Notes

Short expository notes and journal-club summaries.

What capable agents must know: selection theorems under uncertainty

A summary of Section 5 of Nayebi, What Capable Agents Must Know: Selection Theorems for Robust Decision-Making under Uncertainty (arXiv:2603.02491): under partial observability, low average regret on betting goals forces an agent’s memory to preserve the predictive distinctions the evaluation family makes, and threshold bets recover the predictive state and, under a linear-PSR assumption, its dynamics. PDF

General agents contain world models

A summary of Richens, Abel, Bellot, and Everitt (ICML 2025), the follow-up to Robust agents learn causal world models: any agent that generalises to multi-step goal-directed tasks must contain a predictive world model, extractable from its policy, with accuracy forced up by the depth of the goals it can achieve. Includes a step-by-step account of the goal construction behind Theorem 1 and the exact relation between goal depth and the number of encoded trials. PDF

Robust agents learn causal world models

A summary of Richens and Everitt (ICLR 2024): an agent that stays near-optimal across intervention-induced domain shifts must, in an information-theoretic sense, carry an approximate causal model of its environment. PDF

Preconditioned DeltaNet: a curvature and online-regression view

Notes on Tumma, Loo, and Rus, Preconditioned DeltaNet: curvature-aware sequence modeling for linear recurrences (arXiv:2604.21100). The delta rule is online first-order regression on the recurrent state, and preconditioning makes that regression curvature-aware. PDF