Latent Action Reparameterization for Efficient Agent Inference

In three linesLAR (Latent Action Reparameterization) compresses LLM agent action spaces by learning semantic multi-step latent actions. This reduces effective decision horizon and inference costs while preserving expressiveness. Across benchmarks, LAR decreases action tokens and wall-clock inference time without degrading task success rates.

Summary generated by Claude — human-verified