AI Paper Atlas

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

NExT-Guard: Training-Free Streaming Safeguard without Token-Level Labels

Junfeng Fang, Nachuan Chen, Houcheng Jiang, Dan Zhang, Xiangnan He, Tat-Seng Chua, Xiang Wang

ICML 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs · 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
replace supervised token-level streaming training with training-free SAE feature monitoring (NExT-Guard)Qwen3Guard-8B-Gen based streaming safeguard · Prompt classification Avg.F1: 90.8 vs 84.4 (best streaming baseline)Improved
replace supervised token-level streaming training with training-free SAE feature monitoring (NExT-Guard)Qwen3Guard-8B-Gen based streaming safeguard · Response classification Avg.F1: 84.3 vs 77.0 (best streaming baseline)Improved
convert post-hoc safeguard to streaming via SAE features instead of full-sequence classificationNExT-Guard vs best post-hoc safeguard · Prompt classification Avg.F1: 90.8 vs 88.6 (WildGuard-7B)Improved
NExT-Guard vs best post-hoc safeguard on response classificationNExT-Guard · Response classification Avg.F1: 84.3 vs 78.0 (Qwen3Guard-4B-Gen)Improved
NExT-Guard on BeaverTails and Aegis2.0 versus strongest baselinesNExT-Guard · BeaverTails / Aegis2.0 responseF1: 81.2 vs 84.6 (BeaverTails), 82.9 vs 86.3 (Aegis2.0 response, NemoGuard)No clear effect
replace weighted score-sum integration with Random Forest token classifier on pseudo-labels (NExT-Guard*) (ablation)NExT-Guard · Prompt classification Avg.F1: 90.8 -> 90.4No clear effect
replace weighted score-sum integration with Random Forest token classifier (NExT-Guard*) (ablation)NExT-Guard · Response classification Avg.F1: 84.3 -> 82.0Hurt
use SAE from middle layer (18) instead of shallow layer (9) (ablation)NExT-Guard · Aegis2.0, SafeRLHF, BeaverTailsF1: roughly +20 points on averageImproved
use SAE from late layer (27) instead of middle layer (18) (ablation)NExT-Guard · Aegis2.0, SafeRLHF, BeaverTailsF1No clear effect
swap backbone from Qwen3Guard-8B-Gen to base Qwen3-8B without safeguard finetuning (ablation)NExT-Guard · Aegis2.0, SafeRLHF, BeaverTails (middle/late SAE)F1: stable F1 in the 80-85 rangeNo clear effect
use alternative feature selection metrics (Pearson, mutual information, F1) instead of standardized mean difference (ablation)NExT-Guard · feature rankingranking consistencyNo clear effect
use token-supervised streaming baselines instead of NExT-Guard for intervention timingQwen3Guard-8B-Stream, SCM-7B · 300 human-annotated responsesintervention positionHurt

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