Residual Connections and the Causal Shift: Uncovering a Structural Misalignment in Transformers
This addresses a subtle architectural issue in LLMs that could enhance training efficiency and model performance, though it is incremental as it builds on existing Transformer frameworks.
The paper identifies a structural misalignment in autoregressive Transformers where residual connections tie activations to the current token while supervision targets the next token, potentially propagating mismatched information. Experiments show that hidden representations switch from input to output alignment deep in the network, and proposed mitigation strategies like residual attenuation improve performance on benchmarks.
Large Language Models (LLMs) are trained with next-token prediction, implemented in autoregressive Transformers via causal masking for parallelism. This creates a subtle misalignment: residual connections tie activations to the current token, while supervision targets the next token, potentially propagating mismatched information if the current token is not the most informative for prediction. In this work, we empirically localize this input-output alignment shift in pretrained LLMs, using decoding trajectories over tied embedding spaces and similarity-based metrics. Our experiments reveal that the hidden token representations switch from input alignment to output alignment deep within the network. Motivated by this observation, we propose a lightweight residual-path mitigation based on residual attenuation, implemented either as a fixed-layer intervention or as a learnable gating mechanism. Experiments on multiple benchmarks show that these strategies alleviate the representation misalignment and yield improvements, providing an efficient and general architectural enhancement for autoregressive Transformers.