TIME Training Method Solves Qwen's Overthinking Problem

TIME (short context-triggered thinking) is a training methodology that encourages models to perform short reasoning in the middle of a response as the context changes, instead of generating a huge reasoning block at the start of the response.

This focuses on solving the 'overthinking' problem seen in models like QwQ, where the model wastes resources by performing tens of thousands of tokens of reasoning even for simple questions.

Key technical details:

• Target models: Qwen3 (4B, 8B, 14B, 32B)
• Training method: QLoRA and a four-phase curriculum applied
• Evaluation tool: Uses the self-developed TIMEBench benchmark
• Release scope: The dataset, training notebooks, scripts, and eval harness are released on GitHub, making reproduction easy

This research has been accepted at ACL 2026, presenting a new approach to improving the reasoning efficiency of models.