ConceptMoE: Adaptive Token-to-Concept Compression for Implicit Compute Allocation

TL;DR

ConceptMoE introduces a dynamic token merging approach that adaptively compresses sequences into concept representations, enabling more efficient and effective large language models with improved performance and speedups.

Contribution

It proposes a novel adaptive token-to-concept compression method for MoE models, improving efficiency and performance across language and vision tasks.

Findings

Outperforms standard MoE in multiple benchmarks

Reduces attention computation and KV cache usage significantly

Achieves notable speedups in prefill and decoding times

Abstract

Large language models allocate uniform computation across all tokens, ignoring that some sequences are trivially predictable while others require deep reasoning. We introduce ConceptMoE, which dynamically merges semantically similar tokens into concept representations, performing implicit token-level compute allocation. A learnable chunk module identifies optimal boundaries by measuring inter-token similarity, compressing sequences by a target ratio $R$ before they enter the compute-intensive concept model. Crucially, the MoE architecture enables controlled evaluation: we reallocate saved computation to match baseline activated FLOPs (excluding attention map computation) and total parameters, isolating genuine architectural benefits. Under these conditions, ConceptMoE consistently outperforms standard MoE across language and vision-language tasks, achieving +0.9 points on language pretraining, +2.3 points on long context understanding, and +0.6 points on multimodal benchmarks. When converting pretrained MoE during continual training with layer looping, gains reach +5.5 points, demonstrating practical applicability. Beyond performance, ConceptMoE reduces attention computation by up to $R^2\times$ and KV cache by $R\times$. At $R=2$, empirical measurements show prefill speedups reaching 175\% and decoding speedups up to 117\% on long sequences. The minimal architectural modifications enable straightforward integration into existing MoE, demonstrating that adaptive concept-level processing fundamentally improves both effectiveness and efficiency of large language models.