Towards Interpretability Without Sacrifice: Faithful Dense Layer Decomposition with Mixture of Decoders
This addresses the challenge of understanding and editing dense layers in large language models, offering a novel approach that is not incremental but provides a new paradigm for interpretability.
The paper tackles the problem of making dense layers in large language models interpretable without sacrificing accuracy by introducing Mixture of Decoders (MxDs), which expand dense layers into specialized sublayers and outperform state-of-the-art methods on the sparsity-accuracy frontier in models up to 3B parameters.
Multilayer perceptrons (MLPs) are an integral part of large language models, yet their dense representations render them difficult to understand, edit, and steer. Recent methods learn interpretable approximations via neuron-level sparsity, yet fail to faithfully reconstruct the original mapping--significantly increasing model's next-token cross-entropy loss. In this paper, we advocate for moving to layer-level sparsity to overcome the accuracy trade-off in sparse layer approximation. Under this paradigm, we introduce Mixture of Decoders (MxDs). MxDs generalize MLPs and Gated Linear Units, expanding pre-trained dense layers into tens of thousands of specialized sublayers. Through a flexible form of tensor factorization, each sparsely activating MxD sublayer implements a linear transformation with full-rank weights--preserving the original decoders' expressive capacity even under heavy sparsity. Experimentally, we show that MxDs significantly outperform state-of-the-art methods (e.g., Transcoders) on the sparsity-accuracy frontier in language models with up to 3B parameters. Further evaluations on sparse probing and feature steering demonstrate that MxDs learn similarly specialized features of natural language--opening up a promising new avenue for designing interpretable yet faithful decompositions. Our code is included at: https://github.com/james-oldfield/MxD/.