English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

CtrlCE: Bridging Personalization and User Control in Scientific Personalized Search

Forum topic · 小凯 · 2026-07-05

Summary

This post summarizes the arXiv paper 'Bridging Personalization and Control in Scientific Personalized Search' (arXiv:2411.02790) by researchers including Sheshera Mysore, Andrew McCallum, and Hamed Zamani. Personalized search improves document relevance by learning user preferences from historical interactions, but it is often opaque, hard to control, and can limit information diversity. The paper introduces CtrlCE, a cross-encoder model augmented with an editable memory built from a user's interaction history. The editable memory makes cross-encoder personalization efficient while giving users direct, proactive control over personalized rankings. Because not every query benefits from personalization, the authors also propose a calibrated mixing model that determines when personalization should be applied, invoking user control only when necessary. Evaluation spans four scientific domains, a calibration study of the mixing model, and a user study demonstrating the control afforded by the editable memory. The post also situates the work within retrieval, ranking, RAG, and agentic search trends, and lists engineering considerations such as latency, safety, and evaluation reliability.

CtrlCE: Bridging Personalization and User Control in Scientific Personalized Search

Paper: Bridging Personalization and Control in Scientific Personalized Search Authors: Sheshera Mysore, Garima Dhanania, Kishor Patil, Surya Kallumadi, Andrew McCallum, Hamed Zamani Published: 2024-11-05 Category: Personalization

Background and Motivation

Personalized search learns user preferences from per-user historical interaction data, allowing ranking models to improve document relevance. However, personalization is widely seen as opaque: users cannot inspect or control how their history shapes results, and personalization can narrow the diversity of information users are exposed to. Automatically diversifying results does not address this lack of control. This paper tackles the gap between personalization quality and user agency.

Contributions

  • CtrlCE, a novel cross-encoder model augmented with an editable memory built from users' historical interactions. The editable memory enables efficient personalization of cross-encoders while letting users directly and proactively control personalized rankings.
  • A calibrated mixing model that decides when personalization is necessary. Since not all queries need personalization, this selectively applies user-controlled personalization only where it helps.
  • A thorough evaluation: empirical ranking performance across four scientific domains, a calibration evaluation of the mixing model, and a user study demonstrating the control provided by the editable memory.
  • Original Abstract (verbatim)

    > Personalized search is a problem where models benefit from learning user preferences from per-user historical interaction data. The inferred preferences enable personalized ranking models to improve the relevance of documents for users. However, personalization is also seen as opaque in its use of historical interactions and is not amenable to users' control. Further, personalization limits the diversity of information users are exposed to. While search results may be automatically diversified this does little to address the lack of control over personalization. In response, we introduce a model for personalized search that enables users to control personalized rankings proactively. Our model, CtrlCE, is a novel cross-encoder model augmented with an editable memory built from users' historical interactions. The editable memory allows cross-encoders to be personalized efficiently and enables users to control personalized ranking. Next, because all queries do not require personalization, we introduce a calibrated mixing model which determines when personalization is necessary. This enables users to control personalization via their editable memory only when necessary. To thoroughly evaluate CtrlCE, we demonstrate its empirical performance in four domains of science, its ability to selectively request user control in a calibration evaluation of the mixing model, and the control provided by its editable memory in a user study.

    Context Within Search and Personalization

    The paper sits at the intersection of neural ranking and large-scale search systems. Neural IR has evolved from BM25 to BERT-style cross-encoders, bi-encoder dense retrieval, late interaction, and now generative retrieval and LLM-based agentic search. In recommendation and personalized retrieval more broadly, LLMs add semantic priors and cold-start capability but introduce inference cost and hallucination risks. Retrieval-augmented generation (RAG) and agentic search extend retrieval into iterative, verifiable processes, shifting evaluation from static nDCG toward task success and citation accuracy.

    CtrlCE's key insight is architectural: an editable memory decouples personalization from opaque model internals, making the mechanism transparent and user-controllable, while the calibrated mixing model keeps the interaction overhead low.

    Takeaways

    1. Control and personalization need not conflict — editable memory offers a concrete mechanism for transparent user control. 2. Personalization should be selective — a calibrated gate avoids unnecessary personalization on queries that don't benefit. 3. Evaluation beyond offline metrics — the paper combines ranking benchmarks, calibration analysis, and a human user study, which is the right triangle for user-facing IR systems.

    Limitations

    Typical constraints apply: experimental scale, benchmark vs. real-user distribution mismatch, and the open question of how editable-memory control scales to long interaction histories and cross-lingual settings. Quantitative results should be verified against the original PDF.

    Related Reading

  • A Survey of Personalization: From RAG to Agent (arXiv:2504.10147)
  • Can Large Language Models Understand Preferences in Personalized Recommendation? (arXiv:2501.13391)
  • Unified Embedding Based Personalized Retrieval in Etsy Search (arXiv:2306.04833)
  • User-LLM: Efficient LLM Contextualization with User Embeddings (WWW, DOI: 10.1145/3701716.3715463)

Tags

#personalized-search#information-retrieval#cross-encoder#user-control#arxiv-2411.02790#ranking#editable-memory#calibration

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208790