Transformers need glasses! Information over-squashing in language tasks
read the original 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.
This paper has not been read by Pith yet.
Forward citations
Cited by 2 Pith papers
-
From Dispersion to Attraction: Spectral Dynamics of Hallucination Across Whisper Model Scales
The Spectral Sensitivity Theorem identifies a phase transition in Whisper models where scaling causes self-attention to collapse into rank-1 attractors, decoupling output from acoustic evidence.
-
Adaptive Long-term Embedding with Denoising and Augmentation for Recommendation
ALDA4Rec improves sequential recommendation by denoising item-item graphs via community detection and adaptively fusing short-term GCN embeddings with long-term sequence models using GRUs, attention, and MLP weighting.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.