Post

What happens when we can't use recursive belief to compete? We can use anomaly detection instead! Here, we (led by soon-to-be-Dr @nitalon.bsky.social ) devise a multi-agent account where compression & reward expectation are used to notice deception jair.org/index.php/ja...

It augments agents with and ℵ-IPOMDP mechanism: 1. Anomaly Detection mechanisms: "My predictions are consistently wrong." 2. Out-of-Belief Policy: A severely 'defensive' response. In effect: 'I don't know _how_ you're tricking me, but I know you are, so stop or we'll both suffer'

With this, more equitable social outcomes are achieved, and exploitation is reduced when compression & expectation are tuned within normal bounds. It provides a plausible, mechanistic formal framework in which agents can compete without recursion.

The model offers another multi-agent route to paranoia alongside hypermentalising. If the anomaly detector is tuned too tightly, it yields false +ves. The agent becomes overly defensive against benign peers. This connects generalist accounts of delusions to social contexts.