Honest LLaMA: New method could make ChatGPT more truthful
Researchers at Harvard University have developed a technique called Inference-Time Intervention (ITI) to improve the truthfulness of large language models.
ITI shifts model activations along specific network segments during text generation, increasing Alpaca's accuracy in the TruthfulQA benchmark from 32.5% to 65.1%.
The method is minimally invasive, requires little training data and computational power, and does not encourage misleading behavior as reinforcement learning with human feedback can.
https://the-decoder.com/honest-llama-new-method-could-make-chatgpt-more-truthful/
—
@aioftheday — news about artificial intelligence
Researchers at Harvard University have developed a technique called Inference-Time Intervention (ITI) to improve the truthfulness of large language models.
ITI shifts model activations along specific network segments during text generation, increasing Alpaca's accuracy in the TruthfulQA benchmark from 32.5% to 65.1%.
The method is minimally invasive, requires little training data and computational power, and does not encourage misleading behavior as reinforcement learning with human feedback can.
https://the-decoder.com/honest-llama-new-method-could-make-chatgpt-more-truthful/
—
@aioftheday — news about artificial intelligence