Multi-task Bayesian in-context learning

Let evidence set the prior.

Choose related task histories, make a prediction, then watch the same target evidence resolve to different posteriors. This is a transparent Bayesian reference lab for the paper's explicit-prior idea, not its trained transformer.

1. Load related task history

Choose an environment

2. Commit a prediction

Target evidenceHidden until reveal
Bayesian reference posterior-
Mastery0 / 2
A prior is an explicit starting belief assembled from related datasets. Pick a stream, then predict before revealing the target.

Related tasks -> target task -> predictive distribution

Why the result moves

The target is fixed. The starting belief is not.

Prior prefix

Select a related environment to expose the task history before the target examples.

Production rule

alpha = 1 + prior successes + target successes
beta = 1 + prior failures + target failures

Limit to inspect

Severe seasonal shifts can make related history a poor starting belief. More context is not automatically better.

Transfer and export

Can you predict the coastal posterior before the reveal?

Use Coastal with the same 5 / 10 target evidence and choose Low. The reference posterior should become 31.3%, demonstrating adaptation across environments. The downloaded evidence preserves both attempts and this limitation.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.