training
Curriculum learning: why the order of the examples matters Lesson
Training a model on easy examples before hard ones can make it learn faster and end up better than showing the same data in random order, which is why frontier labs now build ladders of increasingly difficult tasks.
Learning rate schedules and warmup Lesson
A learning rate schedule changes how big each training step is over the course of a run, and warmup is the near-universal practice of starting tiny and ramping up -- without it, large transformer training runs frequently diverge in the first few hundred steps.
Loss of plasticity: why a network that trains forever stops being able to learn Lesson
Keep training a neural network on a stream of new tasks and it does not just forget the old ones - it gradually loses the ability to learn new ones at all, until it performs worse than a freshly initialized network.
Credit assignment: figuring out which step deserved the reward Lesson
Credit assignment is the problem of working out which of an agent's many decisions actually caused the eventual outcome, and it is the central difficulty in training anything that acts over a long sequence of steps.
A task factory ran fifteen rounds and broke the model grading it News
A new paper builds harder and harder terminal tasks by recursively rewriting accepted ones, and across fifteen rounds a fixed frontier solver's success rate fell from 90 percent to 2.5 percent, with the authors reporting no ceiling in sight.
Differential privacy: a mathematical promise about what a model can reveal about you Lesson
Differential privacy is a formal guarantee that a system's output would look almost the same whether or not any single person's data was included, enforced by adding carefully calibrated noise. It is the only privacy protection for machine learning that comes with a provable bound rather than a hope.
Model merging: combining two fine-tuned models by averaging their weights Lesson
Model merging combines two or more models that share a common ancestor by arithmetic on their weights - averaging them, adding their differences, or interpolating between them - producing a single model with both sets of skills and no additional training.
Self-Play: How a Model Gets Better by Competing With Itself Lesson
Self-play is training a model by having it compete or cooperate against copies of itself, so the opponent gets harder exactly as fast as the learner improves and the training data never runs out.
Imitation Learning: Teaching Machines by Showing, Not Rewarding Lesson
Imitation learning trains a system by having it copy demonstrations from an expert rather than discover behaviour through trial and error, and behaviour cloning - simply predicting what the expert did in each situation - is the simplest and most widely used version of it.
Training on the best of K guesses is a third scaling axis alongside parameters and data News
A paper from UIUC and Harvard shows that generating several candidate outputs per training example and learning only from the closest match improves sample efficiency 6.2-fold, and that the benefit grows rather than shrinks as models and datasets get bigger.
Catastrophic forgetting: why teaching an AI something new can erase what it knew Lesson
Catastrophic forgetting is the tendency of a neural network to lose an old skill when trained on a new one, because both skills are stored in the same weights and nothing protects the old settings.
A new distillation method lets the teacher model interrupt the student mid-thought News
Researchers found that when a student model starts reasoning down a wrong path, its teacher's next word tends to be a redirection like But or Wait, and turned that disagreement into an automatic trigger for the teacher to briefly take over.
Forward and Reverse KL Divergence Lesson
KL divergence measures how far one probability distribution is from another, and it is asymmetric: swapping which distribution goes first changes whether your model tries to cover everything the target does or commit to one part of it.
Mixed-precision training: why models are trained in half-broken numbers on purpose Lesson
Modern models are trained using 16-bit and even 8-bit numbers instead of the 32-bit standard, roughly doubling speed and halving memory, by carefully keeping full precision exactly where the arithmetic would otherwise fall apart.
Giving experts no memory cuts optimizer state from 50 gigabytes to 1.3 News
A single-author preprint shows that a mixture-of-experts model can drop momentum entirely for its expert layers, shrinking persistent optimizer state from 50.55 gigabytes to 1.29 with almost no effect on final quality.
Distributed training: how one model gets split across thousands of chips Lesson
No single chip can hold a frontier model, so training is split across thousands of them in four distinct ways -- by data, by layer, by tensor, and by expert -- and choosing the right mix is what separates a cluster running at a third of its potential from one running at a tenth.
A Huawei-chip training report shows what leaving CUDA actually costs News
SLAI's technical report documents full-parameter post-training of a DeepSeek-V4 model on Huawei Ascend hardware, and the work list -- rebuilt collectives, converted checkpoints, hand-written kernels -- is the real measure of chip independence.
Softmax and Cross-Entropy: How a Model Turns Scores Into a Confident Guess Lesson
Softmax turns a model's raw scores into a probability distribution, and cross-entropy measures how far that distribution is from the true answer; together they are the training target behind almost every classifier and language model.
Multi-Token Prediction: Teaching Models to Look Several Words Ahead Lesson
Multi-token prediction trains a language model to predict several upcoming tokens at once instead of just the next one, giving a richer training signal and a built-in path to faster generation.
Reinforcement Learning with Verifiable Rewards (RLVR) Lesson
RLVR trains a language model by rewarding answers that an automatic checker can confirm are correct - a passing unit test, a right final number, a valid proof - instead of using human ratings or a learned reward model. It is the training recipe behind the modern wave of reasoning models.
Regularization: why deliberately handicapping a model makes it better Lesson
Regularization is the family of techniques that make a model perform worse on its training data in order to make it perform better on data it has never seen, with dropout and weight decay the two that shaped modern deep learning.
Layer normalization: the rescaling that keeps deep networks trainable Lesson
Layer normalization rescales the numbers flowing through each layer of a network to a consistent range, which stops signals from exploding or vanishing as they pass through dozens of layers -- and it is what makes training a transformer at all possible.
Residual Connections: The Shortcut That Made Deep Networks Possible Lesson
A residual connection is a shortcut that adds a layer's input directly to its output, so the layer only has to learn the change rather than rebuild everything from scratch — a simple trick that lets networks be hundreds of layers deep without collapsing, and the reason modern transformers can be stacked as deep as they are.
Optimizers: how Adam and AdamW turn gradients into learning Lesson
An optimizer is the rule that decides how a neural network changes its weights after each mistake; Adam and its refinement AdamW became the default because they adapt the step size for every weight, making training faster and far less finicky than plain gradient descent.
A giant benchmark tested 24 optimizers - and AdamW's edge held up News
OmniOpt ran a controlled bake-off of more than two dozen modern training optimizers across model sizes from 60M to 1B parameters, and its main lesson is deflating: no challenger cleanly dethrones AdamW, because an optimizer's advantage depends heavily on scale, task, and tuning budget.
The RL 'mirage': the policy you optimize isn't the one you ship News
A top-ranked paper shows that in modern RL training, improving the model you optimize does not guarantee improving the model you actually deploy - because the two run on different engines that disagree on probabilities.
Group Relative Policy Optimization (GRPO) Lesson
GRPO is the reinforcement learning method behind most modern reasoning models: instead of training a separate 'critic' network to judge answers, it samples a whole group of answers to each question and scores each one by how it compares to the group's average.
Grokking: When a Model Suddenly 'Gets It' Long After It Should Have Lesson
Grokking is a training phenomenon where a neural network first memorizes its training data with near-zero understanding, then -- after a long, flat plateau of continued training -- abruptly generalizes and starts solving unseen examples correctly.
Three Popular Ways to Train Reasoning AIs Turn Out to Be One Formula News
A new proof shows that three widely used reinforcement-learning recipes for training reasoning models - GRPO, Dr. GRPO, and DAPO - are all just different operations on a single number, the spread of rewards within a group of sampled answers.
What Is Gradient Descent? Lesson
Gradient descent is the optimization method that trains almost every modern neural network: it repeatedly nudges the model's parameters a small step in the direction that most reduces its error, until the error stops falling.
On-Policy vs Off-Policy Learning Lesson
On-policy learning trains a model on data generated by its own current behavior, while off-policy learning trains it on data generated by something else -- an old version, a different policy, or a fixed dataset -- and the choice shapes how stable, sample-efficient, and reliable the training is.
Fine-tuning and LoRA: teaching an old model a new job without retraining it Lesson
You almost never train an AI from scratch. You take one that already knows language and nudge it toward your specific task - and a trick called LoRA lets you do that by adding a tiny sticky note instead of rewriting the whole brain.
A 35-billion-parameter agent that punches like a trillion-parameter model News
Shanghai AI Lab argues you can reach giant-model performance on long tasks not by adding parameters, but by training on much longer chains of real work.
Backpropagation: how a neural network learns from its mistakes Lesson
The single algorithm behind nearly all AI training - assigning blame for an error backward through millions of dials, so each one knows which way to turn.
An AI's hallucinations turned out to be a map with blank spots News
Researchers showed that when a world-model AI imagines impossible futures, it's usually in places it barely saw in training - and that you can predict and fix those blind spots cheaply.
A wave of new methods trains AI without a human answer key News
Several research groups landed on the same idea at once - improve a model by learning from its own attempts instead of expensive human labels - and the field is debating whether it really removes the labeling burden or just hides it.
Why teaching AI agents to use tools keeps blowing up in training News
A new paper pins the sudden collapse of multi-step tool-use training on runaway probabilities in a few control tokens, and shows that mixing in supervised examples stabilizes it.
Training vs inference: the two very different jobs inside every AI Lesson
Why building an AI model and using it are separate worlds with separate costs, and why that split explains custom chips, model prices, and where the real money in AI actually goes.
Distillation: how a small AI learns from a big one Lesson
Distillation trains a smaller, cheaper model to imitate a larger, smarter one, the idea behind both efficient deployment and the 'copying' accusations now driving AI geopolitics.
Teaching AI with rewards — minus the expensive second model that grades it News
The standard way to polish a model with rewards quietly runs a second 'critic' model alongside it. A new method derives the critic's judgment from the model itself, dropping the extra cost.
Scaling laws — does bigger always mean better? Lesson
For years, AI progress ran on a simple recipe: make the model bigger, feed it more data, get a better model. That pattern is real and predictable — but it has limits and surprises. Here's what scaling laws actually say.
Polishing AI by looking inside its 'mind' instead of just thumbs-up, thumbs-down News
Reward training usually treats the model as a black box — thumbs up, thumbs down, hope for the best. A new method peers inside to see why an answer was preferred, and shapes the lesson on purpose.
The little words that keep AI from getting boring News
Rewarding a reasoning model too hard makes it repetitive — and the casualties are tiny words like "but" and "instead" that let it branch to a better thought. A near-free fix protects them.
Faster AI training by quietly cloning the model News
Teaching a model with rewards is slow because it has to write out endless practice answers. A new trick: make a cheap, shrunk-down copy of the model to crank those out faster.
Crediting an AI for the right steps — without a second model to judge them News
When you reward an AI for a good final answer, it's hard to know which of its steps earned the credit. The usual fix is training a second 'judge' model. This skips that.
slime Tool
The open-source large-scale asynchronous training framework from THUDM that Z.ai used to run the post-training scaling behind GLM-5.3.
prime-rl Tool
Open-source RL post-training stack that splits rollout generation and gradient updates across GPUs, used in this week's widely discussed $500 fine-tune that beat five frontier configurations on a catalog-review workflow. Practical for teams that already have an automatically scored task and want to train a specialist rather than pay per call for a frontier model.
Unsloth Studio Tool
A no-code desktop and web interface for training and running language models on your own machine, with a one-line installer and desktop shortcuts. Training is NVIDIA-GPU centric; inference and export work across Mac, Windows and Linux.
Unsloth (AMD support) Tool
The fine-tuning and RL toolkit now documents AMD support across training, RL, chat, and deployment on Windows, WSL, and Linux, plus a cross-platform Studio beta.
Skill Self-Play Tool
Apache-2.0 release of a system that grows and prunes a library of skill packages, each with routing metadata, examples and an executable validator, then trains a solver on the tasks they generate. Includes benchmark material and training launchers; expects eight visible GPUs.
SLAI T-Rex Tool
The public workflow behind a full-parameter Ascend post-training run on a DeepSeek-V4-family model: FP8-to-BF16-to-Megatron checkpoint conversion, launch templates, and inspectable data-construction pipelines for continued pre-training and supervised fine-tuning. The production engine and custom kernels are withheld.
PyLate Tool
A training and retrieval library for late-interaction models, built on Sentence Transformers, for people who want to fine-tune a retriever on their own corpus rather than use an off-the-shelf embedding API.
MindSpeed-LLM Tool
Huawei's official large-model training toolkit for Ascend NPUs, covering distributed layouts, checkpoint conversion and supported model families. Worth reading its support table honestly -- DeepSeekV4-Flash is currently marked Prototype, its label for not-fully-validated features.