Adaptive Latent Agentic Reasoning

In three linesALAR (Adaptive Latent Agentic Reasoning) is a dual-mode framework alternating between compact latent reasoning and explicit chain-of-thought based on decision difficulty. Trained using agent actions as supervision, it reduces generated tokens by 43.6% in search and 84.6% in tool-use while maintaining accuracy.

Summary generated by Claude — human-verified