junshim/When2Think-1.5B

When2Think-1.5B: Adaptive Hybrid Reasoning

When2Think-1.5B is a 1.5 billion parameter model post-trained from deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B. Its core innovation lies in its ability to adaptively determine whether to reason explicitly (THINK/NOTHINK) and how much reasoning to allocate to a given problem. This allows it to provide direct answers for simpler instances while engaging in multi-step reasoning for more complex ones, optimizing the balance between accuracy and computational efficiency.

Key Capabilities

• Adaptive Think/NoThink Behavior: Learns to dynamically switch between direct answering and explicit multi-step reasoning based on problem difficulty.
• Accuracy-Efficiency Trade-off: Reduces unnecessary computation without compromising performance on challenging problems, leading to significant token savings on easier tasks.
• Standalone Deployment: Functions as a standard causal language model, requiring no external components like routers or verifiers for inference.
• Enhanced Mathematical Reasoning: Demonstrates improved Pass@k scores on benchmarks like MATH-500, GSM-Plus, AIME24, and AIME25, often with reduced token usage.

Good for

• Mathematical Problem Solving: Excels in tasks requiring verifiable reasoning, such as those found in mathematical benchmarks.
• Efficient Reasoning Research: Ideal for studying adaptive reasoning policies and the interplay between direct answers and explicit reasoning.
• Adaptive AI Systems: Suitable for applications where dynamic resource allocation for reasoning is crucial.
• Direct-Answer and Explicit-Reasoning Generation: Provides a single checkpoint capable of both modes of response, adapting to input complexity.

For more details, refer to the research paper: When2Think: Learning When and How Much to Reason.