Summary of Whispering Experts: Neural Interventions For Toxicity Mitigation in Language Models, by Xavier Suau et al.
Whispering Experts: Neural Interventions for Toxicity Mitigation in Language Models
by Xavier Suau, Pieter Delobelle, Katherine Metcalf, Armand Joulin, Nicholas Apostoloff, Luca Zappella, Pau Rodríguez
First submitted to arxiv on: 2 Jul 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed paper tackles the issue of Large Language Models (LLMs) generating toxic language by identifying neurons responsible for toxicity and reducing their activation levels proportionally to this power. The authors introduce AUROC adaptation (AurA), a hyperparameter-free intervention that can be applied to any pre-trained LLM, achieving up to 2.2x reduction in toxicity with minimal perplexity increase. AurA demonstrates effectiveness across models of different scales (1.5B-40B parameters) and preserves common-sense zero-shot abilities. Furthermore, it can counteract malicious adversarial prompts that elicit toxic content. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper tries to make large language models less likely to produce bad words by finding the “bad” parts inside the model and reducing their strength. The authors create a way to do this without needing any special settings, called AUROC adaptation (AurA). This helps reduce bad words by 2.2 times while only slightly affecting how well the model understands language overall. AurA works for different-sized models and keeps the model’s common sense abilities. |
Keywords
» Artificial intelligence » Hyperparameter » Perplexity » Zero shot
