Inferred efficiency of static chains in LRMS
Recent LRMs achieve optimal performance by solving complex tasks using detailed COT inference. However, many of the simple tasks they handle can be solved by smaller models with fewer tokens, which makes unnecessary exhaustive reasoning. This echoes human thinking, where we use fast, intuitive responses to solve easy problems and for complex thinking slower, analytical thinking. Although LRMs mimic slow logical reasoning, they produce significantly longer outputs, increasing computational costs. The current approach to reducing inference steps lacks flexibility, limiting the model to a single fixed inference style. The need for adaptive reasoning is growing, and efforts can be adjusted according to task difficulty.
Limitations of existing training-based and non-training approaches
The latest research on improving the efficiency of LRMS reasoning can be classified into two main areas: training-based and training-free approaches. Training strategies often use reinforcement learning or fine-tuning to limit token use or adjust the depth of reasoning, but they tend to follow a fixed pattern without flexibility. No training methods utilize rapid engineering or pattern detection to shorten the output during inference; however, they also lack adaptability. The latest work focuses on variable-length inference, where the model adjusts the inference depth according to task complexity. Others study “overthinking”, in which case, unnecessarily over-modeling. However, few methods can be used to switch dynamically between fast and thorough reasoning – this article solves directly.
Introduction to Othink-R1: Dynamic Fast/Slow Inference Framework
Researchers at Kwajiang University and Oppo developed Othink-R1, a new approach that enables LRMS to switch between fast and slow thinking as cleverly as humans. By analyzing the inference patterns, they determine which steps are essential and which are redundant. With the help of another model as judges, they trained LRMS to adjust its inference style based on task complexity. Their method reduces unnecessary reasoning by more than 23% without losing accuracy. Using loss function and fine-tuning datasets, Othink-R1 outperforms previous models in various mathematical and questioning tasks.
System architecture: Inference pruning and double reference optimization
The Othink-R1 framework helps LRM switch dynamically between fast and slow thinking. First, it determines that LRMS includes unnecessary reasoning, such as over-interpretation or double checking, and detailed steps are indeed essential. Using this feature, it builds a select training dataset by pruning redundant inference and retaining valuable logic. Then, during fine-tuning, the special loss function balances the two ways of reasoning. This double reference loss compares the output of the model with fast and slow thinking variants, thereby encouraging flexibility. As a result, Othink-R1 can adaptively select the most efficient inference path for each problem while maintaining accuracy and logical depth.
Empirical evaluation and comparison of performance
The Othink-R1 model is tested on simpler quality checks and mathematical tasks to evaluate its ability to switch between fast and slow reasoning. The model demonstrates strong performance using datasets such as OpenBookQa, CommonsenseQA, Asdiv, and GSM8K, producing fewer tokens while maintaining or improving accuracy. Compared with baselines such as Nothinking and Dualformer, Othink-R1 exhibits a better balance between efficiency and effectiveness. Ablation studies confirm the importance of pruning, KL constraints and LLM judges in achieving the best outcome. One case study shows that unnecessary reasoning can lead to overthinking and reduced accuracy, highlighting the strength of Othink-R1 in adaptive reasoning.
Conclusion: Going towards a scalable and effective hybrid inference system
In short, Othink-R1 is a large inference model that adaptively switches between fast and slow thinking patterns for improved efficiency and performance. It solves the problem of unnecessary complex reasoning in large models by analyzing and classifying the inference steps as essential or redundant. By pruning redundancy while maintaining logical accuracy, Othink-R1 can reduce unnecessary calculations. It also introduces a double reference KL-Divergence loss to enhance hybrid inference. Tested by mathematical and quality check tasks, it reduces the redundancy of reasoning by 23% without sacrificing accuracy, indicating the hope of building more adaptable, scalable, and efficient AI inference systems in the future.
Check Paper and github pages. All credits for this study are to the researchers on the project. Also, please feel free to follow us twitter And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.
Sana Hassan, a consulting intern at Marktechpost and a dual-degree student at IIT Madras, is passionate about applying technology and AI to address real-world challenges. He is very interested in solving practical problems, and he brings a new perspective to the intersection of AI and real-life solutions.
