Dynamic Large Concept Models: Latent Reasoning in an Adaptive Semantic Space

Large language models apply uniform computation to all tokens, despite language exhibiting highly non-uniform information density. We propose Dynamic Large Concept Models (DLCM), which learn variable-length semantic concepts from latent representations and perform reasoning in a compressed concept space. By reallocating computation from redundant token processing to concept-level reasoning, DLCM enables adaptive compute allocation aligned with semantic structure. We further introduce a compression-aware scaling law and a decoupled µP parametrization for heterogeneous token- and concept-level modules. With approximately 34\% lower inference FLOPs, DLCM achieves a 2.69\% average improvement across 12 zero-shot benchmarks, with gains concentrated on reasoning-intensive tasks.