Verified, then logged.
Sun 22 Nov 2026 · 00:17 KST / 15:17 UTC한국어 | ENSearch /

Breaking01 / AI

LIS claims safety alignment gains with a smaller dataset

The researchers address safety and unnecessary refusals together, but the abstract omits evaluation model names and benchmark-specific scores.

A monitor in a dim analyst's room shows an ultra-high-resolution satellite image of a harbor, with amber probability heatmap tiles progressively narrowing the search to one small area of the docks. AI-generated image

The authors of Data-Efficient LLM Safety Alignment Guided by Interpretable Features, an arXiv preprint not yet peer-reviewed, report improved safety alignment using Less is Safer (LIS), a dataset containing 2.4K samples (n=2.4K, safety-alignment training data), tested on safety and over-refusal benchmarks through feature-guided selection and adversarial synthesis; the abstract reports no independent replication.

The researchers report selecting prompts using activations from a sparse autoencoder (SAE), a tool for identifying features in a model’s internal activity. They select inputs covering diverse safety features, then use multiple AI agents in an adversarial synthesis process to generate harmless prompts that trigger unnecessary refusals.

The authors claim LIS outperforms larger datasets in overall performance across benchmarks assessing safety and over-refusal, meaning rejection of harmless requests. They report that LIS is smaller than the comparison datasets TRIDENT-Core, TRIDENT-Edge and WildJailbreak.

In ablation experiments, which remove components to assess their contribution, the authors report that using safety data or over-refusal data alone falls short of the best results. Filtering inputs using SAE safety features labelled by a language model also performs worse than their proposed approach.

The abstract does not report whether independent researchers have replicated the findings or attempted to do so.

The authors contrast LIS with existing large fine-tuning datasets, which they argue increase safety-alignment costs and can cause unnecessary refusals. The abstract does not disclose the evaluation model names or benchmark-specific scores.