• token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
• fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.

GSPO stabilizes long-horizon RL training for language models by optimizing at the sequence level rather than per-token, directly addressing the model-collapse and instability failures teams hit when scaling GRPO. If you're doing RL fine-tuning or training reasoning models, it's a concrete algorithmic fix worth understanding before your next long run.

• articleChinese CLIP: Contrastive Vision-Language Pretraining in Chinese
• articleOFASys: Enabling Multitask Learning with One Line of Code!
• articleQwen1.5-MoE: Matching 7B Model Performance with 1/3 Activated Parameters
• articleGeneralizing an LLM from 8k to 1M Context using Qwen-Agent

• clipScaling to long-horizon synthetic environments with GRPOAI Engineer
• articleGRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual SettingsKonstantin Dobler, Federico Scozzafava, Jonathan Janke, Mohamed Ali, Simon Lehnerer
• postWhen reward functions beat labeled data: reinforcement fine-tuning with GRPOAndrew Ng

Checking sign-in…

Loading comments…