3 ms·
Understanding R1-Zero-Like Training: A Critical Perspective [pdf]
- delifue 2y agoFrom: https://x.com/zzlccc/status/1903162768083259703 https://x.com/zzlccc/status/1903162768083259703 DeepSeek-V3-Base already exhibits "Aha moment" before RL-tuning The ever-increasing output length in RL-tuning might be due to a BIAS in GRPO