Aha moment analysis

[yangluo23]

Great work, and thanks for sharing your exploration of the Aha moment!

But, may you provide more details on how you analyze this phenomenon? Specifically, what types of special tokens are considered indicative of an Aha moment? For instance, you mentioned "wait" in your report—have you also statistically analyzed other self-reflection patterns?

Additionally, one puzzling question is that in many open-source multimodal R1 projects, researchers often struggle to achieve this Aha moment. In fact, many small models tend to experience a rapid decline in completion length. However, your report perfectly reproduces the characteristics of DeepSeek-R1. In your view, if possible, what key differences in your training process set your work apart from other unsuccessful attempts—aside from avoiding RL on an SFT model?

[HengguangZhou]

Thank you for your interest. Yes, we did analyze other reasoning patterns, including self-reflection. We conducted a statistical analysis of keywords indicative of self-reflection and exploration of alternative approaches, including “aha”, “check”, “possibl”(possible and possibly), “verify”, “wait”, and “yet”. We visualized the distribution of keywords over time steps in the following plot:

I think it’s quite interesting to see a peak of these keywords between time step 800 and 1000.

Regarding the key differences, we believe that the difficulty of the training data for the model plays a crucial role. Specifically, if a question is too simple, the model may easily provide a short answer without exploring more complex reasoning trajectories. On the opposite, if a question is too difficult, the model would have no clue of where to start.

[yangluo23]

Thanks for youre reply!
Yes, I do think so. It seems that the growth characteristics of R1's completion length are highly related to the dataset and training method. I’ve found this in my recent experiments. At the same time, this might be an interesting point to further explore.

[lerogo]

@HengguangZhou @xirui-li @ruocwang

Thank you for your excellent work.

We replicated your method, and then we tried to search for the results of each step, but we still couldn't find examples of the "aha moment".
We searched using the words you mentioned, such as "aha", "check", etc., but these words are far from the "aha moment".

Could you please provide the training logs? And mark some files where the "aha moment" occurs? This should be important for the community.

Thank you again for your work and looking forward to your reply.

[lerogo]

We tried searching for the word "but wait", but only a pitiful few samples showed up with "aha moment."