mlsanityData to 14 Oct 2026

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

arXiv 2505.14681 ↗First posted 20 May 2025cs.AI cs.CL cs.CVcited by 16

On the map: LLM reasoning → Test-time compute and efficient thinking

also in 1 more
Abstract (arXiv)

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.

Key facts. Cited by 16 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 38 of its 59 references that are in the map. First posted on arXiv 20 May 2025.

Builds on

38 of 59 references.

Closest ancestors · 21

Same direction, newest first

  1. 2025A Survey of Slow Thinking-based Reasoning LLMs using Reinforced Learning and Inference-time Scaling Law5 May
  2. 2025Efficient Reasoning Models: A Survey15 Apr
  3. 2025Reasoning Models Can Be Effective Without Thinking14 Apr
  4. 2025SEAL: Steerable Reasoning Calibration of Large Language Models for Free7 Apr
  5. 2025Retro-Search: Exploring Untaken Paths for Deeper and Efficient Reasoning6 Apr
  6. 2025Harnessing the Reasoning Economy: A Survey of Efficient Reasoning for Large Language Models31 Mar
  7. 2025Effectively Controlling Reasoning Models through Thinking Intervention31 Mar
  8. 2025ThinkEdit: Interpretable Weight Editing to Mitigate Overly Short Thinking in Reasoning Models27 Mar
  9. 2025A Survey of Efficient Reasoning for Large Reasoning Models: Language, Multimodality, and Beyond27 Mar
  10. 2025Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models20 Mar
  11. 2025L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning6 Mar
  12. 2025Dynamic Parallel Tree Search for Efficient LLM Reasoning22 Feb
9 more
  1. 2025The relationship between reasoning and performance in large language models--o3 (mini) thinks harder, not longer21 Feb
  2. 2025LightThinker: Thinking Step-by-Step Compression21 Feb
  3. 2025Mixture of Tunable Experts -- Behavior Modification of DeepSeek-R1 at Inference Time16 Feb
  4. 2025Learning to Stop Overthinking at Test Time16 Feb
  5. 2025The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks12 Feb
  6. 2025s1: Simple test-time scaling31 Jan
  7. 2025Thoughts Are All Over the Place: On the Underthinking of o1-Like LLMs30 Jan
  8. 2024Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs30 Dec
  9. 2024Token-Budget-Aware LLM Reasoning24 Dec

Other references · 17

  1. 2017Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layercited by 3,509
  2. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,879
  3. 2025Kimi k1.5: Scaling Reinforcement Learning with LLMscited by 1,210
  4. 2024DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Modelscited by 867
  5. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 797
  6. 2024From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipelinecited by 500
  7. 2025LIMO: Less is More for Reasoningcited by 498
  8. 2025Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Modelscited by 384
  9. 2024OLMoE: Open Mixture-of-Experts Language Modelscited by 301
  10. 2024OpenMoE: An Early Effort on Open Mixture-of-Experts Language Modelscited by 187
  11. 2025Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketchingcited by 117
  12. 2025Trading Inference-Time Compute for Adversarial Robustnesscited by 61
5 more
  1. 2025Generative AI Act II: Test Time Scaling Drives Cognition Engineeringcited by 15
  2. 2025Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Controlcited by 14
  3. 2025Can Atomic Step Decomposition Enhance the Self-structured Reasoning of Multimodal Large Models?cited by 10
  4. 2025C3PO: Critical-Layer, Core-Expert, Collaborative Pathway Optimization for Test-Time Expert Re-Mixingcited by 7
  5. 2025Unveiling Hidden Collaboration within Mixture-of-Experts in Large Language Modelscited by 3

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 8; 2026: 8.

16 papers.

Sort:

By area:

  1. 2025Multilingual Routing in Mixture-of-Expertscited by 29
  2. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  3. 2025Steering MoE LLMs via Expert (De)Activationcited by 26
  4. 2025A Survey on Agentic Multimodal Large Language Modelscited by 23
  5. 2025SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Modelscited by 10
  6. 2026Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Expertscited by 9
  7. 2026From Data to Behavior: Predicting Unintended Model Behaviors Before Trainingcited by 8
  8. 2025A Theory of Inference Compute Scaling: Reasoning through Directed Stochastic Skill Searchcited by 5
  9. 2025Rewiring Experts on the Fly:Continuous Rerouting for Better Online Adaptation in Mixture-of-Expert modelscited by 3
  10. 2025ReflCtrl: Controlling LLM Reflection Efficiently via Representation Engineeringcited by 3
  11. 2025Ban&Pick: Enhancing Performance and Efficiency of MoE-LLMs via Smarter Routingcited by 2
  12. 2026Certain Head, Uncertain Tail: Expert-Sample for Test-Time Scaling in Fine-Grained MoEcited by 2
4 more
  1. 2026Routing Sensitivity Without Controllability: A Diagnostic Study of Fairness in MoE Language Modelscited by 1
  2. 2026Do Domain-specific Experts exist in MoE-based LLMs?cited by 0
  3. 2026xMIx: High-Performance Serving-Time Platform for Mechanistic Interpretability Appscited by 0
  4. 2026From Routing Signals to Selective Review: Visual regrounding in MoE VLMscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025Understanding and Steering the Cognitive Behaviors of Reasoning Models at Test-Time
  2. 2025Skill-Based Mixture-of-Experts: Adaptive Routing for Heterogeneous Reasoning via Inferred Skills
  3. 2025Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities
  4. 2025Dynamic Experts Search: Enhancing Reasoning in Mixture-of-Experts LLMs at Test Time
  5. 2025Think How to Think: Mitigating Overthinking with Autonomous Difficulty Cognition in Large Reasoning Models
  6. 2025Enhancing Reasoning Abilities of Small LLMs with Cognitive Alignment
  7. 2025Controlling Thinking Speed in Reasoning Models
  8. 2026Cognitive Demand Steering for Adaptive Meta-Reasoning in Large Language Models