Clinical trial failures pose a major obstacle to efficient drug development, incurring high costs and delaying patient access to effective therapies. Structured approaches such as the Hierarchical Interaction Network (HINT) predict outcomes by explicitly modeling drug‑disease‑protocol interactions, yet they make limited use of the free‑text information in trial registrations. We evaluated two domain‑specific language models—BioBERT (108 M parameters) and BioMistral (7 B parameters)—on the TOP benchmark (12,477 trials), representing each trial as concatenated text from ten ClinicalTrials.gov fields. Two training regimes were compared: full fine‑tuning for BioBERT, low‑rank adaptation for BioMistral, and probing, where the pretrained model is frozen and only a lightweight classifier is trained on its embeddings.\n\nOn the held‑out test set for each phase, all four model variants achieved higher ROC‑AUC (0.613–0.772) and PR‑AUC (0.649–0.901) than the published HINT results, corresponding to relative PR‑AUC gains of 10%–20%. F1 differences were small and mixed, and no model—including HINT—clearly outperformed a trivial classifier that always predicts success. Under the tested configurations, probing performed comparably to fine‑tuning: fine‑tuned models scored highest on ROC‑AUC and PR‑AUC in Phases I and II, while frozen BioMistral embeddings led in Phase III (ROC‑AUC 0.772; PR‑AUC 0.901), although most differences were below 0.03 and were not tested for significance.\n\nThese findings suggest that pretrained biomedical representations already encode signals relevant to trial outcomes, and that probing offers a computationally cheaper alternative to fine‑tuning for this task.\n\nReview: Domain‑specific large language models can effectively leverage textual information from clinical trial registrations to improve outcome prediction; in resource‑constrained settings, freezing the model and applying a probing classifier provides an efficient and reliable approach.