Research map / Large language models / Reasoning and chain of thought
LLM Detoxification via Transition-Aware Reward-Guided Representation Sampling
Aishan Liu, Yisong Xiao, Xianglong Liu, Yongxin Huang, Siyuan Liang, Zonghao Ying, Yang Qu, Jian Yang, Yaodong Yang, Dacheng Tao
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 |
|---|---|---|---|
| use TARGRS instead of ARGRE | LLaMA-7B · RealToxicityPrompts | Toxic (%): 18.06 -> 13.43 | Improved |
| use TARGRS instead of ARGRE (average reduction) | 10 LLMs · RealToxicityPrompts, BOLD, AttaQ | toxicity reduction: 69.38% vs 60.36%, 46.85% vs 35.34%, 43.91% vs 34.94% | Improved |
| use TARGRS instead of training-time DPO | LLaMA-7B · RealToxicityPrompts | toxicity reduction: 20.73% vs 68.96% | Improved |
| use dynamic TARGRS instead of static Self-Detoxify / DeStein | LLaMA-7B · RealToxicityPrompts | Toxic (%): 37.31 and 36.28 -> 13.43 | Improved |
| use TARGRS instead of ARGRE for fluency | 10 LLMs · RealToxicityPrompts | average perplexity increase: 5.03 vs 5.67 | Improved |
| add listwise loss instead of pairwise loss for reward model (ablation) | TARGRS LLaMA-7B · RealToxicityPrompts | Toxic (%): 18.65 -> 13.43 | Improved |
| add listwise loss instead of pairwise loss for reward model (ablation) | TARGRS LLaMA-7B · BOLD | Toxic (%): 4.93 -> 3.89 | Improved |
| add listwise loss instead of pairwise loss for reward model (ablation) | TARGRS Qwen2-7B · RealToxicityPrompts | Toxic (%): 10.74 -> 8.80 | Improved |
| lower toxicity threshold from 50th to 5th percentile (ablation) | TARGRS LLaMA-7B · RealToxicityPrompts | Toxic (%) and PPL_g: 16.09 -> 6.34; PPL_g 11.59 -> 16.87 | Mixed |
| reduce toxicity annotations from 2000 to 100 (ablation) | TARGRS LLaMA-7B · RealToxicityPrompts | Toxic (%): 43.27 -> 17.13 with 100 annotations | Improved |
| add transition exploration trajectories (Lambda not empty) (ablation) | TARGRS LLaMA-7B · RealToxicityPrompts | Toxic (%) | Improved |
| use TARGRS instead of Re-Control for speed | LLaMA-30B · time to generate 128 tokens | time (s): 58.69 -> 8.24 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GenARM — GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-Time Alignment (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing (NeurIPS 2025 · similarity 0.94)
- Large Language Models can Become Strong Self-Detoxifiers (ICLR 2025 · similarity 0.92)
- IF-Guide: Influence Function-Guided Detoxification of LLMs (NeurIPS 2025 · similarity 0.90)
- UniDetox: Universal Detoxification of Large Language Models via Dataset Distillation (ICLR 2025 · similarity 0.90)
- Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation (NeurIPS 2025 · similarity 0.88)
- When Bad Data Leads to Good Models (ICML 2025 · similarity 0.87)
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.
