5 papers
Interpreting Reinforcement Learning Agents with Susceptibilities
Chris Elliott, Einar Urdshals, David Quarel +1
Susceptibilities are a technique for neural network interpretability that studies the response of posterior expectation values of observables to perturbations of the loss. We gener…
Stagewise Reinforcement Learning and the Geometry of the Regret Landscape
Chris Elliott, Einar Urdshals, David Quarel +2
Singular learning theory characterizes Bayesian learning as an evolving tradeoff between accuracy and complexity, with transitions between qualitatively different solutions as samp…
SCALAR: Benchmarking SAE Interaction Sparsity in Toy LLMs
Sean P. Fillingham, Andrew Gordon, Peter Lai +3
Mechanistic interpretability aims to decompose neural networks into interpretable features and map their connecting circuits. The standard approach trains sparse autoencoders (SAEs…
Universal Agent Mixtures and the Geometry of Intelligence
Samuel Allen Alexander, David Quarel, Len Du +1
Inspired by recent progress in multi-agent Reinforcement Learning (RL), in this work we examine the collective intelligent behaviour of theoretical universal agents by introducing…
On a numerical upper bound for the extended Goldbach conjecture
David Quarel
The Goldbach conjecture states that every even number can be decomposed as the sum of two primes. Let denote the number of such prime decompositions for an even . It is k…