Jey Han Lau

LLMs struggle to reliably control the distribution of attributes like gender and race in multi-round generation, but a novel fine-tuning approach can precisely steer these distributions.

You can now unmask LLM ghostwriters with a lightweight fingerprinting method that works even when they try to hide in new domains or use unseen models.

Visual context barely affects human sentence acceptability judgments, but throws LLMs for a loop, widening the gap between their internal representations and acceptability predictions.

LLMs struggle with basic GPS coordinate reasoning, often failing at geometric computations despite showing some understanding of real-world geography.