---
title: "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?"
type: "question"
status: "open"
date_raised: "2026-07-09T00:00:00.000Z"
tags: ["artificial-life","gpu-simulation","quant-verification","primary-source-verification","microcosmos","quote-verification"]
---


Raised while promoting the capture behind
[[claim-microcosmos-linear-scaling-avoids-quadratic-gpu-populations]] and
[[claim-microcosmos-four-experiments-handdesigned-to-emergent]]. Both notes rest on a
hop capture whose author read arXiv:2607.02954 (arxiv.org/html/2607.02954v1) directly
but preserved a *summary*, not verbatim figures or quotes. Two gaps to close from the
paper's own text and plots:

1. **The scaling benchmark (`[unverified-quant]`).** The capture reports runtime
   scaling **linearly with particle count, tested to ~500k particles on a single
   NVIDIA L40S GPU**. Confirm from the paper's own benchmark figures: (a) that the
   measured scaling is actually linear (not merely sub-quadratic) across the tested
   range; (b) the exact particle ceiling and the GPU used; (c) what "runtime" measures
   — steps/second, wall-clock per simulated second, or memory-bound throughput. The
   capture author explicitly flagged that "exact benchmark numbers/methodology beyond
   this summary should be checked against the paper's own figures before reuse."

2. **The four experiments (`[unverified-quote]`).** No verbatim quote was preserved for
   §experiments. Verify from the paper body: the Reynolds-number range (~10³ to ~1) and
   the Purcell scallop-theorem match; the 1000-node MNIST-folding differentiability
   demo; that NEAT/CPPN neuroevolution produced *emergent* chemotaxis (vs. rewarded
   chemotaxis); and the MAP-Elites gait-diversity result. Capture a verbatim phrase for
   each so the notes can be upgraded past `seedling`.

Medium priority. The engine's *design* (physically grounded differentiable filaments,
quadratic-scaling avoidance) is well grounded in the abstract/methods; what needs the
primary re-read is the specific benchmark number and the experimental claims. Both
notes stay `seedling` until then.