RT @hankhplee: 🚨Trapped by infinite scroll & autoplay on social media? Our TOCHI paper introduces a system that reduces distraction by 4x b…

RT @scyrusk: In this new pre-print, my student @YuxuanL_ outlines how language agents making shockingly biased decisions, even when their…

@sucholutsky @baixuechunzi Thanks for sharing this work, @sucholutsky! I especially love how the methods draw inspiration from psychological approaches to tackle similar challenges in humans. Excited to see researchers uncovering and addressing this hidden iceberg!

Thanks for your question, @lrzneedresearch! Yes, it could be related in terms of the "mismatched generalization" between question-answering and decision-making. Moreover, our findings suggest an even deeper issue—not just style-level generalization. Specifically, they reveal how aligning LLMs at the level of ‘words’ doesn’t necessarily translate into alignment at the level of ‘actions’.

Takeaway: The adage, “actions speak louder than words”, appears to hold for LLMs just as it does for people: LLMs have gotten better at hiding bias in what they “say,” but it remains visible in what their agents “do.” This calls for novel strategies to address not only explicit, but also implicit biases of LLMs. Huge thanks to my advisors Hiro( and Sauvik @scyrusk! 👉 Please boost this thread if you find it interesting! [n/n]