AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Do Activation Verbalization Methods Convey Privileged Information?

Millicent Li, Alberto Mario Ceballos Arroyo, Giordano Rogers, Naomi Saphra, Byron C Wallace

ICML 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
give verbalizer activations from target model instead of text onlyLIT (Llama3) · Feature extraction averageaccuracy: 0.64 vs 0.64 zero-shotNo clear effect
give verbalizer activations instead of text onlyPatchscopes (Llama3) · Feature extraction averageaccuracy: 0.38 vs 0.64 zero-shotHurt
give verbalizer activations instead of text onlyPatchscopes (Ministral) · Feature extraction averageaccuracy: 0.27 vs 0.68 zero-shotHurt
use different model families for target and verbalizerLIT · Feature extraction country-currencyaccuracy: 0.79 -> 0.00Hurt
replace activations with reconstructed input text via inversion (ablation)Llama3 verbalizer · Feature extraction, all activationsaccuracy: 0.79 vs 0.79 (country-currency)No clear effect
invert single token layer-15 activation instead of verbalizing it (ablation)Patchscopes (layer 15) · Feature extraction single tokenaccuracy: 0.16 -> 0.32 (athlete-sport)Improved
invert all activations instead of using LIT (ablation)LIT · Feature extraction athlete-sportaccuracy: 0.84 -> 0.39Hurt
shuffle persona attributes to remove sociodemographic correlation (ablation)LIT · PersonaQA countryaccuracy: 0.72 -> 0.03Hurt
use fantasy personas unknown to verbalizer (ablation)LIT and Patchscopes · PersonaQA-Fantasyaccuracy: 0.00 on all attributesHurt
finetune verbalizer M2 on the same persona knowledge as M1 (ablation)LIT · PersonaQA-Fantasy countryaccuracy: 0.00 -> 0.20Improved
verbalizer has knowledge the target lacks (ablation)LIT · PersonaQA-Fantasy fav foodaccuracy: 0.45 non-zero despite M1 lacking knowledgeHurt
compare linear probe to LIT on M1 activations (ablation)Logistic probe · PersonaQA-Fantasyaccuracy: 0.38 vs 0.25 (fav food)No clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy