Cell-Based Representation of Relational Binding in Language Models

1. Problem Statement

Understanding discourse requires tracking entities and the relations holding between them across multiple sentences. While Large Language Models (LLMs) perform well on relational reasoning tasks, the internal mechanism by which they "bind" entities, relations, and attributes remains unclear.

• Gap in Literature: Previous research on relational binding in LLMs has largely focused on short contexts involving single entity-attribute pairs or simple state tracking. These studies often fail to address discourse-level relational structures that span multiple sentences, involve multiple entities, and require maintaining complex, multi-dimensional bindings (i.e., which attribute belongs to which entity under which specific relation).
• Core Question: How do LLMs encode and retrieve relational bindings in a low-dimensional, interpretable subspace within their activation space, particularly when dealing with complex, multi-sentence narratives?

2. Methodology

The authors propose a framework called Cell-Based Binding Representation (CBR) and employ a combination of linear probing, causal intervention, and activation patching to validate it.

A. Data Construction: Indexed Relational Scheme (IRS)

To study discourse-level binding, the authors constructed controlled, multi-sentence datasets across five domains (countries, cities, jobs, objects, relations).

• Annotation: Instead of using natural language predicates (e.g., "manufactured in"), they introduced a delexicalized Indexed Relational Scheme (IRS).
• Structure: Every attribute token is annotated with a specific pair of indices: $(e_i, r_i)$, where $e_i$ is the entity index (order of introduction) and $r_i$ is the relation index (order of relation type introduction).
• Goal: This creates a ground truth where attributes are bound to specific $(e_i, r_i)$ "cells."

B. Identifying the CBR Subspace

• Linear Probing: The authors used Partial Least Squares (PLS) regression to map model activations (specifically from middle layers, e.g., Layer 15) to the ground-truth indices $(e_i, r_i)$.
• Hypothesis: If LLMs use a linear subspace for binding, a low-dimensional projection of activations should linearly decode these indices.
• Visualization: They projected attribute token activations onto the top two PLS components to visualize the geometry of the subspace.

C. Causal Interventions (Activation Patching)

To prove the subspace is causally responsible for binding (rather than just correlational), they performed interventions:

1. Logit Landscape Analysis: They sampled points in the CBR subspace and patched activations to these points. They observed that the model's predicted attribute changes systematically as the activation moves across "cells" in the grid-like subspace.
2. Perturbation: They added noise along the CBR directions versus random directions. Perturbing the CBR subspace significantly degraded prediction accuracy, while random perturbations did not.
3. Steering: They explicitly steered activations to shift indices (e.g., changing $r_i$ from 1 to 2). This successfully forced the model to predict attributes associated with the new relation index, confirming the mechanism's controllability.

D. Generalization and Stability Analysis

• Cross-Context Transfer: They tested if a projection matrix learned in one context could decode indices in another. They found that while the absolute position of the subspace varies by context, a translation vector ($\Delta$) can align activations across contexts, enabling cross-context transfer.
• Robustness: They tested stability under surface-level perturbations (shuffling sentence order, ablating attributes) and found the CBR geometry and index prediction remained stable.
• Semantic Sensitivity: They verified that the subspace captures semantic similarity; attributes bound to semantically similar relations cluster together in the CBR space.

3. Key Contributions

1. Discovery of CBR: The paper identifies a Cell-Based Binding Representation (CBR), a low-dimensional linear subspace where relational bindings are organized in a grid-like pattern. Each "cell" corresponds to a unique $(e_i, r_i)$ pair.
2. Delexicalized Indexing: The introduction of the Indexed Relational Scheme (IRS) allows for the systematic probing of how LLMs abstract discourse structure into discrete indices, independent of specific lexical content.
3. Causal Evidence: Through activation patching and steering, the authors provide causal proof that LLMs actively rely on this subspace to retrieve bound attributes. Manipulating the subspace directly alters model predictions.
4. Cross-Context Mechanism: The discovery that context-specific CBR representations are related by translation vectors explains how models generalize relational binding mechanisms across different narratives.
5. Generalizability: The phenomenon is shown to be robust across different model families (Llama3-8B-Instruct, Qwen3-8B), domains, and discourse patterns.

4. Key Results

• Linear Decodability: PLS models achieved near-perfect fits ($R^2 \approx 0.95$) in decoding $(e_i, r_i)$ indices from activations using only 2–5 components.
• Grid Geometry: Visualizations revealed a clear grid structure where one axis corresponds to entity indices and the other to relation indices.
• Causal Impact:
  • Perturbing the CBR subspace caused a sharp drop in attribute prediction accuracy (e.g., from ~0.9 to <0.2).
  • Steering the subspace to a new index $(e_i, r_i)$ successfully changed the model's output to the attribute associated with that new cell.
• Translation Vectors: Applying a learned translation vector allowed a projection matrix trained on one context (e.g., "countries") to accurately predict indices in another (e.g., "cities"), achieving $R^2 > 0.97$ after alignment.
• Real-World Applicability: The CBR signal was successfully identified in the Re-DocRED real-world relation extraction dataset, particularly for samples with high contextual similarity to the training data.

5. Significance

• Mechanistic Interpretability: This work moves beyond "black box" performance to reveal the specific geometric and causal mechanisms LLMs use for discourse understanding. It extends the Linear Representation Hypothesis to complex, multi-entity relational structures.
• Monitoring and Control: The CBR subspace provides a structured way to monitor internal world states and correct binding errors (e.g., in bias scenarios where a model might incorrectly bind "CEO" to "male" instead of the contextually correct "female").
• Downstream Applications: The findings suggest potential improvements for Relation Extraction (RE) and Information Extraction tasks by leveraging these internal representations for more stable and interpretable reasoning.
• Theoretical Insight: It challenges the view that binding is a one-dimensional or purely positional phenomenon, demonstrating that LLMs construct a compositional, grid-like internal state to manage discourse complexity.

In summary, the paper establishes that LLMs do not merely memorize associations but construct a structured, low-dimensional, and causally active "cellular" map to bind entities and relations, enabling them to reason over complex, multi-sentence discourses.