PSI learns label-free causal carriers in LLMs, matching supervised alignment on MIB

TrackingActions · x · 2026-11-19

A new mech-interp preprint (arXiv:2611.21307, Muratore & Mathis) introduces PSI: a low-rank activation subspace learned inside a frozen LLM that carries its choice, so swapping it changes the model's answer as a declared "semantic program" specifies — no counterfactual labels needed. On the MIB benchmark it matches supervised alignment at equal data/budget, and on natural science questions without gold labels it tracks the model's own answers better than a gold-label-trained subspace.

Original post →

More from Research

Research channel →