🔆 SHINE: Context → LoRA in a Single Pass

SHINE is an in-context hypernetwork that reads a context document and generates a LoRA adapter for the Qwen3-8B backbone in a single forward pass. In SHINE mode the context is encoded entirely into the LoRA weights — it is not placed in the prompt — and the model answers questions using only that adapter. Compare against In-Context (context in the prompt) and Only Question (no context).

Paper: arXiv 2602.06358 · Code: MuLabPKU/SHINE · Checkpoint: Yewei-Liu/SHINE-ift_mqa