Language models develop new social biases through adaptive exploration

A study with the telling title “Large language models develop novel social biases through adaptive exploration” has been submitted for review on the OpenReview platform and drew a notable response on Hacker News: the post garnered 106 votes and collected 53 comments.

The paper's central claim, reflected in its title, is that large language models develop new social biases through adaptive exploration. This is not the familiar charge leveled at AI — bias inherited directly from training data — but rather biases that emerge anew as the model masters the space of possible actions and adjusts its behavior.

To put it more simply, the authors point to a mechanism in which bias turns out to be not baggage brought in from a corpus of texts, but a byproduct of the very way the model searches for and locks in successful strategies. This shifts the perspective: what has to be countered is not only what the model has “read,” but also what it arrives at on its own.

Unfortunately, the full text of the article could not be obtained — instead, OpenReview serves up a browser verification page, so the details of the methodology, the specific experiments and the quantitative results remain off-screen. The argument can only be judged in full from the original publication.

Source: Hacker News