Attention Saturation and Gradient Suppression at Inflection Layers: Diagnosing and Mitigating Bottlenecks in Transformer Adaptation

TL;DR

This paper investigates the causes of gradient suppression in transformer layers during fine-tuning, introduces diagnostic metrics to identify inflection layers, and proposes a targeted fine-tuning strategy using LoRA adapters to improve adaptation efficiency.

Contribution

It formalizes the mechanism of gradient suppression at inflection layers, introduces diagnostic metrics for layer identification, and proposes a diagnose-first, inject-light fine-tuning method with LoRA adapters.

Findings

Over-trained models benefit from inflection-layer LoRA injection.

Under-trained models experience performance degradation with inflection-layer injection.

Unblocking inflection layers enhances high-level and low-level feature adaptation.

Abstract

Pre-trained Transformers often exhibit over-confidence in source patterns and difficulty in forming new target-domain patterns during fine-tuning. We formalize the mechanism of output saturation leading to gradient suppression through standard cross-entropy and softmax analysis, showing that gradient suppression at inflection layers confines adaptation to high-level recombination of existing features while preventing low-level reconstruction. We introduce a set of layer-wise diagnostic metrics -- attention entropy (saturation proxy), activation gradient norm, parameter gradient norm, and Delta-CKA under a shared PCA basis -- to identify inflection layers characterized by both low attention entropy and steep gradient decay. Building on these findings, we propose a diagnose-first, inject-light fine-tuning strategy: selectively inserting LoRA adapters at inflection layers to restore suppressed backward signals with minimal parameter overhead. Experiments on BERT-base transfer from SST-2 to Rotten Tomatoes under under-trained and over-trained source regimes reveal that over-trained initialization benefits from inflection-layer LoRA injection, while under-trained initialization suffers performance degradation. When base features are strong, unblocking inflection layers facilitates high-level compositional adaptation; when base features are weak, full-pathway unblocking is required for low-level reconstruction, as supported by joint analysis of layer-wise activation gradients and Delta-CKA dynamics.