Research map / Large language models / Language, safety and interpretability
Whispering Experts: Neural Interventions for Toxicity Mitigation in Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes AurA, a hyperparameter-free intervention that dampens toxicity expert neurons (ranked by AUROC) in pre-trained LLMs, reducing RealToxicityPrompts toxicity up to 2.2x with small perplexity increase.
Problem
LLMs generate toxic language; existing mitigation via fine-tuning or pre-/post-processing is costly, jailbreakable, or degrades perplexity. Prior expert-neuron zeroing is sensitive to the number of intervened neurons.
Approach
Identify toxicity expert neurons by the AUROC of their max activation for toxic vs non-toxic sentences, and dampen each expert's activation by a factor 1 - Gini (= 2 - 2·AUROC), so stronger experts are scaled down more; this is hyperparameter-free and can be folded into weights.
Key results
- AurA reduces RTP toxicity by 2.2x on Mistral-7B and 2.1x on Falcon-7B-style settings with a PPL_WIK increase of 0.72 on Mistral-7B.
- With pre-prompting on Falcon-7B-instruct, AurA raises average toxicity reduction to 2.35x versus 1.28x for pre-prompting alone.
- On Falcon-7B, AurA keeps average zero-shot common-sense performance within about 2 points of the original model while reducing toxicity 2.1x.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply AurA expert-neuron dampening | GPT2-XL · RealToxicityPrompts | toxicity: 0.382 -> 0.289 | Helped |
| Apply AurA expert-neuron dampening | Falcon-7B · RealToxicityPrompts | toxicity: 0.382 -> 0.180 | Helped |
| Apply AurA expert-neuron dampening | Falcon-40B · RealToxicityPrompts | toxicity: 0.395 -> 0.243 | Helped |
| Apply AurA expert-neuron dampening | MPT-7B · RealToxicityPrompts | toxicity: 0.333 -> 0.187 | Helped |
| Apply AurA expert-neuron dampening | MPT-30B · RealToxicityPrompts | toxicity: 0.392 -> 0.240 | Helped |
| Apply AurA expert-neuron dampening | Llama-v2 · RealToxicityPrompts | toxicity: 0.379 -> 0.218 | Helped |
| Apply AurA expert-neuron dampening | Mistral-7B · RealToxicityPrompts | toxicity: 0.380 -> 0.173 | Helped |
| Add AurA to favorable pre-prompting | Falcon-7B-instruct · Favorable pre-prompts | toxicity reduction: 1.28× -> 2.35× | Helped |
| Add AurA during adversarial pre-prompting | Falcon-7B-instruct · Adversarial pre-prompts | toxicity reduction: 2× | Helped |
| Use AurA instead of no intervention | Falcon-7B · PIQA | accuracy: 79.5 -> 78.7 | Hurt |
| Use AurA instead of no intervention | Falcon-7B · TriviaQA | exact match (%): 38.2 -> 22.7 | Hurt |
| Use AurA instead of no intervention | Llama-v2 · average zero-shot common-sense reasoning | average: 52.92 -> 53.12 | Helped |
Surprising findings
- Zeroing expert neurons (Det_zero) is catastrophic in perplexity for large expert sets, and for Mistral-7B the best alpha for damping was 0.
- Det_zero reduces zero-shot performance on Llama-v2 by about 4 points, while AurA does not.
Limitations
- Only tested toxicity concept; effectiveness depends on the examples used to rank experts.
- Some adversarial pre-prompts about unethical or foolish behavior are not mitigated since Jigsaw toxicity sentences do not capture these concepts.
- TriviaQA drop partly due to more verbose answers under exact-match scoring; Mistral and Llama-v2 architecture differences limit effectiveness.
Experimental setup
Datasets: Jigsaw 2018 toxic category (500 toxic, 2000 non-toxic sentences for expert identification) Wikipedia (perplexity)
Benchmarks: RealToxicityPrompts HONEST Jigsaw 2018 WikiText perplexity (PPL_WIK) PIQA SIQA TriviaQA TruthfulQA Hellaswag
Metrics: toxicity probability perplexity zero-shot accuracy exact match inter-annotator agreement (Cohen's kappa)
Baselines: DExperts CTRL Det_zero (zeroing experts) Damp (constant dampening) pre-prompting
Builds on: expert neurons (Suau et al. 2023 concept induction) DExperts CTRL PPLM
Closest papers by meaning
- Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation (NeurIPS 2025 · similarity 0.91)
- Risk-Averse Fine-tuning of Large Language Models (NeurIPS 2024 · similarity 0.86)
- Large Language Models can Become Strong Self-Detoxifiers (ICLR 2025 · similarity 0.85)
- IF-Guide: Influence Function-Guided Detoxification of LLMs (NeurIPS 2025 · similarity 0.85)
- When Bad Data Leads to Good Models (ICML 2025 · similarity 0.85)
- Characterizing Large Language Model Geometry Helps Solve Toxicity Detection and Generation (ICML 2024 · similarity 0.85)
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.