Transformers need glasses! Information over-squashing in language tasks

2024/06/06 by Federico Barbero, Barbero, Federico, Andrea Banino +13 · 17 citations
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Natural Language Processing Techniques #Speech and dialogue systems

paper · pdf · doi:10.48550/arxiv.2406.04267

openalex publication_date 2024/06/06 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

We study how information propagates in decoder-only Transformers, which are the architectural backbone of most existing frontier large language models (LLMs). We rely on a theoretical signal propagation analysis -- specifically, we analyse the representations of the last token in the final layer of the Transformer, as this is the representation used for next-token prediction. Our analysis reveals a representational collapse phenomenon: we prove that certain distinct sequences of inputs to the Transformer can yield arbitrarily close representations in the final token. This effect is exacerbated by the low-precision floating-point formats frequently used in modern LLMs. As a result, the model is provably unable to respond to these sequences in different ways -- leading to errors in, e.g., tasks involving counting or copying. Further, we show that decoder-only Transformer language models can lose sensitivity to specific tokens in the input, which relates to the well-known phenomenon of over-squashing in graph neural networks. We provide empirical evidence supporting our claims on contemporary LLMs. Our theory also points to simple solutions towards ameliorating these issues.

Cited by

• Understanding the Failure Modes of Transformers through the Lens of Graph Neural Networks
• Masks Can Be Distracting: On Context Comprehension in Diffusion Language Models
• Hierarchical Token Prepending: Enhancing Information Flow in Decoder-based LLM Embeddings
• Enhancing Graph Representations with Neighborhood-Contextualized Message-Passing
• Generalizable Insights for Graph Transformers in Theory and Practice
• Attention and Compression is all you need for Controllably Efficient Language Models
• Don't Blind Your VLA: Aligning Visual Representations for OOD Generalization
• Attention Sinks in Diffusion Language Models
• Stream: Scaling up Mechanistic Interpretability to Long Context in LLMs via Sparse Attention
• gLSTM: Mitigating Over-Squashing by Increasing Storage Capacity
• Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin
• Neural Message-Passing on Attention Graphs for Hallucination Detection
• Attention Consistency for LLMs Explanation
• Rethinking Layer-wise Model Merging through Chain of Merges
• Beyond Transcription: Mechanistic Interpretability in ASR
• Training Transformers for Mesh-Based Simulations
• DGP: A Dual-Granularity Prompting Framework for Fraud Detection with Graph-Enhanced LLMs

Related

• Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models
• Attention-likelihood relationship in transformers
• Attention Is Not All You Need: The Importance of Feedforward Networks in Transformer Models
• Equivalent Linear Mappings of Large Language Models
• Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space