Unlocking Multimodal Protein Language Models at Inference Time

TL;DR - This paper systematically studies inference-time sampling for multimodal protein language models across three models and four tasks. It finds that task-tailored inference strategies can substantially improve generation quality without updating model parameters.

• Compares vanilla sampling, task-specific classifier-free guidance, and reward-guided beam search.
• Examines three levels of inference control: sampling distributions, per-step logits, and parallel generation trajectories.
• Shows that default inference protocols are often suboptimal and that preferred exploration-exploitation trade-offs vary by task.
• Reports consistent quantitative gains that raise model performance ceilings and challenge prior conclusions about the underlying models.