Research has found that the reconstruction scores of natural language autoencoders (NLA) cannot verify the truthfulness of individual statements; the model...
Research has found that the reconstruction scores of natural language autoencoders (NLA) cannot
verify the truthfulness of individual statements; the model may rely on 'private codes' rather than actual evidence. The authors propose the RECAP method, which jointly trains a linear head on the target model to keep the specified content decodable. On Pythia-160M, independent probes can reliably distinguish true and false statements (AUC 0.96) and still mark lies under adversarial edits (AUC 0.95).
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs ：https://info.arxiv.org/labs/index.html.
论文提出RECAP，通过在目标模型上联合训练线性头，监督指定内容保持可解码，以检验激活解释是否逐声明忠实。摘要称其在Pythia-160M上取得较高真假声明区分效果。
研究指出，自然语言自编码器的重建分数不能验证逐声明忠实性，模型可能依赖“私密代码”而不是真实依据。RECAP试图以内容可解码性补充这一验证路径。
Aioga判断，RECAP的重点不只是提高重建表现，而是把“指定内容能否被独立解码”作为可检验标准。摘要中的结果显示，该思路值得关注，但材料未说明其适用范围。
若相关方法在更多模型和任务中成立，激活解释评估可能从整体重建分数转向逐声明核验。当前材料仅报告Pythia-160M结果，不能据此推断普遍有效性。 值得关注后续研究是否报告更大规模模型、不同数据与更多对抗编辑设置下的表现，并进一步说明独立探针如何区分真实依据、私密代码与其他可解码信号。
The readable text on this page was extracted from the public source and organized with attribution, publication time and the original link. Copyright remains with the original author and publisher.
Ingestion channel: Summary aggregation · Source domain: arxiv.org
Source: HuggingFace Daily Papers（社区热门论文）
Original link: Open original source
Aioga archive: Open intelligence page
Content record: source-page · Updated: 2026-07-22T00:00:00.000Z
统一接入主流 AI 模型 API，为开发、测试与生产环境提供稳定调用入口。
立即访问 api.w173.com
Scan to open this articleAioga aggregates global AI updates and preserves source information for verification and citation.