AI Science Tracker
EN

Safety probe errors followed predictable scaling across three model families

University of Washington; Anthropic · 22 Oct 2026 · computer science

Discovery

Safety probe errors followed predictable scaling across three model families

Source check: Source checked. Not independent scientific confirmation.

Finding: Researchers fit scaling laws to safety probes trained on smaller language models. The fits predicted the error of the next larger model to within 0.02 in 52 of 72 comparisons at the largest label budget. They tested these held-out predictions in the Qwen2.5, Qwen3.5 and Gemma 4 families.

Technology: Learned activation probes on Qwen2.5, Qwen3.5 and Gemma 4

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
Learned activation probes trained on up to a million labels per setting generate measurement curves; power-law fits on smaller models predict performance on a held-out larger model.
Test and result
The study trained two probe architectures across twenty models and compared held-out next-size forecasts with measured performance. Probe error was one minus AUROC.
Limitation / still unknown
The held-out extrapolation covers three model families and the study’s six safety tasks. The fitted constants and accuracy may differ with other data or probe designs.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.30678 · https://arxiv.org/html/2610.30678v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: arXiv v1 §§3–5.1, Figure 10 and E.6 held-out model-size comparisons, confirmed core claim in v2.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.