TrainingPrime Intellect targets the mismatch between agent rollouts and training
The October 19 release changes the default reinforcement-learning loss and unifies evaluation plumbing; its stress test is deliberately demanding.
Training
Pre-training and post-training advances that change the cost, speed and behavior of AI models.
Pre-training · Post-training
13 articles
Training
TrainingHugging Face’s agent curriculum separates imitation from learning through action
The October 20 guide organizes post-training methods around the evidence each method consumes.
TrainingAi2’s HiRO-ACEv2 sharpens climate simulation while exposing its training boundary
The system reports fine-scale extreme-event statistics against X-SHiELD; other climate regimes remain future work.
TrainingMicrosoft’s Agent Lightning trains through the harness that runs the agent
The v1.0 design reduces the gap between a training loop and deployed orchestration, but harness changes remain part of the experiment.
TrainingNVIDIA’s Nemotron recipe treats lost capabilities as a training problem
Nemotron 3.5 Super combines staged training and teacher consolidation; a reported long-context recovery illustrates why improving one skill can weaken another.
TrainingHugging Face puts the training experiment behind a conversation
ML Intern plans, runs and documents small adaptation jobs; its example costs describe compute charges, not the full cost of building reliable models.
Hover to animateTrainingNVIDIA finds that better speech labels can beat more noisy data
A small trusted-label refinement stage improved a recogniser already adapted on a much larger corpus. For robot voice interfaces, the order of the lessons matters.
Hover to animateTrainingSkild’s soccer robot turns its own improvement into a harder curriculum
The company’s September self-play result uses rival policies to make training harder without scripting every new manoeuvre.
Hover to animateTrainingFireworks makes the training-to-rollout handoff part of its product
The Training API manages the machinery between learning steps. Its most consequential promise is that the trainer and the model collecting experience compute the same thing.
Hover to animateTrainingMistral: As AI improves, its training tasks must get harder
Large 4’s reinforcement-learning account highlights the work of keeping tasks useful as the learner improves. More generated attempts can otherwise repeat mastered skills or stale experience.
Hover to animateTrainingNVIDIA: Cheaper training could pay for more experiments
Nemotron 3 Ultra uses four-bit training with higher-precision exceptions. The opportunity is more useful research per budget, provided stability and downstream quality survive.
Hover to animateTrainingFlashAttention-4: Faster chips still need better software
Expensive GPUs can still spend time waiting on slower operations. Research from Together AI and five other organizations shows how better scheduling can turn more of that hardware into useful work.
Hover to animateTrainingReViSQL: Better tests can teach AI what more examples cannot
ReViSQL targets SQL queries that return the expected answer for the wrong reason. Its lesson for specialist models is to improve the definition of success, not simply expand the dataset.
No articles in this topic yet
Explore the other topics in this desk.
Explore the research library for paper-based analysis.