The BasePeople

Abhishek Mishra

Titles as the sources gave them. In 1 entry, with 2 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“a TF-IDF classifier ranks harm better than the original Latent Guard in every model we test, and linear and MLP probes do about as well as sentence-embedding classifiers trained on the same examples.”
Faithful Latent Harm Monitors Work No Better Than the Models Own Behaviour at a Matched False Alarm RatearXiv, October 22, 2026. Speaking as: Other. From the abstract, on results for activations read at the end of the user instruction. The words are the authors' joint finding.
“The representation just before generation detects more, catching 44 percent of harmful prompts at a matched false-alarm rate against 32 percent for the refusal of the model, but that gain follows the refusal decision itself”
Faithful Latent Harm Monitors Work No Better Than the Models Own Behaviour at a Matched False Alarm RatearXiv, October 22, 2026. Speaking as: Other. From the abstract, on activations read just before the model's reply. The words are the authors' joint finding.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    Faithful Latent Harm Monitors Work No Better Than the Models Own Behaviour at a Matched False Alarm Rate

    A preprint by Abhishek Mishra, Sagnik Chatterjee and Suyash Maniyar tests latent safety monitors, which read a model's internal activations to flag harmful prompts, on reworded and jailbreak-style requests across 15 models. The authors report that simple text classifiers match or beat them and that gains track the model's own refusals.