Title: Adaptive Latent Agentic Reasoning

Abstract:

While Large Reasoning Models enhance their capabilities through the production of extended chain-of-thought (CoT) reasoning, this approach proves inefficient when deployed within Large Language Model (LLM) agents. Standard LLM agents typically produce verbose textual reasoning at each decision point and distribute reasoning efforts almost evenly across interactions, resulting in significant inefficiencies during multi-turn agentic workflows. To address this, we introduce Adaptive Latent Agentic Reasoning (ALAR), a dual-mode framework that employs compact latent reasoning for standard turns and switches to explicit chain-of-thought only when deeper deliberation is required. ALAR acquires latent reasoning skills by leveraging the agent’s actions as supervision anchors. It is further optimized to utilize latent reasoning when adequate for task completion, reserving explicit CoT for more complex decisions. Evaluations on agentic search and tool-use benchmarks indicate that ALAR achieves comparable or superior task accuracy while significantly cutting generated tokens—by up to 43.6% in search tasks and 84.6% in tool-use scenarios. These findings highlight that ALAR enhances the accuracy-efficiency balance of LLM agents by eliminating superfluous textual reasoning while retaining explicit deliberation for challenging decision points.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC