arXiv cs.CL· Giovana Kerche Bon\'as, Thales Sales Almeida, Marcelo da Silva Reis, Rodrigo Nogueira, Marcos Meideiros Raimundo·· 2 天前AI 评分30
PRISM：工具调用语言模型激活干预的迁移与局限
PRISM: Transfer and Limits of Activation Interventions in Tool-Calling Language Models
AI 导读
研究团队提出 PRISM 框架，评估激活干预在精度、语言和规模上的迁移能力。用 NF4 量化副本的行为标签可在极少场景下稳定估计干预方向，五语言 NF4 监督方向无需目标语言微调即可迁移至阿拉伯语，将 Llama-3.1-8B 的不安全违规降低 21.95 个百分点，并借助配对良性激活迁移至最大 70B 参数的模型。实验代码与交互式结果浏览器已公开。
正文 · 原文
Abstract:Recent literature shows that unsafe tool calls can often accompany textual refusals in language models. To investigate this hidden vulnerability, we introduce PRISM, a framework for evaluating the transfer of activation interventions across precision, languages, and scales. Our findings demonstrate that leveraging lower-resource settings (specifically quantized NF4 labels and cross-lingual transfer) offers a robust pathway to expose and improve model alignment. Notably, using behavioral labels from an NF4 copy enables stable intervention estimation even with very few scenarios. Furthermore, these alignment improvements transfer effectively across languages: five-language NF4-supervised directions successfully transfer to held-out Arabic, reducing unsafe violations in Llama-3.1-8B by 21.95 percentage points without target-language tuning. By mapping paired benign activations, we also successfully transfer these safety interventions to much larger models (up to 70B parameters). Ultimately, PRISM highlights how auxiliary supervision from compressed models and multilingual data can uncover and mitigate hazardous tool-calling behaviors across different model families. We release all experimental code and supporting artifacts at this https URL, alongside an interactive explorer of recorded model responses and experimental results at this https URL.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.20808 [cs.CL] |
| (or arXiv:2610.20808v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.20808 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Giovana Bonás [view email]
[v1]
Wed, 14 Oct 2026 18:55:23 UTC (176 KB)
来源：arXiv cs.CL · arxiv.org