Latent Dynamics · 2026-11-03

The (Butter)fly Dream: Linear Separability of World-Model Rollouts in a Connectome-Constrained Visual Model

Junichiro Niimi

arXiv:2611.01866PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors test whether a fixed connectome-constrained fly visual system (trained only for optic-flow estimation) can tell whether stimuli come from real video or from “dreams” generated by world models. Using a classifier two-sample test on temporal statistics of cell-type responses with sequence-level cross-validation, the authors report AUCs 0.85–0.99 for nine dream conditions in the trained observer, but 0.58–0.75 in the untrained observer.

Why it matters

It operationalizes “can a brain tell a world-model dream from reality?” as whether deviations in temporal structure become linearly decodable from a biological, task-trained response space. The setup isolates observer features (fixed brain) from generator imperfections (world models), and adds controls (untrained network, pixel-statistics control) plus graded temporal probes.

Method

  • Use a fixed published connectome-constrained fly visual network (50 independently trained models) as observer; collect per cell-type temporal statistics (type-mean).
  • Generate “dreams” by continuing five real frames into 14 generated frames using persistence random walk, MDN-RNN, and RSSM at three temperatures; compare to reality (MPI-Sintel) rendered to the same retina space.
  • Train a logistic-regression classifier two-sample test on observer temporal statistics with sequence-level disjoint cross-validation; report held-out AUC and do temporal perturbation probes (frame order swap/repeat) and a detection-horizon analysis (first k generated frames).
Abstract (from arXiv)

Zhuangzi dreamt he was a butterfly and, on waking, could not tell which was the dream. We ask a version of his question of a model of the fly brain: placed inside a world generated by a world model, can a fixed brain tell that the world is a dream? We treat it as a question about what a connectome-constrained model trained on one task makes available beyond that task: do its responses make the defects of a dream linearly decodable? The observer is a published model of the fly visual system trained only for optic-flow estimation; reality is the MPI-Sintel movie on the model's 721-column retina; dreams are continuations of five real frames by three small world models (a persistence random walk, an MDN-RNN and an RSSM) at three temperatures. The judge is a classifier two-sample test on temporal statistics of cell-type responses, with sequence-level cross-validation. In the trained observer's responses all nine dreams are separable from reality (AUC 0.85-0.99, mean over 50 trained models), in those of the same network before training much less so (0.58-0.75 for the learned world models; paired difference 0.20-0.37), equally on sequences the observer never saw, and in an order unrelated to pixel error. Because these AUCs are near ceiling, we add graded probes made of each window's own real frames with only their order broken. Swapping one adjacent pair of the 14 generated frames gives 0.73 on the trained responses against 0.54 before training and 0.56 for pixel statistics; the trained curve reaches 0.75 at 1.1 swapped pairs of seven, the pixel statistics at 4.7, and the untrained network not at all. A detection horizon, the test on the first k generated frames, orders the world models where the full-length AUC saturates.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First announced by arXiv: 2026-11-03 · Indexed by Pub-AI: 2026-11-03 UTC

preprint

The (Butter)fly Dream: Linear Separability of World-Model Rollouts in a Connectome-Constrained Visual Model

Junichiro Niimi

First announced by arXiv: 2026-11-03 · Original record ↗

arxiv
2611.01866
Explicit source versions (1)
  • v1 · Date not supplied
  1. Preprint available

    Source date: 2026-11-03 · First detected: 2026-11-03 UTC · Historical baseline

    Source evidence (1)
    • Official evidence ↗ · RSS guid=oai:arXiv.org:2611.01866v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv

Zhuangzi dreamt he was a butterfly and, on waking, could not tell which was the dream. We ask a version of his question of a model of the fly brain: placed inside a world generated by a world model, can a fixed brain tell that the world is a dream? We treat it as a question about what a connectome-constrained model trained on one task makes available beyond that task: do its responses make the defects of a dream linearly decodable? The observer is a published model of the fly visual system trained only for optic-flow estimation; reality is the MPI-Sintel movie on the model's 721-column retina; dreams are continuations of five real frames by three small world models (a persistence random walk, an MDN-RNN and an RSSM) at three temperatures. The judge is a classifier two-sample test on temporal statistics of cell-type responses, with sequence-level cross-validation. In the trained observer's responses all nine dreams are separable from reality (AUC 0.85-0.99, mean over 50 trained models), in those of the same network before training much less so (0.58-0.75 for the learned world models; paired difference 0.20-0.37), equally on sequences the observer never saw, and in an order unrelated to pixel error. Because these AUCs are near ceiling, we add graded probes made of each window's own real frames with only their order broken. Swapping one adjacent pair of the 14 generated frames gives 0.73 on the trained responses against 0.54 before training and 0.56 for pixel statistics; the trained curve reaches 0.75 at 1.1 swapped pairs of seven, the pixel statistics at 4.7, and the untrained network not at all. A detection horizon, the test on the first k generated frames, orders the world models where the full-length AUC saturates.