IC-1496A single frozen transformer block from LLaMA-7B consistently improves performance across diverse visual tasks when appended to existing visual encoders

Ziqi Pang, Ziyang Xie, Yunze Man, Yu-Xiong Wang

SourceFrozen Transformers in Language Models Are Effective Visual Encoder Layers

The paper appends one frozen transformer block from LLaMA-7B (the last block by default) to various visual encoders and trains the rest of the network. Across image classification (ViT), point cloud classification (Point-BERT), action recognition (ViT on SSV2), motion forecasting (VectorNet, MMTransformer), and vision-language tasks (METER, SQA3D), the LLaMA block yields consistent gains. Ablations with a matched-capacity MLP and with randomly initialized LLaMA blocks confirm the improvement comes from the pre-trained LLaMA weights rather than added parameters. Fine-tuning the LLaMA block instead of freezing it degrades performance due to overfitting.

Evidence
correlational
Key metric
ImageNet: ViT-t 72.1→73.2, ViT-s 80.1→80.7, ViT-b 80.6→81.7; ScanObjectNN: 87.4→88.0 (bg), 88.1→88.5 (obj), 83.1→83.8 (t50); SSV2: ViT-s 64.71→65.89, ViT-b 64.97→66.03; Argoverse ADE: VectorNet 0.77→0.76, MMTransformer 0.72→0.71; VQAv2 overall: 69.60→70.23; Flickr30k IR@1: 49.66→50.22; SQA3D EM@1: 47.20→48.09; Capacity ablation: ViT-s-MLP 80.4 vs ViT-s-LLaMA 80.7 vs ViT-s 80.1
Caveat
Experiments are designed for diversity of tasks under fair comparisons rather than state-of-the-art performance; computational resources limited batch sizes (e.g., 24 clips vs 384 in VideoMAE), so absolute numbers are lower than original papers.
Model
LLaMA
Datasets
ImageNet-1k / ImageNet / ImageNet-1k-val / ImageNet-Val [eval], ImageNet-C [eval], ImageNet-A [eval], ImageNet-SK [eval], ImageNet-R / ImageNet-Rendition [eval], ScanObjectNN [eval], ModelNet40 [eval], Argoverse [eval], VQAv2 [eval], Flickr30k [eval], SQA3D [eval]
Methods
DeiT [supporting]
Related work
VideoMAE [compared-to]
Related findings
IC-1497, IC-1498
Extraction
automatic-extraction