Alibaba’s SpecFormer softens embedding spectra to fight attention collapse in recommendation

_reachsumit · x · 2026-07-28

Alibaba's SpecFormer tackles a recommendation-system failure mode the authors call embedding and attention collapse. They argue that directly applying standard Transformers to recommendation often underperforms simpler models because long-tail, heterogeneous recommendation data causes the singular-value spectrum of embeddings to collapse around a few dominant directions.

The paper proposes three mechanisms: a learnable spectral softening module to smooth the singular-value distribution of input embeddings, a spectrum-softened attention module, and a spectral residual position encoding. The goal is to keep the representation space from collapsing so the Transformer can scale better with depth in recommendation settings.

Original post →

More from Research

Research channel →