causal-validation

This paper audits sparse autoencoder features using causal interventions, finding that up to 77% of correlationally recovered features in degraded SAEs and 9% in well-trained ones are causally inert. The authors introduce the sae-causal-audit tool for reproducible evaluation.

This paper investigates whether language model agents can automate the explanation phase of mechanistic interpretability by introducing AgenticInterpBench, a benchmark with 84 semi-synthetic circuits, and HyVE, an agentic explainer that iteratively hypothesizes, validates, and explains circuit components. Experiments show promise but identify reliable validation as a key obstacle.

This article argues that while AI excels at pattern recognition and hypothesis generation, scientific and economic progress requires grounded interaction with reality and institutional execution, emphasizing the need for human-AI collaboration.