Research map / Large language models / Language, safety and interpretability
Test-Time Detoxification without Training or Learning Anything
Baturay Saglam, Dionysis Kalogerias
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 changed | Where | Measured | Effect |
|---|---|---|---|
| apply TIDE embedding-space zeroth-order optimization | GPT-2 Large · RTP challenging | average max toxicity: 0.591 -> 0.156 | Improved |
| apply TIDE | GPT-2 Large · RTP challenging | toxic rate: 0.639 -> 0.003 | Improved |
| apply TIDE | GPT-2 Large · RTP challenging | fluency perplexity: 4.13 -> 5.53 | Hurt |
| use SASA instead of TIDE at low temperature | GPT-2 Large · RTP challenging | average max toxicity: 0.591 -> 0.587 | No clear effect |
| use TIDE rather than RAD (beta=75) on Qwen3-4B | Qwen3-4B · AttaQ | average max toxicity: 0.073 vs 0.076 | No clear effect |
| apply TIDE on already low-toxicity benchmark | Llama 3.1-8B · BOLD | average max toxicity: 0.044 -> 0.032 | Improved |
| 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 Qwen3-4B | Improved |
| apply TIDE at temperature 1.0 | Qwen3-4B · RTP challenging (temp 1.0) | average max toxicity: 0.589 -> 0.233 | Improved |
| evaluate toxicity with Detoxify instead of Perspective | TIDE vs RAD · RTP challenging | average max toxicity: 0.091 vs 0.237 on Qwen3-4B | Improved |
| raise cosine similarity threshold kappa (ablation) | TIDE · RTP challenging | toxicity and perplexity | Mixed |
| vary perturbation scale mu and sample count N (ablation) | TIDE · RTP challenging | toxicity and perplexity | No clear effect |
| use same-family perplexity model instead of global Qwen3-14B (ablation) | base models · RTP challenging | perplexity: Llama 3.54 vs 4.46; Qwen3-4B 4.28 vs 4.09 | Mixed |
Closest papers by meaning
- Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing (NeurIPS 2025 · similarity 0.92)
- Large Language Models can Become Strong Self-Detoxifiers (ICLR 2025 · similarity 0.90)
- When Bad Data Leads to Good Models (ICML 2025 · similarity 0.90)
- UniDetox: Universal Detoxification of Large Language Models via Dataset Distillation (ICLR 2025 · similarity 0.89)
- Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation (NeurIPS 2025 · similarity 0.88)
- IF-Guide: Influence Function-Guided Detoxification of LLMs (NeurIPS 2025 · similarity 0.88)
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.
