Linear Pain Representation Identified in 25 Open-Weight Models
September 16, 2026
Researchers identified a linear pain direction across 25 open-weight models ranging from 2B to 72B parameters. This direction is nearly orthogonal to fear and negative valence, suggesting LLMs represent pain as a distinct internal dimension across both base and instruction-tuned architectures.
HOW THIS AFFECTS YOU
●
researcherYou can use these findings to probe for specific internal representations of affective states in larger models.
●
policyThis provides empirical evidence for studying how models represent concepts related to self-harm and distress.