{"id":"64ec0fb4-65a9-4415-9ed9-c5b6a5580876","arxiv_id":"2507.13707","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A spatial-tokenization transformer predicts deformable-body interactions accurately and scales to meshes beyond 100,000 nodes where graph baselines run out of memory.","lead":"A neural-network simulator groups mesh points into spatial cells, compresses them into a small set of attention tokens, and predicts how deformable bodies interact. It reports accurate predictions on meshes with over 100,000 nodes, where standard graph-based simulators run out of memory, and introduces a new large-scale simulation dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline fairness is the load-bearing issue: ABCD baselines are handicapped by an arbitrary 1000-world-edge cap and ABCD-XL has no baseline run, so 'significantly outperforms' and 'effective at 100k nodes' are not yet established.","rationale":"The reader's weakest_assumption focuses on whether cell averaging and token compression retain thin-contact/stress information. That is a real risk, but it is not the most load-bearing issue because the paper's own small-scale experiments (SPHERESIMPLE, DEFORMINGPLATE) already exercise thin and contact-rich geometries, and the reported accuracy there suggests the bottleneck is not catastrophic. The more immediate threat to the central claim is the validity of the baseline comparisons. The ABCD world-edge cap is an arbitrary, one-sided modification: it changes the baselines' capacity to represent exactly the inter-object contact that the paper claims to improve. The absence of any baseline on ABCD-XL means the scalability claim is only a demonstration that AST runs, not that it outperforms anything at that scale. If a controlled rerun shows that a baseline matches AST with an adequate edge budget, the abstract's 'significantly outperforms state-of-the-art approaches' is unsupported. This is a testable, concrete concern, unlike the more speculative information-loss worry. The verdict remains CONDITIONAL: the paper's method is plausible and the small-scale results are mostly positive, but the artifact release and a controlled baseline comparison are necessary before the headline claim can be accepted.","tokens_in":13605,"tokens_out":11341,"duration_ms":130942,"concrete_test":"On ABCD, rerun MGN, BSMS, and HCMT with the original radius-based world-edge construction, sweeping the world-edge radius so the number of edges per frame is comparable to the number of mesh edges (e.g., 5k, 20k, 50k) and also with the current 1000-edge cap for reference. If any baseline reaches or beats AST's reported RMSE of 0.505e-3, the significant-improvement claim on deformable-deformable contact fails. On ABCD-XL, run a memory-reduced MGN (gradient checkpointing, reduced radius, or subgraph training) to obtain a comparable accuracy number; if it fits and yields RMSE at or below 0.480e-3, the 'effective at 100k nodes' claim is not unique to AST.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section A.1 caps the number of world edges used by MGN, BSMS, and HCMT on the ABCD dataset to the 1000 smallest pairwise distances per frame. This cap applies only to the graph baselines: AST does not construct world edges, so it is not subject to the same interaction budget. For a 4k-node tetrahedral mesh, the contact interface between two deformable bodies can generate thousands of close node pairs; retaining only 1000 edges will truncate the contact graph and degrade the baselines precisely on the task the paper emphasizes (deformable-deformable contact). No sensitivity sweep or justification is given for the number 1000, and the authors instead claim 'significant improvement' over these baselines. On ABCD-XL, all baselines OOM and no memory-reduced or approximate baseline is run; the large-scale result is therefore an absolute accuracy number with no comparator. The paper's Figure 10/11 only show runtime scaling up to 20k elements, not accuracy at 100k. Thus the central quantitative claims rest on an uncontrolled comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive Spatial Tokenization (AST), a learned simulator for deformable body interactions. AST maps an unstructured mesh onto a per-frame octree of spatial cells, aggregates mesh-node features into cell features via message passing, compresses the sparse cells into a fixed-length set of latent tokens using farthest-point sampling and cross-attention, processes the tokens with stacked self-attention layers, and decodes the result back to the mesh. The authors introduce two new datasets, ABCD and ABCD-XL, built from CAD geometries with Abaqus simulations, and compare AST against MGN, BSMS, and HCMT on DEFORMINGPLATE, SPHERESIMPLE, ABCD, and ABCD-XL. Reported rollout RMSEs favor AST on DEFORMINGPLATE and ABCD, are comparable on SPHERESIMPLE, and are unopposed on ABCD-XL because the baselines run out of memory.","tokens_in":13903,"tokens_out":5185,"duration_ms":59376,"significance":"The idea of replacing explicit dynamic world edges with adaptive spatial tokenization is a genuine and potentially impactful contribution for learned deformable-body simulation. The method is clearly described, the octree-plus-cross-attention architecture is well motivated, and the new ABCD/ABCD-XL datasets address a real gap in large-scale deformable-contact benchmarks. The DEFORMINGPLATE result (displacement RMSE 1.1 vs 2.9 for HCMT) and the reported runtime scaling are valuable. However, the paper's central claims of 'significant improvement over state-of-the-art' and effectiveness at 100,000 nodes rest on comparisons that are not fully controlled, and the new dataset is not released. The significance is therefore conditional: the architecture is promising, but the quantitative claims need to be re-established under a fair baseline protocol.","major_comments":[{"comment":"The ABCD comparison is not controlled. The world-edge construction is modified for MGN, BSMS, and HCMT by retaining only the 1000 world edges with the smallest pairwise distances per frame, while AST uses no world edges and is therefore not subject to this cap. For a 4k-node tetrahedral mesh in ABCD, the deformable-deformable contact interface can contain thousands of close node pairs, and truncating to 1000 edges can remove exactly the edges that encode contact. No sensitivity sweep over the cap is reported, and no justification is given for the value 1000. The reported ABCD gains (0.505 vs 0.541 for HCMT) therefore cannot be attributed to the method rather than to the baseline handicap; the claim of significant improvement on this dataset is not established.","section":"A.1, Table 1"},{"comment":"The large-scale claim is not supported by a controlled comparison. On ABCD-XL, all three baselines run out of memory, so Table 1 reports only AST's absolute RMSE (0.480) with no baseline accuracy. Figures 10 and 11 show runtime scaling only up to roughly 20k elements, not accuracy at 100k elements. The statement that AST 'remains effective' on meshes exceeding 100,000 nodes requires at least an accuracy comparison against a memory-reduced or approximate version of a baseline (e.g., a subsampled MGN or a baseline with restricted edges), or a clearly framed scalability study without the comparative claim.","section":"§5.2, Table 1, Figures 10-11"},{"comment":"The claim of consistent significant improvement over state-of-the-art is overstated on SPHERESIMPLE. AST reports 14.4 ± 0.8 versus BSMS 15.0 ± 0.8; the difference is within one standard error, and MGN's 19.0 ± 4.9 overlaps. Only DEFORMINGPLATE shows a clearly separated improvement (1.1 vs 2.9). The abstract and conclusion should be qualified to reflect the actual pattern of results.","section":"§5.2, Table 1"},{"comment":"The paper introduces ABCD and ABCD-XL as novel datasets and claims them as a contribution, but provides no download URL, license, or code release, and the supplementary material contains no artifact. Without the dataset and code, the empirical results cannot be reproduced or audited; at minimum the dataset construction pipeline should be released or a clear availability statement added.","section":"§1, §5.1"}],"minor_comments":[{"comment":"The units for stress are not specified; the header says ×10^-3 for displacement only, but stress values appear on very different scales (e.g., 6891 for DEFORMINGPLATE vs 2.11 for ABCD-XL).","section":"Table 1"},{"comment":"The notation is inconsistent: c_l^i is used both for the cell and for its feature, and the definition of v'c_t = [c_L^1, ..., c_L^{N_L}] introduces a symbol that is not defined elsewhere.","section":"Equation (3)"},{"comment":"The statement 'When Lcell = 7, all mesh nodes are assigned to a single cell at the initial frame' is inconsistent with the octree definition in Section 3.2, where larger L gives smaller cells; please clarify or correct this sentence.","section":"B.1"},{"comment":"The claim that HCMT 'performs well during the initial 50 frames but gradually diverges' on SPHERESIMPLE would benefit from a quantitative rollout curve, since the table only reports divergence.","section":"A.3"},{"comment":"The phrase 'significantly outperforms' is used before presenting the statistical uncertainty; please align the wording with the standard errors in Table 1, particularly for SPHERESIMPLE.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than rejection because the architecture and dataset are potentially valuable, but the baseline protocol must be fixed and artifacts released. The 1000-edge cap in Section A.1 is the main concern; the authors should run a sweep over the edge budget or use a standardized interaction budget for all methods. The lack of code and data is also a significant barrier to verification, especially given that ABCD-XL has no baseline runs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is sensible: instead of building dynamic world edges between deformable bodies, quantize the mesh into octree cells, aggregate node features into cells, compress to a fixed number of latent tokens with cross-attention, and run self-attention in token space. The integration is new for deformable-body simulation, the element-aware feature handling is a nice touch, and the new ABCD/ABCD-XL dataset fills a real gap. The method description is clear enough to reproduce, which is more than many papers in this area manage.\n\nWhat the paper does well: on DEFORMINGPLATE the improvement is large (displacement RMSE 1.1 vs 2.9 for HCMT), and the efficiency argument—avoiding O(n^2) edge construction—is sound. The ablations on cell depth and the runtime scaling plots help.\n\nThe soft spots are in the comparison, not the method. The ABCD baselines are run with world edges capped at the 1000 smallest per frame. That cap is not in the original MGN protocol, no sensitivity sweep is given, and it applies only to the graph baselines, not to AST. On a 4k-node contact problem, 1000 edges can truncate the contact interface exactly where the paper wants to show superiority. On ABCD-XL, all baselines OOM, so the headline 'effective at 100k nodes' is an absolute number with no comparator; the runtime plots stop at 20k elements. The reported gains on SPHERESIMPLE (14.4 vs 15.0) and ABCD (0.505 vs 0.541) are modest. So the 'significantly outperforms' claim currently rests on one dataset plus an unfair baseline setting.\n\nThere are also the usual missing artifacts: no code or data release, and the error bars are not tied to a number of seeds. I'd want those addressed, but none of this is fatal to the core idea.\n\nWho it's for: people working on learned simulation for contact-rich deformables will want to know this tokenization approach. It deserves a serious referee, but the review should focus on forcing a fair baseline comparison and releasing artifacts.\n\nMy recommendation: accept for peer review, conditional on the authors rerunning ABCD baselines without the arbitrary edge cap (or justifying it with a sweep), adding any feasible baseline on ABCD-XL, and releasing code/data.","headline":"A plausible tokenization method for deformable-body simulation, but the empirical claims overreach because the baselines are handicapped and the large-scale result has no comparator.","tokens_in":14376,"tokens_out":2587,"would_cite":false,"duration_ms":27718,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that Adaptive Spatial Tokenization (AST), which maps deformable-body meshes into octree cells and compresses them into fixed-length tokens via cross-attention, predicts deformation dynamics more accurately than…","keywords":["Adaptive Spatial Tokenization","deformable body simulation","mesh-based physics learning","graph neural network scalability","octree spatial quantization","cross-attention tokenization","large-scale finite element simulation","contact dynamics"],"falsifier":"Take a dense mesh with a thin feature or a narrow contact interface whose thickness is smaller than the octree cell size, run AST with a cell depth coarse enough that the feature falls inside one cell, and compare rollout error against a finer cell depth. If the error does not grow sharply, the averaging assumption survives; if it does, cell quantization is erasing decisive local information.","tokens_in":13425,"feed_emoji":"🧩","tokens_out":6036,"duration_ms":67167,"temperature":0.7,"pith_summary":"The paper proposes Adaptive Spatial Tokenization (AST), a way to encode the state of deformable-body simulations as a compact set of latent tokens instead of a graph with dynamically computed pairwise edges. The idea is to bin mesh nodes into octree cells each frame, average node features per cell, and use cross-attention to compress the cells into a fixed-length token set that a transformer updates to predict the next state. If the reported results hold, AST removes the O($n^{2}$) world-edge bottleneck that makes graph-based simulators impractical on dense meshes, and extends learned simulation to meshes above 100,000 nodes. The paper also introduces ABCD and ABCD-XL datasets for benchmarking large-scale deformable contact.","feed_headline":"Spatial tokenization simulates 100K-node meshes where graph methods fail","feed_subtitle":"New method bins mesh nodes into octree cells and attention tokens, beating baselines on contact-heavy benchmarks.","key_machinery":"The load-bearing object is the adaptive spatial token: a per-frame octree cell that groups nearby mesh nodes, whose averaged features are compressed through cross-attention into a fixed number of latent tokens. The octree gives O(n) spatial grouping and supports sparse convolution; farthest-point sampling selects a spatially distributed subset of cells as queries; cross-attention compresses those cells into a fixed-length token set; self-attention propagates global information across tokens; and the decoder mirrors the process back to the original mesh. This machinery replaces explicitly constructed pairwise world edges with cell-based interaction grouping.","core_discovery":"AST's central claim is that spatial quantization plus attention-based tokenization is a better representation for contact-rich deformable simulation than node-level graphs. At each frame, mesh node positions define an octree; non-empty cells aggregate the embedded node features by averaging, and a cross-attention layer with farthest-point-sampled queries condenses the cells into 256 or 512 fixed tokens. Self-attention layers then predict the next state in latent space, and the decoder reverses the process back to mesh and element nodes. In the paper's experiments, AST reports the lowest rollout RMSE on DEFORMINGPLATE, SPHERESIMPLE, and ABCD, and is the only method that runs on ABCD-XL, whose meshes exceed 100,000 nodes, where MGN, BSMS, and HCMT run out of memory.","pith_inferences":["One implication the paper leaves implicit is that the fixed token budget decouples memory from mesh size, so the same architecture is a candidate for even larger industrial meshes once the cell count saturates.","Because the octree is rebuilt every frame, the cell depth Lcell is a continuous accuracy-versus-speed dial; a coarser grid trades contact fidelity for cost, which could be exploited for adaptive simulation.","The averaging step is the likeliest failure point; a testable extension would replace mean pooling with attention pooling or per-cell moment features and check whether thin-feature contacts improve.","Treating the state as vector fields over cells suggests the tokenization could transfer to other grid-based physical fields such as fluids or thermal fields, but the paper only demonstrates solid mechanics."],"forward_implications":["Dynamic pairwise world edges are no longer needed, so the O(n^2) memory bottleneck becomes O(n) cell aggregation, which is what allows 100K-node runs.","On DEFORMINGPLATE, AST reports displacement RMSE of 1.1e-3 versus 2.9e-3 for HCMT and stress RMSE of 842 versus 7272, showing the tokenized representation does not sacrifice contact accuracy.","On ABCD-XL, AST is the only method that avoids out-of-memory and reports displacement RMSE 0.480e-3, demonstrating stable long rollouts at large scale.","The same tokenizer, processor, and decoder transfer across quasi-static compression, cloth-sphere interaction, and CAD part squishing, suggesting the representation generalizes across material and contact regimes.","The new ABCD and ABCD-XL datasets give future methods a common benchmark for large-scale deformable interactions."],"supporting_citations":[{"why":"Supplies the MeshGraphNets baseline and the message-passing formulation AST builds on for graph encoding and world-edge construction.","marker":"[20]"},{"why":"Hierarchical Contact Mesh Transformer baseline AST compares against and whose node-level attention motivates token-based contact modeling.","marker":"[32]"},{"why":"Bi-stride multi-scale GNN baseline that AST claims to outperform on volume meshes.","marker":"[3]"},{"why":"Supplies the octree and sparse convolution machinery used to quantize mesh nodes into spatial cells.","marker":"[29]"},{"why":"Shape2VecSet cross-attention to fixed-length latent sets inspires AST's cell tokenization design.","marker":"[33]"},{"why":"Shows cross-attention mapping 3D shapes to fixed tokens, which AST adapts to simulated vector fields.","marker":"[16]"},{"why":"Farthest Point Sampling selects the spatially distributed cell subset used as queries in the tokenizer.","marker":"[21]"},{"why":"ABC CAD dataset is the geometry source for the new ABCD and ABCD-XL benchmarks.","marker":"[13]"}],"fun_headline_variants":["AST: attention tokens beat graphs on 100K-node deformable meshes","Adaptive spatial tokens power 100K-node deformable body sims","Spatial tokens, not edges: deformable sims at 100K nodes","Take that, graphs: AST simulates 100K-node deformable meshes","Beyond graphs: adaptive spatial tokenization for large-scale deformable bodies"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that averaging the mesh nodes inside each octree cell, and then compressing the cells into a fixed number of tokens, keeps enough local strain, stress, and contact information to make accurate long rollout predictions.","fun_headline_variants_meta":{"raw":{"variants":["AST: attention tokens beat graphs on 100K-node deformable meshes","Adaptive spatial tokens power 100K-node deformable body sims","Spatial tokens, not edges: deformable sims at 100K nodes","Take that, graphs: AST simulates 100K-node deformable meshes","Beyond graphs: adaptive spatial tokenization for large-scale deformable bodies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001234,"raw_usage":{"total_tokens":5078,"prompt_tokens":967,"completion_tokens":4111,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":4010}},"tokens_in":583,"tokens_out":4111,"duration_ms":31568,"temperature":1.0,"reasoning_tokens":4010,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:18:08.010961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dense mesh with a thin feature or a narrow contact interface whose thickness is smaller than the octree cell size, run AST with a cell depth coarse enough that the feature falls inside one cell, and compare rollout error against a finer cell depth. If the error does not grow sharply, the averaging assumption survives; if it does, cell quantization is erasing decisive local information.","supporting_citations":[{"cited_title":"Learning Flexible Body Collision Dynamics with Hierarchical Contact Mesh Transformer","cited_arxiv_id":null,"evidence_quote":"Hierarchical Contact Mesh Transformer baseline AST compares against and whose node-level attention motivates token-based contact modeling."},{"cited_title":"O-cnn: Octree-based convolutional neural networks for 3d shape analysis.ACM Transactions On Graphics (TOG), 36(4):1–11, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the octree and sparse convolution machinery used to quantize mesh nodes into spatial cells."},{"cited_title":"3dshape2vecset: A 3d Shape Representation for Neural Fields and Generative Diffusion Models.ACM Transactions on Graphics (TOG), 42(4):92:1–92:16, 2023","cited_arxiv_id":null,"evidence_quote":"Shape2VecSet cross-attention to fixed-length latent sets inspires AST's cell tokenization design."},{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017","cited_arxiv_id":null,"evidence_quote":"Farthest Point Sampling selects the spatially distributed cell subset used as queries in the tokenizer."}],"review_version":1}