Imbue released Catalyst to find testable explanations of how neural networks work

Imbue released Catalyst to find testable explanations of how neural networks work

On 20 July, Imbue released Catalyst, an open tool for computational research. In theory discovery mode, agents propose several explanations for a single observation, test them through computational experiments, and develop the candidates that perform best under testing.

In an Imbue demonstration, the weights of a multilayer neural network diverged during training from a single line into several rays. The authors studied angular clusters, in which groups of neurons moved in the same direction. The task was to produce an explanation that predicted the outcome of a new computational experiment.

Catalyst begins with several explanations. Each candidate includes a hypothesis, mathematical reasoning, and experimental results. Other agents receive the theory in an isolated context and look for logical errors, gaps, and counterexamples. Each theory must predict the outcome of the next experiment in advance.

Catalyst then compares these predictions with the results and assesses the logical validity and completeness of each explanation. Candidates with high scores are more likely to receive another round of refinement. Here, a computational experiment distinguishes between competing theories by testing what they predicted.

The Catalyst code is publicly available, so this search process can be reproduced for any problem in which the phenomenon can be described and measured with a script. In February, Imbue also released Darwinian Evolver for selecting code and prompts according to an objective function. Catalyst applies selection to explanations: each candidate undergoes criticism and an experiment designed to distinguish between theories.

The authors report that, after five rounds, runs using Gemini 3.5 Flash, Claude Opus 4.8, and Claude Sonnet 4.6 reached the correct clustering mechanism. The run using Claude Haiku 4.5 did not find an explanation within the same five rounds. The output is a sequence of theories and experiments that shows why the system selected each subsequent candidate.

Catalyst is designed for phenomena that can be reproduced and measured with a script. For such a problem, it builds a sequence consisting of an observation, several theories, an experiment, criticism, and the next version of the theory.