Dynamic Latent Routing
For practitioners fine-tuning language models with limited data, DLR offers a method that outperforms standard supervised fine-tuning, while prior discrete-latent baselines underperform.
The paper introduces Dynamic Latent Routing (DLR), a post-training method for language models that jointly learns discrete latent codes, routing policies, and model parameters via dynamic search. In low-data fine-tuning, DLR matches or outperforms supervised fine-tuning across four datasets and six models, achieving a mean gain of +6.6 percentage points.
We investigate the temporal concatenation of sub-policies in Markov Decision Processes (MDP) with time-varying reward functions. We introduce General Dijkstra Search (GDS), and prove that globally optimal goal-reaching policies can be recovered through temporal composition of intermediate optimal sub-policies. Motivated by the "search, select, update" principle underlying GDS, we propose Dynamic Latent Routing (DLR), a language-model post-training method that jointly learns discrete latent codes, routing policies, and model parameters through dynamic search in a single training stage. In low-data fine-tuning settings, DLR matches or outperforms supervised fine-tuning across four datasets and six models, achieving a mean gain of +6.6 percentage points, while prior discrete-latent baselines consistently underperform SFT. Mechanistic analyses and targeted code ablations show that DLR learns structured routing behaviors with distinct causal roles.