MAKING LANGUAGE MODELS MORE TRUTHFUL AT TEST TIME

Summary from the TLDR AI newsletter

Language models are well known for confidently confabulating (sometimes called hallucinating). It turns out that these models have an internal representation for truthfulness and by tweaking the model activations you can elicit a truthful response. The researchers found that by using this technique on the instruction-tuned Alpaca model, they improved TruthfulQA performance from 32.5% to 65.1%.