AI Glossary
What is LoRA (Low-Rank Adaptation)?
LoRA freezes the original model weights and trains tiny low-rank matrices alongside them, cutting fine-tuning cost and memory by orders of magnitude. The resulting adapters are megabytes instead of gigabytes and can be swapped at runtime — the dominant technique for customizing open-weight models and image models alike.
News momentum
Mentions in recent AI news titles and summaries, refreshed from the ingestion stream.
31
-3 vs prior 7d
- Atomic MathReasoner: Instance-Level Composition of Atomic Reasoning Skills for Mathematical LLMs
arxiv-ai · 1d
- New LoRA Adapter Lifts Qwen3.8 Coding Success From 39% to 65%
marktechpost · 1d
- Self-Rewarded Reinforcement Learning for Better-Structured Medical Explanations in Small Language Models
arxiv-ai · 1d
- H3Remover: Reinforcing Video Object Removal without Paired Data
arxiv-ai · 1d