Study Traces Recommender's Last-Item Reliance to Residual Connections in Transformers

_reachsumit · x · 2026-08-17

Finding: Transformer-based causal self-attention recommenders (e.g., SASRec) often rely heavily on the most recent interaction at inference time, but the structural mechanism was unclear.

Conclusion: Research finds that while self-attention aggregates context, residual addition sharply shifts the representation toward the same-position contributions, termed "residual dominance."

Verification: Using inference-time residual scaling as an intervention, the study shows a monotonic trade-off between structural mixing and last-item reliance. Reducing residual strength recovers some hits where non-final positions already ranked the correct item.

Original post →

More from Research

Research channel →