Entity · paper

Analysing Self-Harm Representations in Language Models: a Cross-Architecture Study

paperactiveprovisionalanalysing-self-harm-representations-in-language-models-a-cross-architecture-study-25428cd2·1 events·first seen 4d ago

Aliases: Analysing Self-Harm Representations in Language Models: a Cross-Architecture Study

Co-occurring entities

X-Sensitive Gemma-3-4B-IT SH-Detection

More like this (12)

suicide and self-harm classifier Contrastive-Difference CKA Reveals Concept-Specific Structural Alignment Across Language Model Architectures Same Lesson, Different Story: Cross-Lingual Reconstruction of Cultural Narratives in Large Language Models Modeling Complex Behaviors: Multi-Personality Composition and Dynamic Switching in Vision-Language Models Apparent Psychological Profiles of Large Language Models are Largely a Measurement Artifact Exploring Adversarial Robustness and Safety Alignment in Multilingual Multi-Modal Large Language Models Visually Grounded Self-Reflection for Vision-Language Models via Reinforcement Learning Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models Reasoning Language Models Lesioned Multimodal Language Models Reproduce Aphasic Picture-Naming Patterns From Texts to Scores: Tracing the Emergence of Essay Quality Representations in Large Language Models Where Do Models Find Happiness? Emotion Vectors in Open-Source LLMs

Recent events (1)

5arXiv · cs.CL·4d ago·source ↗

Cross-architecture study of how LLMs internally represent self-harm content

Researchers analyze how four LLMs encode self-harm content using linear probes and contrastive direction extraction across two datasets (X-Sensitive and SH-Detection). A key finding is that self-harm information crystallizes in the final 3–7% of network layers (93–97% depth), and that probe accuracy does not correlate straightforwardly with linear separability. Gemma-3-4B shows a notably different internal representation of the contrastive self-harm direction compared to other models, with implications for detection, intervention, and LLM governance.

Evaluation and Benchmarking AI Safety Research X-Sensitive Gemma-3-4B-IT SH-Detection +1 more