Summary of Large Language Models Can Be Strong Self-detoxifiers, by Ching-yun Ko et al.
Large Language Models can be Strong Self-Detoxifiers
by Ching-Yun Ko, Pin-Yu Chen, Payel Das, Youssef Mroueh, Soham Dan, Georgios Kollias, Subhajit Chaudhury, Tejaswini Pedapati, Luca Daniel
First submitted to arxiv on: 4 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
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 Reducing toxic output from large language models (LLMs) is crucial. Existing methods rely on external reward models or fine-tuning with self-generated data to influence the outcome. In this paper, we demonstrate that LLMs can self-detoxify without additional reward models or re-training. We propose Self-disciplined Autoregressive Sampling (SASA), a lightweight algorithm for toxicity reduction of LLMs. SASA leverages contextual representations from an LLM to learn linear subspaces characterizing toxic and non-toxic output in analytical forms. When auto-completing responses, SASA dynamically tracks the margin of the current output to steer generation away from toxic subspace by adjusting autoregressive sampling strategy. Evaluated on LLLs of different scale and nature with RealToxicityPrompts, BOLD, and AttaQ benchmarks, SASA enhances generated sentence quality relative to original models and attains comparable performance to state-of-the-art detoxification techniques, significantly reducing toxicity level. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about making sure that large language models don’t produce harmful or mean output. Right now, people use special training methods or add extra information to the model to try to make it better. But this paper shows that the model can actually do some of this work on its own without needing extra help. The authors came up with a new way called Self-disciplined Autoregressive Sampling (SASA) that helps the model create more helpful and less toxic responses. They tested it on different language models and found that it worked really well, making sure the output was safer for everyone. |
Keywords
» Artificial intelligence » Autoregressive » Fine tuning
