简体中文 · English

Measuring and Mitigating Identity-Cue Preference Drift in LLM-based Recommender Systems

L1 core mechanism diagnostic. List metrics and postprocessing reranker run. Identity slices and relevance come from public fixtures; no real user profiles are generated.

Paper information

Field	Details
Paper	arXiv 2609.34229
Institution	University of Electronic Science and Technology of China
First public release	2026-09-28 (arXiv v1)
Original source code	No public author repository found as of 2026-09-30.
Adapter	promptshift
Local reproduction code	src/auto_research/reproductions/promptshift/

Original paper summary

Background and main changes

Hold behavior history fixed and compare identity-cued rankings with a reference without identity. Drift and SliceShift measure the resulting shift. Mitigation blends original score and inverse group popularity according to each user's mainstreamness.

flowchart LR
  I[Public input and current state] --> M[PromptShift core mechanism]
  M --> A[Auditable intermediate values]
  A --> O[Output or state update]

Key figure from the paper

Figure 1: Overall method and components. From the paper; copyright belongs to the authors. Click for the source.

Core equations

For reference list \(R\) and identity-cued list \(C\), calculate \(Drift(R,C)\) from set-overlap loss and normalized rank changes for shared items, and \(SliceShift\) from group popularity differences. The reranker uses mainstreamness \(m\) to mix normalized relevance with inverse popularity: \(s'=(1-m)\hat{s}+m(1-p)\).

Offline and online results

Benchmarks, speed and training curves in the paper are the authors' results. The three-seed mini-suite checks only mechanisms and invariants. It neither compares at paper scale nor extrapolates to online gains.

Local reproduction

See metrics/mechanism-seeds42-44.json. The results are marked diagnostic_only=true and cannot enter formal capability rankings.

Comparison protocol: Unmitigated identity-cued ranking versus adaptive inverse-popularity reranking. Relative improvement is not applicable to this L1 mechanism diagnostic.

Scope

Only list metrics and postprocessing reranking execute. Public fixtures supply identity slices and relevance; no real user profiles are generated.