Study: LLMs represent self-harm in final network layers

Researchers have analyzed how language models represent self-harm content, a critical task for intervention and user safety. Their study, which trained linear probes across model layers, found that self-harm information crystallizes in the final 3-7% of a model's layers. The analysis also revealed that Gemma-3-4B represents contrastive self-harm directions in a more intricate manner compared to other tested LLMs. AI

IMPACT Provides insights into LLM capabilities for detecting sensitive content, crucial for safety applications.

RANK_REASON Academic paper on language model capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Gemma 3-4B
• Luis Espinosa Anke
• SH-Detection
• X-Sensitive

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →