claim is not yet established because the comparison to TokenGS is confounded by parameter count and supervision schedule, and the dispersion metric is partly built into the decoding parameterization. I would ask the authors to add a matched-control experiment and to report error bars before accepting. The revisions are substantial in their evidentiary requirements but well within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Name],\n\nWhat you should know: this is a genuine, well-executed advance for query-based feed-forward 3DGS. The authors add a learnable 3D anchor state—center plus support radius—to each token, refine it across decoder layers, use it to bias cross-attention, and decode Gaussians as radius-scaled local offsets. That configuration is not in TokenGS, DAB-DETR, or Scaffold-GS; it's the first time center-plus-radius anchors are used this way in a feed-forward 3DGS token framework. The improvements over TokenGS are consistent on RE10K and DL3DV across token budgets and view counts, with token count and Gaussian budget fixed. The ablations are thoughtful and show each component earns its keep; radius-scaled Gaussian decoding is clearly the biggest lever.\n\nNow the soft spots, in proportion. The central claim—that spatial grounding, not extra capacity or supervision, produces the gains—is weakened by the missing matched control. LocusGS has 241.5M parameters to TokenGS's 222.0M and adds multi-layer rendering supervision plus an extra visibility regularizer on anchor centers. The paper never runs a non-anchor TokenGS with the same added parameters and supervision. That's a real confound. The component ablations don't repair it, because they only turn pieces of the anchor mechanism on and off within LocusGS. Also, the token-level dispersion metric in Table 3 is partly mechanical: with centers parameterized as anchor plus radius times offset, the model can shrink the radius to make within-token dispersion tiny. So low dispersion is not independent evidence of coherent spatial organization. Minor: no error bars anywhere, and code is not released. These are all fixable.\n\nThe paper is honest about its scope, notes the isotropic radius as a limitation, and the related work is appropriate. The writing is clear, and the visualizations are informative.\n\nMy take: this deserves a serious referee. The core idea is new in this subfield, the experiments are broad, and the confound can be addressed with one additional baseline. If that matched-control experiment confirms the anchor effect, the paper becomes a strong acceptance. If not, the engineering value remains, but the abstract's causal claim would need to be scaled back.\n\nWho benefits: people working on feed-forward 3DGS, token-based scene decoding, or anchor mechanisms in transformers. I'd bring it to our reading group, mostly to discuss the attribution problem.\n\nRecommendation: send to peer review, request the parameter/supervision-matched ablation, error bars, and ideally release code.","headline":"Genuine anchor-token advance for feed-forward 3DGS with consistent gains, but the spatial-grounding claim is not isolated from added capacity and supervision.","tokens_in":15383,"tokens_out":3374,"would_cite":true,"duration_ms":32357,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LocusGS claims that equipping each query token with a learnable 3D anchor state—a center and a support radius—improves rendering quality and spatial coherence in feed-forward 3D Gaussian splatting under matched budgets.","keywords":["3D Gaussian Splatting","feed-forward reconstruction","query tokens","spatial anchors","novel view synthesis","cross-attention geometric bias","anchor-centered decoding"],"falsifier":"Train the baseline with the same total parameters (for example, enlarge its prediction heads to reach the same 241.5M count), the same multi-layer rendering supervision, and the same visibility losses on its predicted centers, while keeping its latent queries and free Gaussian decoding; if its PSNR and dispersion metrics match LocusGS, the anchor state is not the active ingredient.","tokens_in":14347,"feed_emoji":"📍","tokens_out":5794,"duration_ms":54149,"temperature":0.7,"pith_summary":"Feed-forward 3D Gaussian splatting systems that use learnable queries to decode groups of Gaussians suffer from a spatial decoupling: the Gaussians produced by one query scatter across the scene instead of covering one local region. LocusGS claims this happens because queries are purely latent, and that giving each query an explicit 3D anchor—a center plus a support radius—fixes it. The anchor is refined layer by layer, steers which image features the query attends to, and localizes the decoded Gaussians around the anchor. The paper reports that this single change improves novel-view rendering quality over the query-based baseline under identical token and Gaussian budgets, and that it sharply reduces token-level Gaussian dispersion. A sympathetic reader would take the paper as establishing that spatial grounding, not just representation capacity, is what organizes query-based scene tokens.","feed_headline":"Grounding each query in 3D sharpens novel views","feed_subtitle":"A center-plus-radius anchor cuts token Gaussian scatter and beats the query baseline at equal budgets.","key_machinery":"The load-bearing object is the anchor state $a_i^l=(\\mu_i^l,r_i^l)$ attached to each Gaussian token: a 3D center and a scalar support radius, both learned and progressively refined across decoder layers by residual updates. It is used in three places: anchor-aware self-attention injects a positional embedding from the center; anchor-to-ray cross-attention adds a geometric bias $b_{ij}^l=-\\tfrac12\\left(D(\\mu_i^l,\\ell_j)/(\\sigma_0 r_i^l)\\right)^2$ to the content logits, where $D$ is the point-to-ray distance, so queries prefer image tokens whose rays pass near the anchor; and anchor-centered decoding predicts Gaussian centers as $\\mu_{i,k}^G=\\mu_i^L+r_i^L\\delta_{i,k}$, local offsets scaled by the radius. Together these convert a latent query into a spatial hypothesis that selects evidence and organizes primitives.","core_discovery":"The central claim is that an explicit, learnable 3D anchor state—rather than any new loss or larger model—is what makes query-based feed-forward 3D Gaussian prediction spatially coherent. On RealEstate10K and DL3DV, the method outperforms the strongest query-based baseline with the same number of tokens and the same total Gaussian budget, for example 28.89 versus 28.41 PSNR on RealEstate10K at 4096 tokens and 24.80 versus 23.44 at four views on DL3DV. The learned anchors form a coarse spatial scaffold over the scene, with smaller radii near detailed regions and larger radii in weakly constrained areas. The paper also reports that anchor-aware cross-attention produces localized, view-consistent evidence aggregation, and that the decoded Gaussians from one token are far more compact, with the mean token-level dispersion score dropping from about 5.1 to 0.20 on RealEstate10K.","pith_inferences":["The anchor-to-ray bias should tolerate pose noise reasonably well: replacing exact rays with a distribution over rays would turn the geometric bias into an expected-distance term, a natural extension the paper does not test.","The center-plus-scalar-radius state is isotropic; an anisotropic or covariance-based support would likely capture elongated structures such as wires and poles better than the current scalar radius.","If spatial grounding is the active mechanism, the same anchor state could transfer to other token-based 3D representations, such as tokens decoding neural radiance fields or signed distance functions, not only Gaussian splats.","A parameter-matched baseline with the same auxiliary losses but no anchor-conditioned attention would settle whether the improvement comes from the anchor or from added capacity; the paper does not run this control."],"forward_implications":["At equal token and Gaussian budgets, rendering quality improves on both benchmarks and across 2-, 4-, and 6-view test settings, so the spatial-grounding design transfers to unseen context lengths.","Token-level Gaussian dispersion falls by roughly an order of magnitude, meaning each query's Gaussians describe one local surface patch rather than many floating pieces.","Cross-attention becomes localized around the anchor, so the query gathers evidence from geometrically consistent image regions, which helps in texture-rich and cluttered scenes.","Middle-layer rendering supervision is helpful and dense supervision is not; anchoring plus supervision at the middle and final layers suffices."],"supporting_citations":[{"why":"defines the 3D Gaussian splatting primitive and differentiable rasterizer that the method predicts and renders.","marker":"[4]"},{"why":"supplies the query-based feed-forward Gaussian-token formulation whose token and Gaussian budgets LocusGS matches as its primary baseline.","marker":"[8]"},{"why":"provides the dynamic-anchor-as-query idea that LocusGS generalizes from 2D object localization to 3D scene reconstruction.","marker":"[20]"},{"why":"shows Gaussians organized around sparse 3D anchors, the optimization-based counterpart to the feed-forward anchor state.","marker":"[21]"},{"why":"is the RealEstate10K benchmark and its two-view reconstruction protocol used for the main evaluation.","marker":"[22]"},{"why":"is the DL3DV benchmark used for multi-view novel view synthesis and cross-view generalization tests.","marker":"[23]"}],"fun_headline_variants":["3D anchor per query grounds Gaussians, sharpens novel views","Explicit 3D anchors stop token scatter in feed-forward splatting","LocusGS: Anchor-aware queries give spatially coherent 3D","Learnable 3D anchor per token yields more stable 3D scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the extra anchor machinery is the cause of the gains, since the baseline has fewer parameters and lacks the intermediate rendering and anchor losses; if a matched-capacity baseline with the same supervision closes the gap, the spatial-grounding claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["3D anchor per query grounds Gaussians, sharpens novel views","Explicit 3D anchors stop token scatter in feed-forward splatting","LocusGS: Anchor-aware queries give spatially coherent 3D","Learnable 3D anchor per token yields more stable 3D scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1574,"prompt_tokens":972,"completion_tokens":602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":523}},"tokens_in":588,"tokens_out":602,"duration_ms":6137,"temperature":1.0,"reasoning_tokens":523,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:32:05.756514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the baseline with the same total parameters (for example, enlarge its prediction heads to reach the same 241.5M count), the same multi-layer rendering supervision, and the same visibility losses on its predicted centers, while keeping its latent queries and free Gaussian decoding; if its PSNR and dispersion metrics match LocusGS, the anchor state is not the active ingredient.","supporting_citations":[{"cited_title":"Tokengs: Decoupling 3d gaussian prediction from pixels with learnable tokens.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026","cited_arxiv_id":null,"evidence_quote":"supplies the query-based feed-forward Gaussian-token formulation whose token and Gaussian budgets LocusGS matches as its primary baseline."},{"cited_title":"DAB-DETR: Dynamic anchor boxes are better queries for DETR","cited_arxiv_id":null,"evidence_quote":"provides the dynamic-anchor-as-query idea that LocusGS generalizes from 2D object localization to 3D scene reconstruction."}],"review_version":1}