A study examines the pitfalls of reading LLM judges' verdicts from their first generated token, showing that this approach overstates position bias and can mislead auditors. The researchers recommend reporting the rate at which a judge leads with a verdict token, which can be done at a lower computational cost.

Researchers compared correcting and deleting harmful training data to mitigate emergent misalignment in language models. They found that correcting the data is more effective than deleting it, with a third reduction in emergent misalignment and improved answers on held-out medical questions.

Researchers propose Alignment Forecasting, a method to predict alignment failures in language models before training. They introduce a benchmark, ALIGNMENTFORECASTBENCH, and a forecasting scaffold that combines an LLM's rating with a learned model's assessment to predict misbehavior. The approach shows promising results but more progress is needed before it can be used in practice.