🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Timely work!

hmmm if you never push up, you maintain more entropy by not doing excessive sharpening. These guys might be onto something🧐

Oh man this is a gorgeous idea. Training *against* negative samples but not towards positive ones maintains entropy in the model, therefore increases pass@high k during RL.

We find that a large base LM can be boosted to match its RL-tuned version📈 without training—simply by transferring the logit difference between a small RL-tuned model and its base at the inference time! 🤯 Check out the 🧵👇

New paper by Andre He: Rewarding the Unlikely: Lifting GRPO Beyond Distribution Sharpening arxiv.org/abs/2506.02355 Tired of sharpening the distribution? Try unlikeliness reward to learn new things from the roads less traveled

Only ding a model for making mistakes! It gives better results in RL and avoids mode collapse. We still understand so little about RL! But we’re learning. Your science dollars at work.

🚀 Check out our new #ICML2025 paper led by Zhepei Wei! Achieve 1.73× faster LLM decoding — no draft model needed, and no discrepancy from vanilla decoding!

🚀 Interesting work by Taiqiang Wu! 💡 Quick takeaway: If you collect additional instruction-following data for SFT, it's better to fine-tune the Base model and then graft the weights onto its corresponding Instruct model — rather than continuing to train the Instruct model!