A paper that should be generating more conversation dropped this week on arXiv. Pokharel and Dantu's research on hidden anchors in multi-agent LLM deliberation finds that when AI agents exchange and revise answers across multiple rounds, early outputs function as anchors that systematically bias the final consensus. The first agent to speak shapes what every subsequent agent says, even when those agents are nominally independent. The paper calls these hidden anchors. Behavioral economists have called the same phenomenon something simpler: whoever goes first, wins.
From Dice RPGs to AI Boardrooms
The Verge's review of Moves of the Diamond Hand, an RPG built around strange conversations and dice-based uncertainty, is an accidental gloss on the anchoring problem. The game's central mechanic is that you never quite know how a conversation will resolve, and the dice introduce genuine noise into social outcomes. What Pokharel and Dantu's paper demonstrates is that LLM deliberation systems have removed that noise, replacing it with the illusion of independence while the first output quietly colonizes the possibility space. Multi-agent AI is less a parliament than a game where one player rolls before the others can see the board.
Curriculum Design Has the Same Bug
A second arXiv paper this week on curriculum alignment in computer science education documents how CS2013 guidelines anchored CS2023 revisions in ways that constrained genuine updating. The same hidden anchor effect, institutionalized in a decade-long standard. Whether the system is a committee of AI agents or a committee of curriculum designers, the pattern holds: deliberation that starts from a shared document converges toward that document, regardless of what evidence accumulates in the interim. The Toy Story reference in David Pierce's Installer column this week noted that the franchise's best insight is that toys do not change their fundamental nature regardless of what world they're placed in. The anchor is always the first story someone told about what the thing is for.