The BaseResearchLessWrong

Quine can predict cross-propensity spillover. (imperfectly but significantly better than chance)

According to the excerpt, megasilverfist reports results of a challenge to predict how fine-tuning a model for one propensity changes others; an entry called Quine predicted spillover imperfectly but better than chance, with a rules ambiguity noted over blank cells.

Open the source ↗

Why it matters

The same thing, for different readers.

For researchers

A SPAR group's challenge tests whether people or pipelines can predict which behaviours change when a model is trained on another; the winner did better than chance, not well.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
A 2 x 24 x 24 experiment of propensity fine-tuning spillover: training for each of 24 propensities, in either direction, and measuring the change in each of 24 evaluated propensities; participants submitted predictions.
What it found
Per the title, the winning entry, Quine, predicted spillover better than chance but imperfectly.
What it does not show
The post notes an ambiguity in the scoring rules about blank cells; only groups that submitted before the data went public are covered.
Status
A report

The record

Questions
Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice?
People
megasilverfist (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of November 12, 2026, via lesswrong (tier 1: the source’s own channel or an official record)
Cite this
LessWrong, “Quine can predict cross-propensity spillover. (imperfectly but significantly better than chance)”, November 12, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-11-12-quine-predicts-propensity-spillover
Our notesLight lane: excerpt only; size of the effect is not in the excerpt. The group's automated pipeline uses a Claude model, mentioned only in passing.