How Post-Training Shapes Biological Reasoning Models

Post-training stages in biological reasoning models reshape generalization in distinct, sometimes opposing ways, according to a study that trained and evaluated more than 100 models across genomics, transcriptomics, and proteins [1][2]. The research, submitted to arXiv on 15 June 2026, examines how continued pre-training, supervised fine-tuning, and reinforcement learning affect both in-domain and out-of-domain performance [1][2]. Biological reasoning models combine large language models with foundation models trained on multimodal biological data, including DNA, RNA, and proteins [1][2]. Foundation models are machine learning or deep learning models trained on vast datasets so they can be applied across a wide range of use cases [4]. Large language models are typically based on transformer architecture and are pre-trained to predict the next word before being fine-tuned for specific tasks [3]. The study found that continued pre-training improves downstream performance by aligning models with biological language [1][2]. Supervised fine-tuning consistently increases in-domain performance but causes out-of-domain performance to peak early and then decline as models fit the training distribution [1][2]. Reinforcement learning, when applied to strong supervised fine-tuning checkpoints with aligned rewards, improves out-of-domain performance and partially recovers generalization [1][2]. These findings indicate that biological reasoning does not improve monotonically with additional supervision or compute [1][2]. Instead, performance depends on how training stages are composed [1][2]. Under fixed post-training budgets, the strongest in-domain to out-of-domain trade-off comes from brief supervised fine-tuning, larger reinforcement learning allocations, and asymmetric adaptation capacity across stages [1][2]. The paper was posted on arXiv, an open-access repository of electronic preprints that is not peer reviewed but is moderated before posting [10]. The repository, which began in 1991, surpassed two million articles by the end of 2021 and now receives about 24,000 submissions per month [10]. The study appears under the machine learning category and is accessible through arXivLabs, a framework that allows community collaborators to develop and share experimental tools on the platform [7][8].