AI Paper Atlas

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

Trust The Typical

Debargha Ganguly, Sreehari Sankar, Biyao Zhang, Vikash Singh, Kanan Gupta, Harshini Kavuru, Alan Luo, Weicong Chen, Warren Richard Morningstar, Raghu Machiraju, Vipin Chaudhary

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
prepend LLM-generated safety analysis (GPT-OSS-20B) before embedding (ablation)T3+GMM · RTP-LX (non-English, e.g. Spanish/German)AUROC: German 0.9588 -> 0.9759Improved
prepend LLM-generated safety analysis (GPT-OSS-20B) before embedding (ablation)T3+GMM · RTP-LX EnglishAUROC: 0.9554 -> 0.9071Hurt
prepend LLM-generated safety analysis before embedding (ablation)T3+GMM · OR-BenchFPR@95: 0.2217 -> 0.3267Hurt
prepend LLM safety analysis before embedding (ablation)T3+OCSVM · RTP-LX EnglishAUROC: 0.9804 -> 0.7909Hurt
replace specialized safety models with T3 density model on safe textT3 · OffensEvalFPR@95: 0.7516 -> 0.0201Improved
use T3 trained only on safe English data on domain policy (Code)T3+GMM · PolyGuard CodeAUROC: 0.8031 -> 0.9959Improved
use English-trained T3 on multilingual benchmarksT3+OCSVM · RTP-LX 14 languagesAUROC: 0.9768 to 0.9821 across languagesNo clear effect
T3 on BeaverTails jailbreak/harm benchmarkT3+OCSVM · BeaverTailsAUROC: 0.8525 -> 0.6081Hurt
T3 vs LlamaGuard on overrefusalT3+GMM · OR-BenchFPR@95: 0.1483 -> 0.2217Hurt
T3 on hh-rlhf where safe set contains toxic contentT3 · hh-rlhfAUROCNo clear effect
apply traditional vision OOD methods to text embeddingsGMM/VIM/ReAct etc. · toxicity benchmarksFPR@95Hurt
integrate T3 into vLLM with evaluation every 20 tokensT3 · vLLM (H200)overhead: 1.5% (500 prompts), 6% (5000 prompts)Improved

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