Condensation theory refines objectivity with almost perfect latent models

Condensation theory, introduced by Sam Eisenstat in a 2025 paper and expanded in this 2026 survey, addresses a core problem in AI alignment: how to decompose a world model—represented as a random variable model—into conceptually meaningful latent parts that are nearly unique and objective. The post defines 'almost perfect condensation' as a set of four properties a latent variable model must satisfy: size (number of latents bounded), reconstruction (each latent predicts its cluster well), Markov (independence between clusters), and well-separatedness (clusters don't overlap much). These properties ensure that if two different condensations exist for the same underlying world model, they are essentially equivalent—there is a bijection between their latents that preserves the structure. This objectivity theorem is the key result: it guarantees that good condensations are not arbitrary but converge on the same conceptual decomposition.

Eisenstat distinguishes between random variable models (probabilistic) and Kolmogorov (algorithmic-information) condensation, which will be covered in a follow-up post. He also outlines open research directions: proving that interesting real-world random variable models actually possess such condensation properties, and extending the theory to string models. The work is part of the Agent Foundations research program and is supported by contributions from Kaarel Hänni, James Cook, and Jeremy Gillen. For AI alignment researchers, condensation offers a principled way to identify natural abstractions in neural network representations, potentially enabling safer, more interpretable AI systems by making 'concepts' mathematically precise and practically verifiable.