You can make a frozen AI model smarter at inference time without retraining it — it works for classifying ocean creatures

Researchers found that a computer vision model frozen in place can still improve its answers by running some of its internal layers twice during inference, rather than once. For underwater species classification with limited labeled data, this trick closes the gap between cheap frozen models and expensive fully-trained ones by 1.4 percentage points, and on some species it beats the fully-trained baseline outright.

The constraint in marine biology is annotation cost — labeling underwater images is expensive and time-consuming. Standard practice is either pay for full annotation and fine-tune a model, or use a cheap frozen model and accept lower accuracy. This shows a third option: keep the model frozen but change how you run it at prediction time. The practical implication is that labs with limited budgets can get closer to fully-supervised accuracy without hiring annotators or running GPU-intensive training. The structural change is small but real — it decouples inference-time compute from fine-tuning, meaning inference-optimizing tricks designed for language models now work for vision too.