Disentangling Language Effects on Tool Admission and Post-Retrieval Safety

Key points

• It remains unclear whether language affects tool admission, post-retrieval generation, or both.1
• We independently vary request, retrieved-page, and required response languages in a controlled URL-retrieval setting across ten models and seven languages.1
• We identify Safety Loss through Interlingual Processing (SLIP).1
• Requiring answers in the page language increases pooled harmfulness on matched tool calls from 50.3% to 57.4% across four closed-weight models, mainly in Chinese and Hindi.1

Sources (1)

• [1]Disentangling Language Effects on Tool Admission and Post-Retrieval Safety
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 16, 08:05 AM

  LLM agents increasingly use retrieval tools across languages, yet their safety evaluation remains English-centric.

  It remains unclear whether language affects tool admission, post-retrieval generation, or both.

Extractive summary: sentences quoted from the sources.