Sakana AI Doc-to-LoRA: Hypernetwork Turns Any Document into a LoRA Adapter in Under 1 Second, Cuts KV-Cache from 12GB to Under 50MB

Sakana AI Research·high signal

Sakana AI's Doc-to-LoRA uses a once-trained hypernetwork to generate LoRA adapters that internalize document content directly into model weights rather than KV-cache, reducing memory overhead from over 12GB to under 50MB for long documents. The approach handles sequences 4x the native context window of the base LLM with near-perfect recall and generates adapters in sub-second latency after one-time meta-training — enabling instant, memory-efficient document knowledge injection without inference-time context stuffing. GitHub repo at SakanaAI/doc-to-lora.