Search papers, labs, and topics across Lattice.
This paper investigates how LLMs perform relational binding in discourse, revealing a Cell-based Binding Representation (CBR): a low-dimensional subspace where cells correspond to entity-relation pairs. Using partial least squares regression on controlled multi-sentence data, the authors decode entity and relation indices from attribute-token activations, identifying a grid-like geometry within the CBR subspace. Activation patching confirms that manipulating this subspace causally affects relational predictions, demonstrating the importance of CBR for relational binding.
LLMs use a surprisingly structured "Cell-based Binding Representation" to track entities and relations in discourse, opening the door to targeted interventions and improved relational reasoning.
Understanding a discourse requires tracking entities and the relations that hold between them. While Large Language Models (LLMs) perform well on relational reasoning, the mechanism by which they bind entities, relations, and attributes remains unclear. We study discourse-level relational binding and show that LLMs encode it via a Cell-based Binding Representation (CBR): a low-dimensional linear subspace in which each ``cell''corresponds to an entity--relation index pair, and bound attributes are retrieved from the corresponding cell during inference. Using controlled multi-sentence data annotated with entity and relation indices, we identify the CBR subspace by decoding these indices from attribute-token activations with Partial Least Squares regression. Across domains and two model families, the indices are linearly decodable and form a grid-like geometry in the projected space. We further find that context-specific CBR representations are related by translation vectors in activation space, enabling cross-context transfer. Finally, activation patching shows that manipulating this subspace systematically changes relational predictions and that perturbing it disrupts performance, providing causal evidence that LLMs rely on CBR for relational binding.