Any language gradient into a Gumbel discrete bottleneck collapses symbols or blocks learning; a gradient-cut plus non-parametric memory table plus DP-Means splitting achieves 79–100% grounding with zero collapse.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Write-Protected Discrete Bottlenecks for Language-Grounded World Models: A Structural Limitation and Sufficient Fix
Any language gradient into a Gumbel discrete bottleneck collapses symbols or blocks learning; a gradient-cut plus non-parametric memory table plus DP-Means splitting achieves 79–100% grounding with zero collapse.