Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Training

TL;DR

This work addresses cognitive inefficiencies in Mixture-of-Experts–based large reasoning models by identifying a subset of experts dubbed cognitive experts through normalized Pointwise Mutual Information (nPMI) and applying inference-time reinforcement to their routing. The approach, called Reinforcing Cognitive Experts (RICE), improves reasoning depth and accuracy on math and science benchmarks (AIME and GPQA Diamond) for both DeepSeek-R1 and Qwen3-235B, while preserving general instruction-following and modestly increasing verbosity. Cognitive experts generalize across domains and can yield targeted gains when domain-specific, though they may trade gains in some areas for others. Compared with prompting and decoding-constraint baselines, RICE provides the strongest average improvements, offering a lightweight, interpretable lever for enhancing cognitive control in large MoE reasoning systems with minimal training overhead.

Abstract

Mixture-of-Experts (MoE) architectures within Large Reasoning Models (LRMs) have achieved impressive reasoning capabilities by selectively activating experts to facilitate structured cognitive processes. Despite notable advances, existing reasoning models often suffer from cognitive inefficiencies like overthinking and underthinking. To address these limitations, we introduce a novel inference-time steering methodology called Reinforcing Cognitive Experts (RICE), designed to improve reasoning performance without additional training or complex heuristics. Leveraging normalized Pointwise Mutual Information (nPMI), we systematically identify specialized experts, termed ''cognitive experts'' that orchestrate meta-level reasoning operations characterized by tokens like ''<think>''. Empirical evaluations with leading MoE-based LRMs (DeepSeek-R1 and Qwen3-235B) on rigorous quantitative and scientific reasoning benchmarks demonstrate noticeable and consistent improvements in reasoning accuracy, cognitive efficiency, and cross-domain generalization. Crucially, our lightweight approach substantially outperforms prevalent reasoning-steering techniques, such as prompt design and decoding constraints, while preserving the model's general instruction-following skills. These results highlight reinforcing cognitive experts as a promising, practical, and interpretable direction to enhance cognitive efficiency within advanced reasoning models.

Figures (1)

• Figure 1: (a) Illustration of cognitive experts identified across domains. (b) Reinforcing only the top two experts (in red color) can improve reasoning accuracy without additional training.