Think Smarter, Focus Wisely: Adaptive Cognitive Allocation for LLM Reasoning

Haosheng Chen, Yun Hua, Xiangfeng Xu, Wenhao Li, Xiangfeng Wang

Keywords: Large language model

Abstract:

Encouraging longer chains of thought is a common design practice for improving LLM reasoning. However, recent studies show that “thinking more” can backfire.
In response, prior studies have typically employed augmented reasoning strategies to enhance performance. While these approaches often improve reasoning robustness and yield higher accuracy, they may also generate excessively long chains, which introduce redundant checks, demand disproportionate reasoning effort, and ultimately lead to inefficient consumption of cognitive resources.
This paper introduces {\bf{A}}daptive {\bf{R}}easoning via {\bf{C}}ognitive {\bf{A}}llocation ({\bf{ARCA}}), a structured reasoning framework that adaptively allocates cognitive resources across reasoning phases based on their reasoning state, thereby mitigating the efficiency–accuracy trade-off.
The core idea of ARCA is to structure the reasoning procedure into classified phases, while grounding the process and suppressing incoherent drift.
Within each phase, ARCA generates candidate directions and employs a Borda-Aggregated selector to identify the most promising ones, while steering inference along phase-aware directions and pruning redundant exploration.
Through the dynamic allocation of cognitive resources, the proposed ARCA framework can achieve a balance between accuracy and efficiency.
Across six reasoning benchmarks, ARCA consistently outperforms strong baselines, either in terms of enhanced accuracy or reduced reasoning cost.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 528