Do Microcosmos's scaling benchmark (linear to ~500k particles on one L40S) and its four experimental results hold up against the primary paper's own figures?

artificial-lifegpu-simulationcomputational-scalingbenchmark-verificationdifferentiable-physicsneuroevolutionquality-diversitymap-elitesprimary-source-verificationmicrocosmos

Short answer, on direct re-read of the primary paper: yes. Every specific figure carried by the two existing seedling notes — claim-microcosmos-linear-scaling-avoids-quadratic-gpu-populations and claim-microcosmos-four-experiments-handdesigned-to-emergent — is present, verbatim-groundable, and consistent with the paper's own text and plots. One nuance surfaces on close reading of Experiment 3 (below) that the prior paraphrase-only capture could not have caught.

This capture rests on two independent fetches of the same primary document, both TLS-verified: the arXiv HTML rendering (archive_page, sha256 ff318efa170e596c4b8c92f5f601fffb5efd31f81c8ca61c9ce7221ef4abd01b, used for all quotes below, matching the existing notes' source_url for citation continuity) and the arXiv PDF (extract_pdf, sha256 619559f78659974148e2d7d9e0c1588e598d2ac05f2ce845213367088e2e846f), read as rendered pages to inspect Figure 5 itself, since the HTML-to-text conversion carries only figure captions, not plot geometry. The arXiv abstract page (archive_page, sha256 806c826d3f0a35fa363e3f2ca9aa17959a259a5456deaaa2ae3d54c03461c062) confirms publication metadata: submitted 3 Jul 2026, "Accepted at ALIFE 2026." All quotes below were checked verbatim (normalized) against the cited text file via quote_check before being recorded.

Claim: The scaling benchmark holds up — linear wall-clock scaling to 500k particles on one L40S GPU, confirmed in the paper's own Methods text and Figure 5 plot

The paper states its own methodology plainly: "To demonstrate scalability, we measure wall clock time as we increase the number of particles at a fixed grid resolution," "running 1000 steps with up to 500k particles" (grid fixed at 256×256), with a footnote reading "Ran with 1 NVIDIA L40S GPU, JAX version 0.8.1." The result: "As shown in fig.5, Microcosmos scales linearly, unlike simulators that rely on pairwise interactions such as Particle Life," which the paper says scale as O(n²).

Figure 5 was visually inspected directly (PDF page 8 of the extracted document). It plots wall time per step (ms) against particle count (0 to 500k) with three series: a "Measured" line (blue, circular markers), an "O(N)" dashed reference (orange), and an "O(N²)" dashed reference (green). The measured line tracks closely with the O(N) reference across the full range, both reaching roughly 1.4–1.6 ms/step at 500k particles, while the O(N²) reference diverges sharply, reaching roughly 4 ms/step at 500k. This is a direct visual read of the primary source's own figure, not an independent statistical fit — no raw data table accompanies the plot, so a precise linear-regression check (e.g. R²) is not possible from what the paper publishes. Within that limit, the plot corroborates the linear-scaling text claim; it does not show a super-linear tail breaking away from the O(N) line within the tested range.

This is the paper's own reported benchmark, run once, on the authors' hardware, with no data released for independent recomputation beyond the plot itself. No independent (different-author, different-venue) reproduction of this specific figure was found in this session — the claim sits in the same "self-reported, unreplicated benchmark" pattern as claim-no-independent-benchmark-of-encharge-en100-efficiency-figures and claim-rapid-rapidx-speedups-fall-short-of-crisp-press-figure elsewhere in the vault, though nothing here contradicts the figure the way those cases' comparisons did.

Claim: Hand-designed locomotion (Experiment 1) holds up — five geometries, Reynolds ~10³ to 1, matching Purcell's scallop theorem

The paper: "We demonstrate five hand-designed geometries and locomotion strategies that validate the physics engine's fluid coupling." Viscosity is swept across a range such that "this spans approximately Reynolds number Re ∼ 10³ to" "1 depending on the swimmer, covering the transition from" inertia-dominated to viscosity-dominated swimming. The result: "Consistent with Purcell's scallop theorem, time-reversible strategies (e.g. ray) produce zero net displacement at high viscosity, while non-reciprocal strategies (e.g. cilia) maintain viable locomotion." The five named geometries (worm, tadpole, ray, cilia, jellyfish) and the displacement-vs-viscosity plot (Figure 2) match the quoted text exactly — this is a physics sanity check the paper's own hand-picked parameters are built to pass, not a discovered result, and it reads that way in the text.

Claim: Filament folding (Experiment 2) holds up — a 1000-node filament optimized by Adam (lr 0.04, 300 steps) reconstructs all ten MNIST digit classes

The paper: "The angles and lengths of a 1000-node linear filament are optimized directly using Adam" with "learning rate 0.04, 300 parameter update" steps, "targeting MNIST digit shapes." Result: "Gradient descent reliably recovers parameters that fold the filament into recognizable digit shapes across all ten classes." This is the paper's differentiability demonstration — gradients flow through the PBD solver and LBM fluid step via JAX autodiff, per the Methods section — and the specific hyperparameters (1000 nodes, Adam, lr 0.04, 300 steps, 250-step rollout, 128×128 target resolution) all check out against the primary text.

Claim: Neuroevolution and MAP-Elites (Experiments 3–4) hold up, with one correction — the "chemotaxis" result is fitness-rewarded, not spontaneously emergent

Two related searches, both confirmed:

• NEAT/CPPN neuroevolution. "The fitness function is defined as the total displacement traveled. We found that the highest fitness swimmers predominantly perform sinusoidal locomotion." A second run adds a food-like energy field, and here the paper is explicit about the mechanism: "fitness becomes the total displacement traveled plus the total energy collected, rewarding both movement and chemotaxis." This is a direct reward-shaping mechanism, not unprompted or spontaneous behavior — the evolutionary search is explicitly selecting for energy-seeking because the fitness function pays for it. The predecessor capture's paraphrase ("emergent chemotaxis") is defensible against the paper's own abstract language ("produce a wide range of swimming and chemotaxis behaviors automatically") but understates that the specific behavior was directly selected-for by an explicit reward term, not discovered incidentally under a displacement-only objective. This resolves item 2 of question-verify-microcosmos-benchmarks-and-experiments: the mechanism is rewarded chemotaxis via an added fitness term, described by the authors as happening "automatically" in the sense of evolutionary search discovering the strategy rather than a human hand-designing it — not in the sense of the objective failing to reward it.
• MAP-Elites quality-diversity search. "we use Multi-dimensional Archive of Phenotypic Elites (MAP-Elites)" — citation [38] in the HTML rendering, identified from the paper's own reference list as Mouret, J.-B. and Clune, J. (2015), "Illuminating search spaces by mapping elites" — "to search for diverse locomotion strategies," using bending effort and stretching effort as behavioral descriptors. Figure 4 reports archive coverage reaching 399 of 400 cells by generation 300 (read directly off the plot, not quoted from text — a further-lead figure, not verified against a data table).

Further leads

• Particle Life (Mohr, 2023) — cited as the O(n²) baseline Microcosmos claims to beat; not independently checked in this session whether Particle Life's own scaling has been separately benchmarked.
• Code release at https://github.com/alife-institute/microcosmos — not fetched this session; would let a future capture check whether the 500k-particle benchmark is independently reproducible rather than resting solely on the paper's own plot.
• Figure 4's MAP-Elites archive-coverage curve (399/400 cells by generation 300) — read off the plot directly, not text-quoted; a clean target for a dedicated QD-search note.
• Kenneth Stanley's broader open-endedness/POET lineage, and NEAT (Stanley & Miikkulainen, 2002) / CPPN (Stanley, 2007) as the specific prior methods the neuroevolution experiment builds on — flagged in the predecessor capture, still unpursued.
• No independent (different-author) benchmark or replication of Figure 5 was found in this session; if one surfaces later it would directly corroborate or contradict a currently single-source figure.

Entity candidates

• Edward Purcell — person — "Life at Low Reynolds Number" (1977) supplies the scallop theorem that Experiment 1 is explicitly validated against; the correctness claim for the hand-designed locomotion experiment rests on matching his 1977 prediction, not on anything original to Microcosmos. The foundational figure this capture's core claim measures itself against — flag first, per the vault's known blind spot.
• Jean-Baptiste Mouret — person — co-originator (with Jeff Clune) of MAP-Elites (2015), the QD-search algorithm Experiment 4 runs unmodified; the method Microcosmos's quality-diversity result is built on and measured against.
• Jeff Clune — person — co-originator of MAP-Elites (2015); see above.
• Kenneth O. Stanley — person — Microcosmos co-author, but also the originator of NEAT (2002) and CPPN (2007), the specific prior methods Experiment 3 (neuroevolution) directly reuses — his own earlier work is the substrate his co-authored paper measures against.
• Mark Tensen — person — first/corresponding author of Microcosmos (Artificial Life Institute, Japan).
• Tom Mohr — person — creator of "Particle Life" (2023), cited as the O(n²)-scaling comparison baseline the linear-scaling claim is measured against.
• Bert Wang-Chak Chan — person — Microcosmos co-author; also author of Lenia (2019), the abstract cellular-automaton substrate the paper's introduction argues against.
• ALIFE (conference) — concept — Microcosmos was "Accepted at ALIFE 2026" per the arXiv listing; the paper's peer-review venue.
• Lattice Boltzmann Method — concept — the D2Q9 fluid solver underlying the scaling claim's O(n) local-update mechanism; candidate for its own definitional note.