Scholarlink

<p>Large language models (LLMs) have improved the fluency of explanations in recommender systems. However, existing research largely focuses on mitigating factual hallucinations or preference misalignment, leaving a trust failure underexplored. In practice, these systems often exhibit a biased promoter persona: they generate explanations that are factually correct and preference-aligned, yet strategically incomplete by concealing item defects. We term this phenomenon Selective Disclosure, which produces true-but-incomplete explanations that erode long-term user trust. To address this, we propose STRATA, a risk-aware plan-and-generate framework that explicitly decouples disclosure decisions from language realization. Specifically, we introduce a Risk-Aware Preference Memory (RAPM) to distinguish between negotiable preferences and non-negotiable constraints. Guided by RAPM, a Strategic Plan Selector performs hierarchical disclosure planning by assigning attribute-level stances, while an Attribute-Grounded Generator translates these strategies into controlled natural language. An offline learning pipeline combining supervised warm-up and KL-regularized reinforcement learning optimizes the trade-off between disclosure sufficiency and evidence consistency. Experiments on three benchmark datasets show that STRATA reduces the selective disclosure of risk-relevant negatives and improves human-perceived trustworthiness and helpfulness over existing baselines. Our code is available at <a href="https://anonymous.4open.science/r/STRATA-C040">https://anonymous.4open.science/r/STRATA-C040</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mitigating selective disclosure in recommender explanations via risk-aware memory and hierarchical planning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月26日

查看全文
  • Chengkai Wang,
  • Baisong Liu

摘要

Large language models (LLMs) have improved the fluency of explanations in recommender systems. However, existing research largely focuses on mitigating factual hallucinations or preference misalignment, leaving a trust failure underexplored. In practice, these systems often exhibit a biased promoter persona: they generate explanations that are factually correct and preference-aligned, yet strategically incomplete by concealing item defects. We term this phenomenon Selective Disclosure, which produces true-but-incomplete explanations that erode long-term user trust. To address this, we propose STRATA, a risk-aware plan-and-generate framework that explicitly decouples disclosure decisions from language realization. Specifically, we introduce a Risk-Aware Preference Memory (RAPM) to distinguish between negotiable preferences and non-negotiable constraints. Guided by RAPM, a Strategic Plan Selector performs hierarchical disclosure planning by assigning attribute-level stances, while an Attribute-Grounded Generator translates these strategies into controlled natural language. An offline learning pipeline combining supervised warm-up and KL-regularized reinforcement learning optimizes the trade-off between disclosure sufficiency and evidence consistency. Experiments on three benchmark datasets show that STRATA reduces the selective disclosure of risk-relevant negatives and improves human-perceived trustworthiness and helpfulness over existing baselines. Our code is available at https://anonymous.4open.science/r/STRATA-C040.