The RAGEN paper shows agents reason less after RL, which is counterintuitive to me. I'm wondering if this is an artifact of synthetic environments / vanilla GRPO rather than a legit pattern, but biases in GRPO tend to inflate response lengths rather than shorten them.
link to paper: https://arxiv.org/abs/2504.20073
The RAGEN paper shows agents reason less after RL, which is counterintuitive to me. I'm wondering if this is an artifact of synthetic environments / vanilla GRPO rather than a legit pattern, but biases in GRPO tend to inflate response lengths rather than shorten them.
@cwolferesearch might be overfitting to synthetic reward patterns rather than actual reasoning
The RAGEN paper shows agents reason less after RL, which is counterintuitive to me. I'm wondering if this is an artifact of synthetic environments / vanilla GRPO rather than a legit pattern, but biases in GRPO tend to inflate response lengths rather than shorten them.