Eliciting Fine-Tuned Transformer Capabilities via Inference-Time Techniques