The authors constructed Less is Safer (LIS), a dataset of 2,400 examples, using signals from inside a language model to guide selection.
Language models can better balance safety and usefulness with less data
Li and colleagues report that 2,400 carefully constructed examples outperform much larger safety datasets on the overall balance between blocking harmful requests and answering harmless ones.
Chinese Tech
Hongyu Li · Zhan Yu · Zhe Zhao · Kimmo Yan · Xiong Kuang · Renfen Hu
Beijing Normal University · Tencent
Research Digest··2 min read
Before you read this
Three ideas this paper builds onsafety alignmentProcess of training models to recognize and refuse harmful requests, ensuring outputs remain safe.108 papers →sparse autoencoderA neural network that learns a compact, interpretable representation of data by activating only a small subset of its neurons at once.9 papers →synthetic data generationCreating training examples artificially, often using an LLM, to overcome real-data scarcity or privacy constraints.84 papers →
Li and colleagues built a compact training dataset that combines examples for handling harmful requests with harmless prompts that provoke unnecessary refusals.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.