flamingo-cross-attention-over-frozen-llm
IN premise — summaries/2026/08/24/wiki-Large_language_model-chunk-5.md
Created 2026-08-24T17:11:17+00:00
Flamingo achieves few-shot visual instruction following by adding cross-attention layers over a frozen language model backbone, without modifying the LLM weights
Summary
Flamingo demonstrates that you can give a large language model the ability to follow visual instructions with only a few examples by bolting on lightweight cross-attention modules, rather than retraining the model. This keeps the expensive, already-trained language backbone completely untouched, meaning multi-modal capability becomes a cheap plug-in feature instead of requiring a full system redesign.