Emergent Ordinal Geometry in Transformers Trained on Local Comparisons

Technical Summary: Emergent Ordinal Geometry in Transformers Trained on Local Comparisons

Problem Statement
The paper investigates whether Transformer models, when trained exclusively on local (adjacent) comparisons within a hidden total order, acquire the same primitive mechanism used by biological systems for transitive inference (TI). TI is the ability to infer a non-adjacent relationship (e.g., $A < C$) from known adjacent relations ($A < B, B < C$). While biological cognition solves this via an "analog mental number line"—evidenced by the symbolic distance effect (where distant comparisons are processed faster/more accurately than adjacent ones)—Transformers are not explicitly pressured to represent order linearly. The study asks if this specific geometric representation and behavioral signature emerge spontaneously in neural networks trained on the classic Bryant and Trabasso [1971] paradigm.

Methodology
The authors constructed a synthetic task using a hidden total order of $N$ entities ($e_0 \prec e_1 \prec \dots \prec e_{N-1}$).

• Training Data: The model is trained only on adjacent pairs in both directions ($D_{train} = \{(e_i, e_{i+1}), (e_{i+1}, e_i)\}$).
• Evaluation Data: The model is evaluated on a disjoint set of distant pairs ($D_{OOD} = \{(e_i, e_j) : |i-j| > 1\}$), requiring the reconstruction of the global order from local data alone.
• Model Architecture: A single-layer Transformer with entity embeddings, one attention head (in the abstract) or two (in the results section), a 4d-wide GELU MLP, and a linear unembedding head. The relation is decoded from the final-position residual.
• Metrics:
  1. Geometric Structure: Principal Component Analysis (PCA) on entity embeddings to measure linearity (variance on the first principal component, $\rho_1$) and rank alignment (correlation between projected position and true rank).
  2. Behavioral Signatures: Decision confidence (logit margin) and geometric separation ($g_{ij}$) are measured against rank distance ($|i-j|$) to detect the symbolic distance effect.

Key Contributions

1. Synthetic Task Isolation: The paper introduces a controlled task that isolates ordinal generalization from local comparisons, strictly mirroring the biological TI paradigm.
2. Emergent 1D Manifold: It provides evidence that Transformers spontaneously organize entity embeddings into a one-dimensional manifold. The principal axis of this manifold recovers the hidden rank order with near-perfect fidelity.
3. Computational Analogue of the Distance Effect: The study demonstrates that even when accuracy reaches a ceiling, the model exhibits the symbolic distance effect: both decision confidence and geometric separation scale monotonically with the rank distance between items.

Results
Experiments were conducted on sets of $N=10, 15,$ and $20$ entities.

• Grokking-like Dynamics: The emergence of out-of-distribution (OOD) generalization coincides with a geometric reorganization of the embeddings. This transition is sensitive to optimization (specifically weight decay) and produces transient dynamics similar to "grokking."
• Geometric Recovery: As the model learns, embeddings collapse onto a line where the principal component aligns with the true rank order.
• Behavioral Mirroring: At peak OOD accuracy, the model's internal representations and output confidence strictly follow the symbolic distance effect. Distant pairs are not only correctly classified but are associated with higher confidence and greater geometric separation in the embedding space compared to adjacent pairs.

Significance
The paper claims to ground a 50-year-old behavioral regularity (the symbolic distance effect in TI) in the geometry of learned neural representations. By demonstrating that Transformers spontaneously develop a "mental number line" without explicit supervision, the work offers a mechanistic account of transitive inference that bridges cognitive science and modern neural network interpretability. It suggests that the ability to perform abstract relational reasoning may be a natural byproduct of optimizing for local constraints within a total order.