Frozen-backbone slot binding exceeds feature clustering, with a characterized ceiling

Resolution (supersedes the earlier "negative" report)

An earlier report on this line concluded that frozen-backbone slot binding did not exceed naive clustering of the same features (an apparent negative). That conclusion is now superseded: the apparent negative was traced to an implementation bug, and with it fixed the result flips — slot learning on a frozen visual backbone genuinely exceeds clustering. This note records the resolved result and its honest ceiling.

The bug (root cause of the apparent negative)

The reconstruction objective normalized its target with a per-dimension z-score (whitening) — a step the reference object-centric method does not use. Whitening forces the decoder to reconstruct low-variance / noise dimensions on equal footing, which starves the gradient that drives object grouping. Removing it (reconstructing the raw frozen features) flips the outcome: within a few thousand steps the binding moves from below the no-training baseline to clearly above it.

What is now established

• A frozen visual backbone + iterative slot attention + a transformer (mask-competition) decoder learns an object partition that exceeds k-means of the same frozen features, with a margin over both the k-means baseline and a shuffled-mask control.
• The learned partition is independent of clustering — its agreement with the k-means partition is low (around 0.35), so it is not merely reproducing clustering; it is a genuinely different, better object grouping.
• Verified across mask source (decoder and encoder readouts), evaluation resolution (patch-grid and full-image), and a real instance-segmentation gold (not box-derived).

The ceiling (characterized, honest)

• The learned binding plateaus early (by a few thousand steps) at a low-0.2 range alignment with real objects (around 0.24), modestly above the clustering baseline (around 0.21). It does not reach the reference method's reported ~0.4.
• This ceiling was ruled out as: gold-yardstick (box-derived vs. true segmentation), evaluation protocol (patch-grid vs. full-resolution), feature source (multiple frozen backbones cluster to a similar level), and caching / augmentation (a non-cached run with real image augmentation plateaus at the same level).
• It is therefore an architecture / recipe ceiling. Closing the remaining gap to the reference would require either very-large-scale training or deep reference-implementation fidelity work — out of scope for this line.

Honest framing (the durable conclusion)

Visual binding is not impossible — but frozen-feature reconstruction yields only a weak-to-medium binding signal. It should not be treated as full object discovery. For a state/world-model substrate it is best used as a semantic carrier / weak binding prior, not a complete perceptual segmenter.

What this changes for the research line

The open question is no longer "can binding be learned" (yes) but whether this weak binding prior is useful downstream: does feeding it into a semantic-state graph improve the model's downstream semantic closure? That — not a higher segmentation score — is the next milestone. Chasing the segmentation number further would be benchmark-chasing.