[R] NL-ITI: modifying LLM internal representations to make it more truthful

Hi, here you can find our recent publication (along with code) in which we modify LLM internal representations to make it more truthful. In short, we optimized ITI method ( [2306.03341.pdf (arxiv.org)](https://arxiv.org/pdf/2306.03341.pdf) ) and achieved significant performance improvement. Evaluation was performed mostly on TruthfulQA, though we also tested generalization beyond it (MMLU, ARC, OpenBookQA). We used KL and CE metrics, to measure how invasive is intervention.

https://paperswithcode.com/paper/nl-iti-optimizing-probing-and-intervention

machinelearning