The field of large language models is moving towards more efficient and effective reasoning capabilities. Recent research has focused on mitigating overthinking, a common issue where models generate excessively long and redundant reasoning chains. Various approaches have been proposed to address this issue, including dynamic compression, conditional token selection, and manifold steering. These methods aim to reduce computational overhead and improve the accuracy of large language models. Notably, some papers have introduced novel frameworks and techniques, such as Auto Long-Short Reasoning and State Machine Reasoning, which enable models to adaptively control their reasoning depth and generate more concise reasoning paths. Noteworthy papers include Amplify Adjacent Token Differences, which proposes a novel approach to mitigate Cyclical Reasoning, and TrimR, which introduces a verifier-based framework for dynamic CoT compression. ConciseRL is also notable for its conciseness-guided reinforcement learning framework, which guides models toward generating correct and concise reasoning traces.
Efficient Reasoning in Large Language Models
Sources
Amplify Adjacent Token Differences: Enhancing Long Chain-of-Thought Reasoning with Shift-FFN
TrimR: Verifier-based Training-Free Thinking Compression for Efficient Test-Time Scaling
ConciseRL: Conciseness-Guided Reinforcement Learning for Efficient Reasoning Models
Select2Reason: Efficient Instruction-Tuning Data Selection for Long-CoT Reasoning
Reasoning Meets Personalization: Unleashing the Potential of Large Reasoning Model for Personalized Generation
Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models
Fast Quiet-STaR: Thinking Without Thought Tokens
Don't Overthink it. Preferring Shorter Thinking Chains for Improved LLM Reasoning
Not All Tokens Are What You Need In Thinking
VeriThinker: Learning to Verify Makes Reasoning Model Efficient
Extended Inductive Reasoning for Personalized Preference Inference from Behavioral Signals
Don't Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models
CoThink: Token-Efficient Reasoning via Instruct Models Guiding Reasoning Models
THINK-Bench: Evaluating Thinking Efficiency and Chain-of-Thought Quality of Large Reasoning Models
Chain-of-Thought for Large Language Model-empowered Wireless Communications
Mitigating Overthinking in Large Reasoning Models via Manifold Steering
Scaling Reasoning without Attention
Emotion-o1: Adaptive Long Reasoning for Emotion Understanding in LLMs
AutoL2S: Auto Long-Short Reasoning for Efficient Large Language Models
Scalable Complexity Control Facilitates Reasoning Ability of LLMs
From Token to Action: State Machine Reasoning to Mitigate Overthinking in Information Retrieval
Revisiting Overthinking in Long Chain-of-Thought from the Perspective of Self-Doubt