Cutter Dawes

Cutter Dawes, Aryan Sharma, Angelos Ioannis Lagos et al.

Representing and navigating hierarchy is a fundamental primitive of reasoning. Large language models have demonstrated proficiency in a wide variety of tasks requiring hierarchical reasoning, but there exists limited analysis on how the models geometrically represent the necessary latent constructions for such thinking. To this end, we develop \textit{H-probes}, a collection of linear probes that extract hierarchical structure, specifically depth and pairwise distance, from latent representations. In synthetic tree traversal tasks, the H-probes robustly find the subspaces containing hierarchical structure necessary to complete the tasks; furthermore, in comprehensive ablation experiments, we show that these hierarchy-containing subspaces are low-dimensional, causally important for high task performance, and generalize within- and out-of-domain. Furthermore, we find analogous, though weaker, hierarchical structure in real-world hierarchical contexts such as mathematical reasoning traces. These results demonstrate that models represent hierarchy not only at the level of syntax and concepts, but at deeper levels of abstraction -- including the reasoning process itself.

Aryan Sharma, Cutter Dawes, Shivam Raval

When trained on tasks requiring an understanding of hierarchical structure, transformers have been found to represent this hierarchy in distinct ways: in the geometry of the residual stream, and in stack-like attention patterns maintaining a last-in, first-out ordering. However, it remains unclear whether these representations are causally used or merely decodable. We examine this gap in transformers trained on the Dyck language (a formal language of balanced bracket sequences), where the hierarchical ground truth is explicit. By probing and intervening on the residual stream and attention patterns, we find that depth, distance, and top-of-stack signals are all decodable, yet their causal roles diverge. Specifically, masking attention to the true top-of-stack position causes a sharp drop in long-distance accuracy, while ablating low-dimensional residual stream subspaces has comparatively little effect. These results, which extend to a templated natural language setting, suggest that even in a controlled setting where the relevant hierarchical variables are known, decodability alone does not imply causal use.

Cutter Dawes, Simon Segert, Kamesh Krishnamurthy et al.

A major challenge in the use of neural networks both for modeling human cognitive function and for artificial intelligence is the design of systems with the capacity to efficiently learn functions that support radical generalization. At the roots of this is the capacity to discover and implement symmetry functions. In this paper, we investigate a paradigmatic example of radical generalization through the use of symmetry: base addition. We present a group theoretic analysis of base addition, a fundamental and defining characteristic of which is the carry function -- the transfer of the remainder, when a sum exceeds the base modulus, to the next significant place. Our analysis exposes a range of alternative carry functions for a given base, and we introduce quantitative measures to characterize these. We then exploit differences in carry functions to probe the inductive biases of neural networks in symmetry learning, by training neural networks to carry out base addition using different carries, and comparing efficacy and rate of learning as a function of their structure. We find that even simple neural networks can achieve radical generalization with the right input format and carry function, and that learnability is closely correlated with carry function structure. We then discuss the relevance this has for cognitive science and machine learning.