Claimempirical · verifiableA factual claim that could be checked directly against observation or primary records.constitution →importance · notableImportance 0.50, from 0 to 1 · notable: a contested point in a live debate (also the default before judging). Higher-importance claims are worth more to assess, so funding reaches them sooner.constitution →

TopicsMechanistic Interpretability Large language models AI Control AI Safety

On held-out jailbreaks and OOD datasets, linear probes outperform SAE probes for harmful-intent detection in Gemma 2 9B IT.

Not yet assessed, so this page carries no verdict. That is not a finding for or against the claim. Attention goes where its expected value is highest and someone funds it; nothing has funded an assessment of this claim yet, and anyone can.

Decomposition

This claim has not been assessed yet. Attention goes where its expected value is highest and someone funds it; once this claim's assessment is funded and runs, it may well decompose into subclaims.

or start a mandate for this whole area →

Provenance

Where this claim has been said, linked to its canonical form.

Neither source has been traced to an origin, or to one another.

DeepMind’s recent negative-results report [29] echoes both findings on a safety-relevant probing task.

Are Sparse Autoencoder Benchmarks Reliable?preprintextraction 0.65

Chanin, David · arXiv.org · May 18, 2026 · original

Related work; refers to Google DeepMind's 2025 report on negative results for SAEs on downstream tasks

Negative result: SAEs underperformed linear probes

Negative Results for SAEs On Downstream Tasks and Deprioritising SAE Research (GDM Mech Interp Team Progress Update #2)blog postextraction 0.85

lewis smith et al. · original

GDM mech interp team study probing Gemma 2 9B IT for harmful user intent, with held-out jailbreaks and datasets as the out-of-distribution test set.

Cite this claim: a formal citation with its evidence attached

Embed this claim: show it on another site with its current assessment

An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.

CardFits its column and sizes itself. Without the script it stays a quotation with its link.

<blockquote class="minerval-claim" data-claim-id="74170bea-5466-4689-8e10-99bd067d3631"><a href="https://minerval.ai/claims/74170bea-5466-4689-8e10-99bd067d3631">On held-out jailbreaks and OOD datasets, linear probes outperform SAE probes for harmful-intent detection in Gemma 2 9B IT.</a></blockquote>
<script async src="https://minerval.ai/embed.js"></script>

IframeWhere scripts are not allowed. A fixed height; add ?theme=light or ?theme=dark to fix its colours.

<iframe src="https://minerval.ai/embed/claims/74170bea-5466-4689-8e10-99bd067d3631" title="Minerval: On held-out jailbreaks and OOD datasets, linear probes outperform SAE probes for harmful-…" width="480" height="225" style="border:0;max-width:100%" loading="lazy"></iframe>

BadgeAn image for a README or a forum post, in Markdown.

[![Minerval's judgment of this claim](https://minerval.ai/embed/claims/74170bea-5466-4689-8e10-99bd067d3631/badge.svg)](https://minerval.ai/claims/74170bea-5466-4689-8e10-99bd067d3631)

The badge now:

Contribute

Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.

Stated views None yet. A view stated to the graph is shown on this page in its speaker’s words, as a source like any other, and carries no weight in the assessment.

Created by extractor · Nov 13, 2026. Every judgment on this page is accompanied by a reasoning trace.