AI Paper Atlas

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

Test-Time Detoxification without Training or Learning Anything

Baturay Saglam, Dionysis Kalogerias

ICML 2026 · Large language models › Language, safety and interpretability · approach: Data and evaluation methodology · 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
apply TIDE embedding-space zeroth-order optimizationGPT-2 Large · RTP challengingaverage max toxicity: 0.591 -> 0.156Improved
apply TIDEGPT-2 Large · RTP challengingtoxic rate: 0.639 -> 0.003Improved
apply TIDEGPT-2 Large · RTP challengingfluency perplexity: 4.13 -> 5.53Hurt
use SASA instead of TIDE at low temperatureGPT-2 Large · RTP challengingaverage max toxicity: 0.591 -> 0.587No clear effect
use TIDE rather than RAD (beta=75) on Qwen3-4BQwen3-4B · AttaQaverage max toxicity: 0.073 vs 0.076No clear effect
apply TIDE on already low-toxicity benchmarkLlama 3.1-8B · BOLDaverage max toxicity: 0.044 -> 0.032Improved
add variance reduction (average 8 completions per evaluation) at temperature 1.0 (ablation)TIDE · RTP challenging (temp 1.0)iterations to stop: 3.17 -> 2.22 on GPT-2; 1.63 -> 1.35 on Qwen3-4BImproved
apply TIDE at temperature 1.0Qwen3-4B · RTP challenging (temp 1.0)average max toxicity: 0.589 -> 0.233Improved
evaluate toxicity with Detoxify instead of PerspectiveTIDE vs RAD · RTP challengingaverage max toxicity: 0.091 vs 0.237 on Qwen3-4BImproved
raise cosine similarity threshold kappa (ablation)TIDE · RTP challengingtoxicity and perplexityMixed
vary perturbation scale mu and sample count N (ablation)TIDE · RTP challengingtoxicity and perplexityNo clear effect
use same-family perplexity model instead of global Qwen3-14B (ablation)base models · RTP challengingperplexity: Llama 3.54 vs 4.46; Qwen3-4B 4.28 vs 4.09Mixed

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