causal
Amortized inference explains LLM Bayesian-like behavior
LLMs approximate Bayesian posteriors not by explicit factorized hypothesis updating but by amortizing inference through long learning, mapping novel contexts into Bayesian-like approximators in their activations after weights are frozen.
causalpending
Speaker
Chandra SripadaEvidence Quote
“It turns out that you can approximate the Bayesian posterior... not by explicitly going through a factorized representation, but in a sense by modeling it or learning it”
Created: 11/17/2026, 6:53:48 AM
Your notes
Loading notes...