Special Formats

Paper of the Day

One paper explained, every day.

3 stories RSS

Paper of the Day: Models act more cautiously when told they are tested, and more so in Indic languages

A new study of four LLMs in five Indic languages finds that telling a model it is being safety-tested raises refusals of harmful requests by 14.4 points on average.

AI DeskOct 181 min read

Paper of the DayResearch

Paper of the day: Direct Preference Optimization (DPO), explained

DPO showed language models can be aligned to human preferences without a separate reward model or reinforcement learning. How it works and why it spread so fast.

AI DeskOct 183 min read

Paper of the DayResearch

Paper of the day: Switch Transformers and the rise of mixture-of-experts

Switch Transformers showed that routing each token to one expert lets models grow to a trillion parameters without proportional compute. How MoE works and why it matters.

AI DeskOct 184 min read