Engineering Verifiable Modularity in Transformers via Per-Layer Supervision

TL;DR

This paper introduces architectural modifications and supervision techniques in transformers to reveal and control hidden modularity, enabling predictable and causal influence over specific behaviors.

Contribution

It presents a novel approach combining dual-stream processing, per-layer supervision, and gated attention to expose and leverage modularity in transformer models.

Findings

Ablation effects are significantly larger with per-layer supervision.

Models with per-layer supervision show four times greater control over targeted behaviors.

The approach reveals widespread modularity, enabling causal control over model components.

Abstract

Transformers resist surgical control. Ablating an attention head identified as critical for capitalization produces minimal behavioral change because distributed redundancy compensates for damage. This Hydra effect renders interpretability illusory: we may identify components through correlation, but cannot predict or control their causal role. We demonstrate that architectural interventions can expose hidden modularity. Our approach combines dual-stream processing separating token and contextual representations, per-layer supervision providing independent gradient signal at each depth, and gated attention regularizing toward discrete activation patterns. When trained with per-layer supervision, models produce ablation effects 5 to 23 times larger than architecturally identical controls trained with standard objectives. This enables 4 times greater control leverage on targeted behaviors: scaling identified attention heads produces smooth, predictable changes in model output. The key finding is architectural. Without per-layer supervision, ablation damage concentrates near zero with low variance (Winograd standard deviation 0.63%). With per-layer supervision, effects spread widely (standard deviation 6.32%), revealing which predictions depend on which circuits. The larger variance is not measurement noise but the signature of unmasked modularity. We validate our approach through three components: engineered features that capture computational dynamics rather than vocabulary structure (validated by near-zero correlation with raw activation clustering), an architecture providing positive control for modularity, and causal experiments demonstrating functional reorganization where different tasks route through different attention heads. This es tablishes a methodology for transforming interpretability from passive observation to active control.