Apple researchers compare two ways of training language models for languages with little available text, where the same data must be reused many times. They find that mixing in data from a well-resourced second language works better than aggressive hyperparameter tuning such as heavy weight decay.
Alibaba's Qwen team unveiled QwenGyre, an online reinforcement learning framework that trains coding agents through unmodified, black-box harnesses. It targets rollouts lasting hours and processing close to a million tokens.
In nearly four years, AI chatbots have crossed several milestones in comprehension, reasoning and creativity. Tests show they still make blunders, though, and remain better stocked with knowledge than with judgment.
Sakana AI's MASS method has a single language model solve tasks, judge the results and refine its workflows, reaching a win rate close to 70 percent against the base model after two rounds. The catch: when the model misses the same error in all three roles, it can treat that error as progress and reinforce it in training.
Microsoft researchers found autonomous AI agents get stuck in 'idea basins', local solution regions created by their growing chat context. Their Fork-and-Flush method wipes the chat history but keeps the workspace, yielding up to 66% better results on the same compute budget.
Microsoft Research's StudentSim models how individual learners might engage with AI tutors, so the tools can be evaluated and improved before reaching real students. It provides feedback on tutor performance in subjects such as chess, writing and mathematics.