{"id":"b59e53e1-f8d6-4cc6-adfc-d2a9d664a9a8","arxiv_id":"2610.01036","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Streamed-edge execution, checkpoint-specific code generation, and tiled memory use let an unchanged MACE checkpoint run 3 to 5 times faster and scale to 11.24 million atoms on one A100 GPU.","lead":"An inference engine, Symmetrix-XL, runs pretrained MACE machine-learning potentials up to 5 times faster and raises the largest system one GPU can hold from about 24,000 to over 11 million atoms, without changing the trained weights. It streams edge computations, generates checkpoint-specific code, and tiles memory, so large molecular-dynamics runs need fewer GPUs and cheaper hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The blanket fidelity claim is anchored on MPA-0 float64 plus one SrTiO3 OMAT-0 float32 config, while the flagship 11.24M-atom claim extrapolates the spline/preservation guarantee to unmeasured checkpoint-precision-scale combinations.","rationale":"The reader identified fidelity generalization from the finite test battery to untested downstream observables as the weakest assumption. My stress-test sharpens that to a more concrete and checkable gap: the fidelity battery is essentially MPA-0 float64, while the headline throughput, capacity, and scaling claims are OMAT-0 medium float32; only one SrTiO3 config in float32 is checked end-to-end, and the flagship 11.24M-atom scale and the case-study chemistries have no reference comparison. This is load-bearing for the abstract's blanket fidelity sentence specifically, and it is testable with a bounded computational sweep. I nevertheless recommend UNCHANGED because (i) the §3 streamed reassociation is algebraically exact and the spline is a per-checkpoint, gated approximation with documented error bounds; (ii) the one float32 OMAT-0 check shows errors ~1e-5–1e-4 in physical units, orders below physically meaningful thresholds; (iii) the 6.0 vs 6.5 A graph asymmetry disclosed in Fig. 2 penalizes, not favors, Symmetrix-XL, so the speedup claim is conservative; and (iv) the paper is unusually explicit about capacity-probe, single-trajectory, internal-validation, and platform-dependence limits (b79, b214, b391, b410). No internal inconsistency or algebraic error was found. The suggested sweep is a verification step, not evidence of a defect.","tokens_in":31412,"tokens_out":6884,"duration_ms":54026,"concrete_test":"Run a fidelity sweep for MACE-OMAT-0 medium in float32: compare Symmetrix-XL against mace-torch on ~200 randomly sampled MP/WBM structures spanning >25 elements and varied coordination (including stretched and near-transition-state geometries), reporting max energy-per-atom, max force-component, and max stress-component differences. Separately re-run the Appendix B spline gate for this checkpoint and precision at 256 nodes and report the worst internal derivative error in force units via an end-to-end atomic-force comparison. Pass criterion: max |ΔF| consistent with the 6.53e-5 eV/A SrTiO3 result (within ~1e-3 eV/A) and spline gate acceptance. If deviations degrade on any chemistry, the abstract's blanket fidelity claim needs a precision/checkpoint qualifier.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The architecture-preserving claim has two parts: an algebraic reassociation (streamed edges, §3) that is exact, and a radial-spline projection (Appendix B) that is a genuine approximation whose gate tolerances are checkpoint- and precision-specific. The measured worst derivative error of 7.23e-3 (Appendix B) is in internal learned units, so its physical force impact must be verified pointwise, not assumed. The fidelity battery in §7.5 and Appendix A is nearly all MACE-MPA-0 medium in float64: WBM relaxation/discovery, thermal conductivity, 87 diatomic pairs, DynaMat, all MPA-0. For the checkpoint and precision actually used in the headline throughput and capacity numbers — MACE-OMAT-0 medium, float32 — the end-to-end evidence is one SrTiO3 5,000-atom task (energy 1.200e-5 eV/atom, max |F| 6.53e-5 eV/A, §7.5) plus an NVE drift comparison on the same structure. The abstract's claim that 'Energy, force, stress ... reproduce reference MACE behavior within measured tolerances' is therefore asserted over a coverage region (OMAT-0 float32 at 1.25–11.24 M atoms, varied chemistries such as the Li–Al, SEI, and Au–Sn case studies) where no reference comparison exists. This is not an internal inconsistency: the spline gate is per-checkpoint and the SrTiO3 spot check shows errors orders below physically meaningful thresholds. But the preservation guarantee at the advertised scale rests on extrapolation from that thin base, and the paper's own case studies concede internal-only validation (b391). The recent optimized-deployment literature (refs 33–34) also means the magnitude of the 3.1–5.0x speedup is baseline-dependent; that is disclosed (§7, Fig. 2 caption) and conservative in direction, so it does not threaten the central thesis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper presents Symmetrix-XL, an inference engine for pretrained MACE checkpoints that combines streamed-edge execution (avoiding graph-wide materialization of expanded edge intermediates), model-specialized code generation (compiling checkpoint-specific operators), and tiled execution over a bounded workspace. The central claims are: complete LAMMPS-step inference is 3.1–5.0× faster than ML-IAP + cuEquivariance; single-A100 capacity for MACE-OMAT-0 medium rises from 24,565 to 11.24 million atoms; the backend weak-scales to 703 million atoms on 64 A800 GPUs at 93.8% efficiency; and energy, force, stress, MD stability, and Matbench Discovery results reproduce reference MACE behavior within measured tolerances. The fidelity evidence consists of an extensive float64 MACE-MPA-0 battery (WBM relaxation/discovery, thermal conductivity, 87 diatomic pairs, DynaMat, NVE conservation) plus a single float32 OMAT-0 SrTiO3 spot check, with downstream-level agreement against published results reported in Table 2. Seven case studies (Li–Al polycrystals, a cathode/electrolyte interface, shocked β-HMX, thiirane polymerization, an SEI full cell, polycrystalline Li6PS5Cl, and Au–Sn dissolution) demonstrate realistic large-scale workflows.","tokens_in":31799,"tokens_out":5262,"duration_ms":35559,"significance":"If the results hold, the paper establishes post-training execution as a genuine third scaling axis for atomistic foundation models, complementary to architecture redesign and distillation, requiring no retraining or weight changes. The engineering claims are well evidenced: the benchmark protocol (§6) is unusually careful (five warm-up steps, 20 measured steps, three fresh-process trials, medians, zero neighbor-list-rebuild requirement, twice-reproduced capacity boundaries), the comparisons use matched checkpoints, coordinates, precision, and properties, and the distributed-scaling numbers are internally consistent with the reported atom-per-GPU counts. The fidelity checks validate against external published workflows and the reference mace-torch implementation rather than against paper-defined quantities, and the numerical agreement in Table 2 is quantitatively close (e.g., WBM discovery F1 0.835664 vs 0.835569; diatomic MAE 5.67e-6 eV/dimer). The paper is also unusually honest about its own limitations, repeatedly scoping case-study conclusions as single-trajectory or internally validated.","major_comments":[{"comment":"The sentence \"Energy, force, stress, molecular-dynamics stability, and Matbench Discovery evaluations reproduce reference MACE behavior within measured tolerances\" (b6) is not scoped to the configurations actually verified. The end-to-end fidelity battery in §7.5, Table 2, and Appendix A is almost entirely MACE-MPA-0 medium in float64; for the checkpoint and precision used in the headline throughput and capacity numbers (MACE-OMAT-0 medium, float32), the reference comparison is a single 5,000-atom SrTiO3 task (energy 1.2e-5 eV/atom, max |F| 6.53e-5 eV/A, stress 8.82e-8 eV/A^3) plus an NVE drift comparison on the same structure. Because the radial-spline gate in Appendix B is explicitly checkpoint-specific (worst internal derivative error 7.23e-3 in internal learned units, with physical fidelity assessed only by end-to-end tests), the preservation guarantee as advertised at 1.25–11.24 million atoms is an extrapolation from one chemistry and one structure. This is not an internal inconsistency, and the paper is transparent about the per-checkpoint nature of the gate, but the unqualified abstract phrasing overstates the coverage. I recommend either adding float32 OMAT-0 parity checks on one or two additional chemistries (the Li–Al, Li6PS5Cl, or OFF23 systems already used in case studies would be cheap) or explicitly restricting the fidelity sentence to the tested configurations.","section":"Abstract and §7.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The capacity-boundary timings in Table 1 each correspond to a different system size, and the paper states this clearly, but the headline comparison of tiled versus standard streaming at their respective capacity boundaries (\"increases the per-atom Loop time by approximately 4% ... while accommodating approximately nine times as many atoms\") conflates a tiling-effectiveness claim with a capacity claim. A reader could take the 4% overhead as a same-size tiling cost; it is not, since the tiled run operates at ~9× the atom count. The authors disclose this in the caption, but the §7.2 text should state equally prominently that a matched-size comparison of the two streaming modes is not reported anywhere in the paper.","section":"§7.2 / Table 1","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"Several references are listed twice under different numbers: [10] duplicates [2], [12] duplicates [39], [28] duplicates [23], and [34] duplicates [24]. These should be consolidated to avoid citation clutter.","section":"References","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The caption labels two panels as \"(e)\" (panels labelled e in the figure legend and in the sentence about DPn/DPw); one of the panel identifiers should be re-lettered for consistency with the in-text references.","section":"Figure 9 caption","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The sentence \"Fig. 6 shows the four sample cells; Fig. 6 resolves the 10M grain-boundary network at true atomic scale\" cites the same figure twice; the second reference should point to the specific panel (Fig. 6b) to avoid ambiguity.","section":"§8.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The ASE comparison uses an exact 6.0 A graph for PyTorch and a retained 6.5 A effective-cutoff graph for Symmetrix-XL. This asymmetry is disclosed in the caption and is conservative (it loads more edges onto Symmetrix-XL), but the main text should state the conservatism explicitly so readers do not treat the matched-LAMMPS claim as applying unchanged to the ASE boundary.","section":"§7.1 / Figure 2a","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Tiled streaming uses a fixed 8,192-workspace-receiver parameter, which appears to be a free design choice. A sentence reporting the sensitivity of capacity and per-atom Loop time to this parameter (or stating that results were insensitive over a tested range) would strengthen the claim that tiling is robust.","section":"§7.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Subscripts and superscripts are rendered inconsistently in running text (SrTiO3, Å3, cm2/s, eV/Å3, 10^4). This is likely a typesetting issue, but the camera-ready version should enforce uniform formatting of chemical formulae and units.","section":"Throughout","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"minor_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Worth your time. This is a well-executed systems contribution, not a modeling paper, and it does not pretend otherwise. The three mechanisms are individually grounded in cited prior work—streamed-edge execution is the Sobek formulation [49], the specialization program extends Tan/Witt [33,34], tiling is a bounded-memory extension—so the novelty is the integration on top of Symmetrix plus the measured results: 11.24M atoms on one A100 versus the 24.5k baseline, 3.1–5.0x LAMMPS speedup, 703M atoms weak-scaling at 93.8%. Those numbers are mutually consistent between abstract and body, which is more than you usually get.\n\nWhat it does well: the benchmark protocol is careful. Five warm-up, 20 measured steps, medians over fresh processes, zero neighbor-rebuild requirement enforced, capacity boundaries reproduced twice. The graph-construction asymmetry in the ASE comparison (6.0 Å exact for PyTorch, 6.5 Å retained skin for Symmetrix-XL) is disclosed, and it cuts against Symmetrix-XL by adding edges, so it is conservative. Fidelity is checked against genuinely external published results—Matbench Discovery, published MACE-MPA-0 workflows—not against quantities the paper defines, so the circularity burden is low. And the limitations section is unusually honest: capacity probes explicitly flagged as not sustained MD [b410], platform dependence, single-trajectory case studies.\n\nThe stress-test note is right about the one real soft spot: the fidelity battery is almost all MACE-MPA-0 float64, while the checkpoint and precision behind the headline capacity numbers is MACE-OMAT-0 float32. For that regime the end-to-end evidence is one 5,000-atom SrTiO3 task plus an NVE comparison. The spline gate is per-checkpoint and per-precision, and the reported worst derivative error (7.23e-3) is in internal learned units, so physical force impact is pointwise, not assumed. That means the preservation guarantee at 1.25–11.24M atoms across the varied chemistries in the case studies rests on extrapolation from a thin base. But it is not a fatal gap: the SrTiO3 spot check shows errors orders below physically meaningful thresholds, the spline qualification is per-checkpoint by design, and the case studies themselves concede internal-only validation where it applies. I would ask for one more OMAT-0 float32 fidelity point at genuinely large scale—say a million-atom energy/force comparison against the reference, or matched sustained MD—before banking on the blanket claim. I would not block the paper on it.\n\nWho it is for: anyone deploying equivariant atomistic foundation models and anyone working on the accuracy-cost frontier of MLIP inference. It deserves a serious referee; accept with the fidelity-coverage refinement in mind.","headline":"A genuinely solid systems paper: the integration and the measured results are new, honestly reported, and the one real soft spot—thin OMAT-0 float32 fidelity coverage at the scale the headline numbers actually claim—is acknowledged and does not break the central thesis.","tokens_in":32504,"tokens_out":2392,"duration_ms":18484},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"An already-trained MACE checkpoint can be executed far faster and on far more atoms per GPU by changing only the inference engine—streaming edges, compiling checkpoint-specific operators, and bounding workspace with tiles—while keeping…","keywords":["equivariant atomistic foundation models","MACE","inference engine","streamed-edge execution","model-specialized code generation","tiled execution","molecular dynamics","single-GPU capacity"],"falsifier":"Run the optimized backend on a task outside the tested battery—say a nudged-elastic-band transition-state or rare-event kinetics trajectory, or a checkpoint other than the MACE-OMAT-0-medium and MACE-MPA-0-medium ones used here—and compare forces and long-trajectory observables against mace-torch; deviations beyond the tested tolerances would weaken the unchanged-weights claim. A matched-cutoff benchmark equalizing graph construction at 6.0 Å on both sides would similarly test whether the 3.1–5.0× speedup survives removal of the disclosed skin asymmetry.","tokens_in":32825,"feed_emoji":"⚛️","tokens_out":5764,"duration_ms":39091,"temperature":0.7,"pith_summary":"This paper tries to establish that post-training execution is a third scaling axis for atomistic foundation models, alongside designing cheaper architectures and distilling to smaller models. Its claim is that a fixed, pretrained MACE checkpoint runs dramatically faster and on far more atoms per GPU when only the inference engine changes—streamed edge processing, compiled checkpoint-specific operators, and bounded tiled workspaces—with the learned weights untouched. If the claim holds, the hardware threshold for million-atom simulations drops sharply, and dedicated inference backends become a practical lever that can be combined with model redesign and compression. The novelty is not a new architecture but evidence about how much of the accuracy–cost frontier is determined by execution rather than by the model itself.","feed_headline":"Trained MACE model runs 11 million atoms on one A100","feed_subtitle":"Streamed, compiled inference cuts LAMMPS-step time 3–5× without touching learned weights, matching reference accuracy.","key_machinery":"The central object is a reassociation of the equivariant message-passing contraction. Each expanded edge weight is written as a projection of a compact radial embedding, $W_{e,\\pi a} = \\sum_q \\varphi_{eq} B_{q,\\pi a}$, and the receiver message $m_i = \\sum_{e:r(e)=i} \\sum_{\\pi,a} W_{e,\\pi a} T_{\\pi a}(h_{s(e)}, Y_e)$ is formed by generating each contribution near its point of use, accumulating it into the receiver, and discarding it. This cuts persistent graph-dependent storage from $\\mathcal{O}(E C P) + \\mathcal{O}(E W_R) + \\mathcal{O}(N C G)$ to $\\mathcal{O}(E K_{\\mathrm{geom}}) + \\mathcal{O}(N C K_{\\mathrm{node}}) + \\mathcal{O}(K_{\\mathrm{work}})$, i.e. graph-plus-node state with a bounded kernel-local workspace, with the reverse pass recomputing radial and angular factors under a chosen schedule. The same streamed schedule is combined with code generation that bakes the checkpoint's fixed sparse coupling structure into compiled CPU or GPU kernels, and with tiled execution that partitions receivers and incident edges into bounded tiles so the large temporary workspaces are tile-sized rather than graph-sized.","core_discovery":"The paper's central claim is that an already-trained equivariant atomistic foundation model can be scaled at execution time instead of at training or architecture time. Symmetrix-XL keeps the MACE checkpoint's learned weights and architectural contract unchanged and replaces how the fixed inference program executes: expanded edge contributions are formed, accumulated into their receiver, and discarded rather than materialized graph-wide; the checkpoint's fixed sparse Clebsch–Gordan paths, coefficients, and layouts are compiled into model-specific kernels; and receivers can be partitioned into bounded tiles that reuse a fixed device-memory workspace. On complete LAMMPS steps the backend is 3.1 to 5.0 times faster than the ML-IAP with cuEquivariance baseline, lifts the single-A100 capacity for MACE-OMAT-0 from 24,565 to 11,240,455 atoms, and weak-scales to 703 million atoms on 64 A800 GPUs at 93.8 percent efficiency. Energy, force, stress, NVE conservation, thermal conductivity, and Matbench Discovery results all reproduce the reference mace-torch behavior within the stated tolerances, although the implementations are not bitwise identical.","pith_inferences":["The disclosed graph-construction asymmetry—6.0 Å exact for PyTorch versus a 6.5 Å effective cutoff with a retained 0.5 Å skin for Symmetrix-XL—means part of the end-to-end speedup may come from trading a slightly larger effective cutoff against fewer neighbor-list rebuilds; a matched-cutoff benchmark would isolate the pure model-evaluation gain.","A natural extension is distributing tiled execution across devices within a node, since tiling already partitions receivers; whether tiles can map across GPUs without the graph-wide $H_1$ state would determine if tiled capacity composes with scale-out.","The spline approach that holds at 256 nodes with worst internal derivative error $7.23\\times10^{-3}$ suggests the same project-onto-spline trick could be pushed to fewer nodes for other models, trading measurable accuracy for further memory and speed."],"forward_implications":["Million-atom molecular dynamics becomes feasible on a single consumer GPU: tiled streaming reaches 4.15 million atoms on one RTX 5090 and an estimated 1.7–2.9 million on 16 and 24 GB devices, respectively, for the tested workload.","Distributed LAMMPS runs gain larger local domains per rank, since tiled streaming weak-scales to 703 million atoms on 64 A800 GPUs, lowering the minimum GPU count for a given system.","Dedicated inference backends join architecture redesign and distillation as a third lever on the accuracy–cost frontier, and the three can be combined since each leaves the deployed model untouched in different ways.","The streaming, specialization, and tiling principles are claimed to extend to other equivariant graph neural networks whose operators admit compatible schedules, subject to per-architecture implementation and numerical qualification."],"supporting_citations":[{"why":"Defines the MACE architecture, directed radius graph, and checkpoint contract that Symmetrix-XL executes unchanged.","marker":"[40]"},{"why":"Supplies the streamed equivariant tensor-product reassociation on which streamed-edge execution builds.","marker":"[49]"},{"why":"The vendor-optimized cuEquivariance operator library that, with ML-IAP, forms the baseline whose complete-step time Symmetrix-XL must beat.","marker":"[44]"},{"why":"LAMMPS is the molecular-dynamics engine inside which both the baseline and Symmetrix-XL throughput and capacity are measured.","marker":"[46]"},{"why":"The reference mace-torch PyTorch implementation against which Symmetrix-XL numerical fidelity is qualified.","marker":"[43]"},{"why":"Matbench Discovery is the downstream benchmark whose WBM relaxation and discovery metrics must reproduce reference results.","marker":"[1]"},{"why":"The prior Symmetrix inference backend that Symmetrix-XL extends with streaming, model specialization, and tiling.","marker":"[38]"}],"fun_headline_variants":["One A100 runs 11M atoms from a trained MACE model","Trained MACE scales to 11M atoms on a single GPU","Symmetrix-XL runs MACE on 11M atoms with one A100","Post-training execution lifts MACE capacity 450-fold","No retraining: MACE hits 11M atoms on one A100"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The claim that unchanged weights give reference behavior rests on the fidelity of streamed, reordered contractions and radial-spline interpolation to genuine MACE outputs, verified only on the finite battery of tested systems and checkpoints; an untested observable or checkpoint more sensitive than the aggregate metrics could deviate beyond tolerance.","fun_headline_variants_meta":{"raw":{"variants":["One A100 runs 11M atoms from a trained MACE model","Trained MACE scales to 11M atoms on a single GPU","Symmetrix-XL runs MACE on 11M atoms with one A100","Post-training execution lifts MACE capacity 450-fold","No retraining: MACE hits 11M atoms on one A100"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000199,"raw_usage":{"cost":0.000198968,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000198968,"upstream_inference_prompt_cost":0.000125048,"upstream_inference_completions_cost":7.392e-05},"total_tokens":1362,"prompt_tokens":1098,"completion_tokens":264,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":167}},"tokens_in":1098,"tokens_out":264,"cost_source":"openrouter_usage","duration_ms":2282,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":167,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-02T20:21:05.815066+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Run the optimized backend on a task outside the tested battery—say a nudged-elastic-band transition-state or rare-event kinetics trajectory, or a checkpoint other than the MACE-OMAT-0-medium and MACE-MPA-0-medium ones used here—and compare forces and long-trajectory observables against mace-torch; deviations beyond the tested tolerances would weaken the unchanged-weights claim. A matched-cutoff benchmark equalizing graph construction at 6.0 Å on both sides would similarly test whether the 3.1–5.0× speedup survives removal of the disclosed skin asymmetry.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the MACE architecture, directed radius graph, and checkpoint contract that Symmetrix-XL executes unchanged."},{"cited_title":"Sobek: Streaming Equivariant Tensor Product Convolutions","cited_arxiv_id":"2607.18074","evidence_quote":"Supplies the streamed equivariant tensor-product reassociation on which streamed-edge execution builds."},{"cited_title":"PyTorch: An imperative style, high-performance deep learning library.Advances in Neural Information Processing Systems, 32:8024–8035, 2019","cited_arxiv_id":null,"evidence_quote":"The vendor-optimized cuEquivariance operator library that, with ML-IAP, forms the baseline whose complete-step time Symmetrix-XL must beat."},{"cited_title":"LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales","cited_arxiv_id":null,"evidence_quote":"LAMMPS is the molecular-dynamics engine inside which both the baseline and Symmetrix-XL throughput and capacity are measured."},{"cited_title":"mace-torch","cited_arxiv_id":null,"evidence_quote":"The reference mace-torch PyTorch implementation against which Symmetrix-XL numerical fidelity is qualified."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Matbench Discovery is the downstream benchmark whose WBM relaxation and discovery metrics must reproduce reference results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior Symmetrix inference backend that Symmetrix-XL extends with streaming, model specialization, and tiling."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}