Post-Training Models
Hiteshwari Patel · NIT Surat · Nov 1, 2026
When a model learns tasks one after another, it needs a way to choose the right knowledge for each new input, not only to protect what it has already learned. This talk reviews Latent-LoRA, a paper on replay-free continual learning. Hiteshwari Patel explains how each task gets a very small adapter built on a frozen decomposition of the model's weights, how an orthogonality penalty limits interference between tasks, and how a router based on a Gaussian mixture over average token embeddings picks the right adapters without any gradient training. The talk also covers how the paper is evaluated, the results it reports, and its limitations.
Talks like this are given by GroupBrain students and researchers. They pick a topic, read the papers, run experiments and present what they learn to the group. You can do the same.