Scholarlink

A central question arises when budgeted cascades for automated claim verification are created: should additional computation be chosen based on a model’s confidence or its faithfulness to its explanation? In this study, a two-stage framework is proposed where an interpretable linear classifier produces a necessity-sufficiency faithfulness score to gate allocation, following the ERASER protocol. High faithfulness cases are resolved cheaply by the first stage; others require a stronger model, with minimal overhead associated with the routing decision. Evaluated on PubHealth and LIAR under matched budgets (20 to 60%), our method reduces median latency by 1.5 to 3x compared to a strong confidence-gated baseline. It matches or slightly outperforms confidence on PubHealth, and it meets or slightly exceeds faithfulness on LIAR, even though the macro-F1 differences are not statistically significant. These findings demonstrate that explanation faithfulness provides a practical alternative to confidence for affordable verification, offering an optimal trade-off between efficiency, performance, and transparency.

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

Faithfulness-Gated Cascades for Low-Latency Explainable Claim Verification Using Interpretable Linear Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bahareh Daneshvar,
  • Rahim Pasbanigoloojeh,
  • Ahmed Alrowaili,
  • Nadhmi A. Gazem,
  • Zeyad Ghaleb Al-Mekhlafi,
  • Mujeeb Ur Rehman

摘要

A central question arises when budgeted cascades for automated claim verification are created: should additional computation be chosen based on a model’s confidence or its faithfulness to its explanation? In this study, a two-stage framework is proposed where an interpretable linear classifier produces a necessity-sufficiency faithfulness score to gate allocation, following the ERASER protocol. High faithfulness cases are resolved cheaply by the first stage; others require a stronger model, with minimal overhead associated with the routing decision. Evaluated on PubHealth and LIAR under matched budgets (20 to 60%), our method reduces median latency by 1.5 to 3x compared to a strong confidence-gated baseline. It matches or slightly outperforms confidence on PubHealth, and it meets or slightly exceeds faithfulness on LIAR, even though the macro-F1 differences are not statistically significant. These findings demonstrate that explanation faithfulness provides a practical alternative to confidence for affordable verification, offering an optimal trade-off between efficiency, performance, and transparency.