Scaling Laws for Safety Probes
Researchers from the University of Washington and Anthropic show that the error rate of activation-based safety probes follows predictable power-law scaling with both model size and number of training labels. Tested across 20 base models, five model families, and six newly built safety datasets, the resulting laws predict a next-size-up model's probe error within 0.02 in 72% of cases and correctly identify the cheaper scaling path (more labels vs. bigger model) in 73% of cases.
Why it matters
Safety probes monitoring model activations for harmful content currently require ad hoc retraining and evaluation every time a new model or harm category appears, with no way to predict performance in advance. A scaling law lets safety teams budget labels and compute for a target probe accuracy instead of trial-and-error, which could make real-time safety monitoring cheaper and more systematic to maintain as models evolve.
What is verified
Researchers from University of Washington and Anthropic trained two probe architectures on 20 base models from five model families, using 32 to 10^6 labels per model, across six newly constructed safety datasets. [1]
Probe error (1 - AUROC) follows two power-law scaling relationships: it falls as a power law with model size, and decreases toward a floor as the number of labeled examples increases, with the slope of that decrease growing with model size. [1]
When fitted on the smallest models in a family, the scaling laws predict the probe error of the next larger model to within 0.02 in 72% of cases, and correctly identify the cheaper path (more labels vs. larger model) to a target error in 73% of cases. [1]
The laws also select the better of two probe architectures, incurring only 0.006 average extra error when they pick the worse one. [1]
What is not
- No independent replication or peer review is reported; this is a single arXiv submission.
- Whether these scaling laws generalize to safety probes used in production monitoring systems beyond the tested model families is untested.
- The paper does not state whether the datasets or probe architectures are being released publicly.
- No comparison to non-probe-based safety monitoring approaches is given.
Sources
- 1arXiv cs.LGarxiv.org · Primary source
Original source published 22 October 2026 at 05:43.
An AI-assisted briefing by Coding Partners. We link the evidence behind the reporting so you can explore the original work and assess the claims. How CP Signal works.