AI Paper Atlas

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

Beyond Linear Probes: Dynamic Safety Monitoring for Language Models

James Oldfield, Philip Torr, Ioannis Patras, Adel Bibi, Fazl Barez

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 6 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
replace linear probe with 5th-order truncated polynomial classifierlinear probe (gemma-3-27b-it) · WildGuardMix testF1: 88.03 -> 88.86Improved
replace linear probe with TPClinear probe (Llama-3.2-3B) · WildGuardMix testF1: 83.24 -> 84.48Improved
replace linear probe with TPClinear probe (Qwen3-30B) · WildGuardMix testF1: 85.53 -> 85.57No clear effect
TPC vs parameter-matched MLPMLP probe (gpt-oss-20b) · WildGuardMix testF1: 87.86 -> 88.05Improved
progressive degree-wise training instead of standard joint training (ablation)TPC · WildGuardMixF1 per truncationImproved
cascade with early exit using confidence threshold tauTPC · WildGuardMix (gemma-3-27b-it L40)F1 vs net parametersImproved
replace linear probe with TPC on shifted distributionlinear probe · ToxicChatF1: 46.27 -> 57.66Improved
replace linear probe with TPC on shifted distributionlinear probe · OpenAI-ModerationF1: 66.27 -> 69.02Improved
use activation probe instead of external LLM monitorgpt-4o-mini monitor · WildGuardMix testF1: 86.63 -> 88.50Improved

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