← Toutes les offres
CW

Solomonoff induction with reflective oracles

Cole Wyeth

Global
Publié le 19 novembre 2026

Description du poste

Linkpost for https://arxiv.org/html/2611.14454v1 Our PIBBSS fellow Damini Kusum wrote an excellent paper on prediction with reflective oracle Solomonoff Induction (rOSI). Reflective oracles are best known for their game-theoretic applications, where they are often dismissed because "why would all players choose the same reflective oracle?" It turns out that reflective oracles can also be used to construct a version of Solomonoff induction (called rOSI) with very nice sequence prediction properties - roughly speaking, it can learn to extract computable structure from (partly) uncomputable sequences. In fact, its properties are quite reminiscent of logical induction, but this paper is much shorter and more accessible. In this context, there is only one "player" to choose an oracle, which is just a fixed-point (though you can sort of consider the set of hypotheses entertained by that player as subplayers who share the oracle). I think this paper is the clearest exposition of a very neat but previously obscure idea in agent foundations: self-reflection can overcome unrealizability.