Unlocking Multimodal Protein Language Models at Inference Time

Multimodal protein language models (pLMs) like ESM3 and DPLM-2 have emerged as powerful tools for protein sequence-structure modeling. However, prior research has overwhelmingly focused on training and fine-tuning, leaving the inference-time sampling strategies—which dictate how these models navigate their learned probability landscapes—largely underexplored. This paper investigates whether the performance of these models is limited by their default inference protocols rather than their underlying capabilities.

The authors establish a three-stage investigation framework to explore the inference design space across three representative pLMs and four fundamental tasks: unconditional sequence-structure co-generation, motif scaffolding, protein structure prediction, and inverse folding. The framework organizes inference controls into three levels:

1. Vanilla Sampling: Benchmarking the foundational sampling distribution by optimizing hyperparameters like diffusion steps, unmasking strategies, and temperature settings.
2. Classifier-Free Guidance (CFG): Steering per-step logits to amplify condition-aligned signals, effectively shifting the sampling trajectory toward more plausible protein states.
3. Reward-Guided Beam Search: Navigating parallel sampling trajectories using model-internal quality signals (e.g., pTM scores) to select superior outputs.

The study reveals that default inference protocols are often suboptimal and act as a hidden confounder in model evaluation. By standardizing and optimizing these protocols, the authors observe substantial quantitative gains across all tasks. Notably, ESM3, which was previously considered an underperforming baseline in certain contexts, achieves state-of-the-art performance when equipped with the proposed inference strategies. The authors identify that inference preferences are task-oriented: exploratory tasks (like unconditional generation) benefit from longer trajectories and higher randomness, while exploitation-biased tasks (like structure prediction) favor shorter, deterministic trajectories. These findings suggest that the perceived limitations of base models are often a byproduct of inadequate inference-time steering.