Engineering Verifiable Modularity in Transformers via Per-Layer Supervision

• URL: http://arxiv.org/abs/2603.18029v1
• Date: Sun, 08 Mar 2026 05:18:14 GMT
• Title: Engineering Verifiable Modularity in Transformers via Per-Layer Supervision
• Abstract summary: Ablating an attention head identified as critical for capitalization produces minimal behavioral change because distributed redundancy compensates for damage.<n>We demonstrate that architectural interventions can expose hidden modularity.<n>This es tablishes a methodology for transforming interpretability from passive observation to active control.
• Score: 0.0
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• 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.

Related papers

• Align, Unify, Suppress, Route: A Coherentist View of Transformer Computation [0.0]
  Coherentist Probabilistic Compositionalism grounds transformer computation in coherentist theories of interpretation.<n>Across 15 models from five architecture families, the suppression, unification, and routing weight-space signatures correlate with held-out activation-level role measures.<n>Explicit contradictions significantly shift a layerwise coherence proxy in 14 models; after removing shared residual covariance, the gap has the predicted direction in every model.
  arXiv  Detail & Related papers  (2026-08-22T16:29:20Z)
• Backdoor Directions in Vision Transformers [56.382912038371046]
  This paper investigates how Backdoor Attacks are represented within Vision Transformers (ViTs)<n>We identify a specific trigger direction'' in the model's activations that corresponds to the internal representation of the trigger.<n>Using this direction as a diagnostic tool, we trace how backdoor features are processed across layers.
  arXiv  Detail & Related papers  (2026-03-11T14:13:48Z)
• OrthoFormer: Instrumental Variable Estimation in Transformer Hidden States via Neural Control Functions [0.0]
  Transformer architectures excel at sequential modeling yet remain fundamentally limited by correlational learning.<n>We propose OrthoFormer, a causally grounded architecture that embeds instrumental variable estimation directly into Transformer blocks via neural control functions.
  arXiv  Detail & Related papers  (2026-03-08T03:05:16Z)
• Steering at the Source: Style Modulation Heads for Robust Persona Control [1.842570462253593]
  Activation steering offers a computationally efficient mechanism for controlling Large Language Models (LLMs) without fine-tuning.<n>In this work, we identify a sparse subset of attention heads that independently govern persona and style formation.<n>We demonstrate that intervention targeting only these specific heads achieves robust behavioral control while significantly mitigating the coherency degradation observed in residual stream steering.
  arXiv  Detail & Related papers  (2026-02-24T07:26:50Z)
• Same Answer, Different Representations: Hidden instability in VLMs [65.36933543377346]
  We introduce a representation-aware and frequency-aware evaluation framework that measures internal embedding drift, spectral sensitivity, and structural smoothness.<n>We apply this framework to modern Vision Language Models (VLMs) across the SEEDBench, MMMU, and POPE datasets.
  arXiv  Detail & Related papers  (2026-02-06T12:24:26Z)
• Disentangling Recall and Reasoning in Transformer Models through Layer-wise Attention and Activation Analysis [3.1526281887627587]
  Distinguishing recall from reasoning is crucial for predicting model generalization.<n>We use controlled datasets of synthetic linguistic puzzles to probe transformer models at the layer, head, and neuron level.<n>Our results provide the first causal evidence that recall and reasoning rely on separable but interacting circuits in transformer models.
  arXiv  Detail & Related papers  (2025-10-03T04:13:06Z)
• GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs [56.93583799109029]
  GrAInS is an inference-time steering approach that operates across both language-only and vision-language models and tasks.<n>During inference, GrAInS hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale.<n>It consistently outperforms both fine-tuning and existing steering baselines.
  arXiv  Detail & Related papers  (2025-07-24T02:34:13Z)
• Backdoor Cleaning without External Guidance in MLLM Fine-tuning [76.82121084745785]
  Believe Your Eyes (BYE) is a data filtering framework that leverages attention entropy patterns as self-supervised signals to identify and filter backdoor samples.<n>It achieves near-zero attack success rates while maintaining clean-task performance.
  arXiv  Detail & Related papers  (2025-05-22T17:11:58Z)
• Learning Action-based Representations Using Invariance [18.1941237781348]
  We introduce action-bisimulation encoding, which learns a multi-step controllability metric that discounts distant state features that are relevant for control. We demonstrate that action-bisimulation pretraining on reward-free, uniformly random data improves sample efficiency in several environments.
  arXiv  Detail & Related papers  (2024-03-25T02:17:54Z)
• Is Disentanglement enough? On Latent Representations for Controllable Music Generation [78.8942067357231]
  In the absence of a strong generative decoder, disentanglement does not necessarily imply controllability. The structure of the latent space with respect to the VAE-decoder plays an important role in boosting the ability of a generative model to manipulate different attributes.
  arXiv  Detail & Related papers  (2021-08-01T18:37:43Z)
• Unsupervised Controllable Generation with Self-Training [90.04287577605723]
  controllable generation with GANs remains a challenging research problem. We propose an unsupervised framework to learn a distribution of latent codes that control the generator through self-training. Our framework exhibits better disentanglement compared to other variants such as the variational autoencoder.
  arXiv  Detail & Related papers  (2020-07-17T21:50:35Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.