reasoning length = thinking length + solution length?

[Jarvis33yu]

Thank you for your great work on the L1 paper. I have one question I'd like to clarify. In the paper, it’s described that LCPO allows control over the reasoning chain length. However, when I look into the code, it seems that the length being controlled is actually the total output length — the thinking length + solution length.

Could you please clarify whether LCPO is controlling just the reasoning chain length, or the entire output length (reasoning + solution)?

[lvtao65535]

Hello, I have the same question. I noticed that in main_ppo.py, the process_item method of the RewardManager class defines the processing of responses, where valid_response_length seems to be the length of the complete response rather than CoT, which will be used for reward calculation subsequently.