简体中文 · English
Not All Is Lost: Repairing Lossy User Preference States of Personalization Encoders¶
L1 core mechanism diagnostic. Frozen-cache evidence selection and residual correction run. User-approved academic/Evolve exception; no production A/B.
Inclusion exception: The user explicitly approved this academic/Evolve mechanism on 2026-10-02. The paper has no quantitative online A/B and does not contribute to industrial evidence claims.
Paper information¶
| Field | Details |
|---|---|
| Paper | arXiv v1 |
| Institution | The first page does not list the first author's affiliation. |
| First public release | 2026-10-01 (arXiv v1) |
| Original source code | No public author implementation found as of 2026-10-02. |
| Adapter | repair-state |
| Local reproduction code | src/auto_research/reproductions/repair_state/ |
Original paper summary¶
Background and main changes¶
Freeze the encoder and task head. Compare per-timestep representations in the existing forward cache against the current preference state, select corrective evidence from long-term, recent or locally concentrated behavior, and apply a residual repair.
flowchart LR
I[Input and current state] --> M[Frozen-cache evidence selection and repair]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Paper figure: Core architecture, training flow or protocol. From the paper; copyright belongs to the authors. Click for the source.
Core equation¶
\(z'=z+\sum_{t\in TopK}softmax(q^T(h_t-z))(h_t-z)\).
Offline and online results¶
All 12 recommendation hosts improved in the paper. Mamba4Rec on MovieLens gained +3.96 MRR points, versus +0.19 for head-only. Personalized generation gains reach +25.23%.
Local reproduction¶
Comparison protocol: Disabled or default mechanism versus enabled core operator. Relative percentages are not applicable; metrics verify invariants, gradients or state transitions, not paper-scale effects.
- Three-seed diagnostic:
metrics/mechanism-seeds42-44.json. diagnostic_only=true; excluded from formal capability rankings.
Scope¶
Frozen-cache evidence selection and residual correction execute. This is a user-approved academic/Evolve exception without online A/B.