Live intelligence

Latest AI updates

2 updates

Research

Scaling Laws for SAE Training Data

This paper investigates how the number of unique activations affects sparse autoencoder (SAE) training when the total training budget is fixed. It finds that quality improves rapidly with buffer size but then plateaus, where replaying the buffer is nearly as effective as using fresh activations. The authors propose a scaling law for limited data and validate it across various models, layers, and SAE variants. They provide a practical rule: choose the smallest buffer that reaches the quality plateau and replay it, reducing activation storage by 8-64x with negligible quality loss.

arxiv14h agoScore 77

sparse autoencodersscaling lawsmodel interpretability

Research Editor selected

Logprob Reduction Fails to Verify Causal PII Suppression in LLMs

This paper shows that SAE-based feature interventions for suppressing PII leakage in LLMs, when evaluated only by log-probability reduction, often produce false positives. Across 36 cases, 47% appeared successful under logprob-only evaluation, but only 2.8% passed a multimetric framework with random direction controls. The authors identify two failure modes and propose causally grounded PII suppression via circuit-level localization and causal mediation analysis.

Why it matters: Provides concrete evidence that a common evaluation metric is unreliable)Skip and proposes a more rigorous methodology, directly impacting AI safety research.

arxiv2d agoScore 85

sparse autoencodersinterpretabilityprivacy