![[2602.12662] Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.12662] Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents
Summary
This paper introduces CogRouter, a framework for large language models (LLMs) that enables dynamic adaptation of cognitive depth, enhancing decision-making efficiency in multi-turn tasks.
Why It Matters
As LLMs become integral in autonomous decision-making, optimizing their cognitive processes is crucial. CogRouter's approach addresses the inefficiencies of fixed cognitive patterns, potentially improving performance in complex tasks across various applications.
Key Takeaways
- CogRouter allows LLMs to adjust cognitive depth dynamically, enhancing task efficiency.
- The framework is based on ACT-R theory, incorporating four hierarchical cognitive levels.
- Two-stage training (CoSFT and CoPO) improves decision-making by focusing on confidence-aware actions.
- Experiments show CogRouter outperforms existing models like GPT-4o and OpenAI-o3 with fewer tokens.
- This research could influence future developments in AI decision-making frameworks.
Computer Science > Artificial Intelligence arXiv:2602.12662 (cs) [Submitted on 13 Feb 2026] Title:Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents Authors:Ruihan Yang, Fanghua Ye, Xiang We, Ruoqing Zhao, Kang Luo, Xinbo Xu, Bo Zhao, Ruotian Ma, Shanyi Wang, Zhaopeng Tu, Xiaolong Li, Deqing Yang, Linus View a PDF of the paper titled Think Fast and Slow: Step-Level Cognitive Depth Adaptation for LLM Agents, by Ruihan Yang and 12 other authors View PDF HTML (experimental) Abstract:Large language models (LLMs) are increasingly deployed as autonomous agents for multi-turn decision-making tasks. However, current agents typically rely on fixed cognitive patterns: non-thinking models generate immediate responses, while thinking models engage in deep reasoning uniformly. This rigidity is inefficient for long-horizon tasks, where cognitive demands vary significantly from step to step, with some requiring strategic planning and others only routine execution. In this paper, we introduce CogRouter, a framework that trains agents to dynamically adapt cognitive depth at each step. Grounded in ACT-R theory, we design four hierarchical cognitive levels ranging from instinctive responses to strategic planning. Our two-stage training approach includes Cognition-aware Supervised Fine-tuning (CoSFT) to instill stable level-specific patterns, and Cognition-aware Policy Optimization (CoPO) for step-level credit assignment via confidence-aware advantage reweighting. The k...
