Small transformers track Bayesian evidence for latent common causes via a context-invariant mechanism

2026-09-28

AI Summary

This study investigates whether small-scale Transformers can exhibit emergent Bayesian reasoning and achieve cross-context generalization. To this end, the project innovatively decouples the model's internal causal mechanisms from the true data-generating structure. Leveraging interpretable Transformer architectures and Bayesian inference theory, it focuses on context-invariant inference of latent common causes alongside natural language predictive modeling. The findings reveal that small Transformers can accumulate Bayesian evidence for latent common causes through context-invariant mechanisms. Furthermore, this work validates the internal representational mechanisms that support cross-context generalization to novel test cases. These insights offer a new perspective for understanding the emergence of Bayesian reasoning in small models.

Abstract

We present an in-depth investigation of how a form of Bayesian reasoning about common causes can emerge as a cross-contextual generalization in small, tractable transformers. Incrementing on recent work, our set-up (i) disentangles causal mechanisms in the model from the causal structure of the true data-generating process, (ii) orients more towards natural language prediction by considering inference of latent common causes, and (iii) considers whether and how Bayesian evidence accumulation for latent common causes can be implemented in representations and mechanisms that allow for cross-context generalization to novel test cases.

Problem

• Bayesian reasoning
• latent common causes
• cross-context generalization
• small transformers
• causal mechanisms

Innovation

• Small transformers
• Bayesian evidence accumulation
• Latent common causes
• Cross-context generalization
• Causal mechanism disentanglement