MovieLens historical-association effect

Determine whether preserving the correct assignment of historical ratings to items improves rating prediction on MovieLens for the evaluated Qwen and Phi readers, beyond information contained in the rating multiset and item metadata.

Background

The study tests historical-association use by permuting ratings within each user’s history while preserving the rated items, item metadata, history length, and exact rating distribution. On MovieLens, the permutation increases mean absolute error by only 0.020 for Qwen and 0.029 for Phi, and the family-adjusted intervals include zero for both readers.

Accordingly, the paper treats MovieLens as an unresolved boundary case rather than as a confirmed replication of the positive Coat association effect. Further evaluation is needed to determine whether the observed assignments contain reliably useful item-specific information under this MovieLens representation and sample.

References

MovieLens supplies an important boundary: its association effects remain inconclusive after adjustment, and simple numerical readers remain strong.

— A Controlled Audit of Personal AI Memory for Rating Prediction  (2610.02764 - Gupta, 2 Oct 2026) in Section 4.2, “Association use depends on the domain”; Section 5, Conclusion