A comparison at inference time
A computational preprint reports different benchmark figures for protein language models across three inference protocols. The evaluation covered three representative models and four tasks: unconditional sequence-structure co-generation, motif scaffolding, structure prediction and inverse folding. The reported evidence is computational and tied to those named model-task settings.
The evaluation used three stages. It began with standardized vanilla sampling, with a grid search across six protocol dimensions. The next stage was task-specific classifier-free guidance, or CFG, which operates on the model's per-step logits - the scores used while generating. The final stage used reward-guided beam search, keeping multiple ESM3 generation paths in parallel and selecting among them with global rewards produced inside the model.
For unconditional co-generation, the evaluation sampled 100 proteins at each of five target lengths: 100, 200, 300, 400 and 500 residues. For motif scaffolding, it generated 100 candidate scaffolds for each of 24 benchmark problems. Unless otherwise indicated, the reported results combined five independent runs with distinct random seeds and were summarized as means with standard deviations or as means with minimum-to-maximum ranges.
What the ESM3 comparisons reported
On ESM3 unconditional co-generation, optimized vanilla sampling reported 174.8 designable samples and 64.4 unique designable clusters. Default sequence-to-structure sampling reported 52.8 designable samples and 34.4 unique designable clusters.
On the same task, ESM3 with CFG reported pLDDT of 88.748 and scTM of 0.931, alongside 325.0 designable samples and 116.6 distinct designable clusters. The corresponding optimized-vanilla figures were 174.8 samples and 64.4 clusters.
On motif scaffolding, ESM3 with CFG solved 23 of 24 benchmark cases in four of five independent runs. Its average success rate was 37.5%, and its average number of unique solution clusters was 179.4, compared with 94.2 for the default setup and 110.0 for optimized vanilla sampling.
For structure prediction, ESM3 with CFG reported RMSD/TM-score pairs of 5.144/0.870 on CAMEO2022 and 3.969/0.889 on the PDB Date split. For inverse folding, average scTM was 0.910 on CAMEO2022 and 0.954 on the PDB Date split.
A further search step
In the final search stage, multiple ESM3 trajectories were maintained in parallel and selected with global, model-internal rewards. The reported beam-search procedure used threshold-based random selection among those paths.
On unconditional co-generation, the reported designable-cluster figures were 116.6 under CFG and 139.0 with beam search. On motif scaffolding, beam search solved 23 of 24 problems in every run; its reported success rate was 53.3%, versus 37.5% under CFG, and its average solution-cluster count was 245.2, versus 179.4 under CFG. For inverse folding, beam-search scTM was 0.914 on CAMEO2022 and 0.956 on the PDB Date split.
A trade-off in computing cost
The efficiency analysis reported a computational trade-off. CFG was reported with one to three times runtime overhead. For ESM3 unconditional co-generation, runtime was about 1.46 times the corresponding vanilla setting, while designable clusters were about 1.81 times as high. Beam search was reported with substantial additional inference cost.
The scope of the comparison
The findings are bounded by the computational scope: three representative protein language models and four controlled tasks. They concern benchmark performance within those named settings.
The document is an arXiv version 1 preprint dated 26 Aug 2026. Its abstract states that code is available at github.com/EchoChou990919/mplm_inference.
Paper data and sources
Original title: Unlocking Multimodal Protein Language Models at Inference Time
Authors: Yi Zhou, Qipeng Wang, Yunqing Liu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text