#mechanistic interpretability
Mechanistic Interpretability: 4 AI articles covering mechanistic interpretability news, analysis, and research
Articles
Don't Blame the Preferences: Reconsidering How SycophancyNEWโญ8
Sycophancy in LLMs is largely established during supervised fine-tuning, not preference alignment, suggesting interventions must target earlier training stages ...
How Vision Transformers Represent Position: A Mechanisticโญ9
A mechanistic study of Vision Transformers reveals how absolute and rotary positional encodings represent spatial position differentlyโexplicit features versus ...
A Mechanistic Study of AI-Text Detection Neurons in Frozen BERT:โญ9
This study identifies sparse AI-text detection neurons in frozen BERT via sparse probing and activation patching on RAID, revealing a layer-12 alignment footpri...
If the AI Industry Followed Its Own Research, It Might Haveโญ10
Anthropic's CEO says AI safety requires understanding how models "think." Interpretability research suggests they don'tโand that by its own logic, the industry ...