News
MIT researchers double the training speed of reasoning models
Adaptive Drafter uses otherwise-idle GPUs during the RL training phase and speeds up reasoning training by 70 to 210 percent.
Training & Fine-Tuning
Verfahren zum Trainieren und Anpassen von Sprachmodellen.
News
On-Policy Distillation: 30 times cheaper to train your own reasoning model
Thinking Machines Lab shows how small models learn reasoning without full RL. Qwen3-8B hits 70 percent on AIME'24 at one-thirtieth the compute cost.
DPO (Direct Preference Optimization) Fine-Tuning LoRA PEFT (Parameter-Efficient Fine-Tuning) QLoRA RLHF SFT (Supervised Fine-Tuning) Synthetic Data Adapter (PEFT) Catastrophic Forgetting Continued Pretraining Instruction Tuning Knowledge Distillation ORPO (Odds Ratio Preference Optimization) Prefix Tuning