Train Only Projector to Add New Modalities Without Regressing LLM Capabilities

rohanpaul_ai · x · 2026-08-25

The paper 'Projector Is All You Train' demonstrates that adding a new modality (e.g., 3D) to an LLM requires training only the projector mapping encoder outputs to the embedding space, leaving the backbone frozen. This approach trains twice as fast as joint fine-tuning and matches or outperforms it on 3D tasks. Conversely, joint fine-tuning caused the Llama backbone's GSM8K accuracy to collapse from 86.96% to 0.61%.

Original post →

More from Models

Models channel →