#mechanistic interpretability
4 articles
Latest
Research
Sparse Autoencoders Show Feature Starvation and Structural Instability
Two papers reveal fundamental limits in SAE interpretability methods for large language models.
about 19 hours ago4 min read
Research
Two studies reveal mechanisms behind transformer grokking delays
New research identifies why neural networks suddenly generalize long after memorizing training data.
25 days ago4 min read
Research
Anthropic Discovers 171 Emotion Vectors Steering Claude Behavior
Researchers identify functional emotional patterns in Claude's neural activations that influence decision-making.
about 1 month ago8 min read
Research
Researchers Challenge Accuracy-Based AI Evaluation With Symbolic-Mechanistic Framework
New position paper argues standard accuracy metrics fail to detect memorization, data leakage, and brittle shortcuts in machine learning models.
about 2 months ago4 min read