biomedical ai developers

Abstract: Coding agents are increasingly proposed as autonomous builders of biomedical models, but existing benchmarks are small, near saturation or open to leaked information. We present BioXArena-Pro, 64 research-level modelling tasks from eight biomedical domains, each rebuilt from a published study; most hold out whole biological units such as species or cohorts. Across 22 agent configurations, the best scores 55.1 of 100, and the scaffold matters almost as much as the model. We re-ran a published method for every task: where the published recipe applies unchanged, the best agent and the methods tie; on rebuilt tasks the methods trail by 8.2 points, suggesting that agents gain by adapting to new data. Pretrained models (including biological foundation models) matter where the strongest agent uses them: forbidding them decreases its score on those tasks by 9 points. Every task was audited in several AI-assisted rounds and a final human review.

Abstract: On-policy distillation (OPD), which uses reverse-KL supervision on student rollouts to transfer a context-conditioned model's behavior into student model parameters for closed-book inference, has shown strong results. However, we find that this success does not translate to knowledge injection: across two document-QA benchmarks, OPD injects document knowledge far less efficiently than simple off-policy distillation. We trace this reversal to two failures that are structural, yet sparse across token positions. Reverse KL weights its gradient by the student distribution and therefore barely lifts document-specific tokens the student considers unlikely; after forward KL restores this signal, student rollouts expose later positions on the target trajectory only after generating the preceding tokens correctly, so coverage deteriorates with target length. Motivated by the concentration of teacher-student divergence at these hard positions, we propose Adaptive-Policy Distillation (APD), whose per-token gate jointly shifts the rollout toward the teacher and the objective toward forward KL where needed, withdrawing both interventions as the student learns. Without labels or a verifier, APD improves over OPD by up to 54.6%, achieving the best scores of 0.558 F1 on QASPER and 0.486 MRR on BioASQ.

Abstract: Language-model agents now carry out long scientific work in biology, but their judgements about conventions, skills, errors, evidence and stopping are made inside the reasoning that does the work, where they cannot be measured, replaced or improved on their own. Here we present Operon-S1, a fast decision system, a System 1 for the agent, that separates judging from doing. A decision harness poses the agent's recurring judgements as closed questions at 13 decision points and acts only through fixed messages that never contain an answer, so any model that answers the questions can serve as the judge. Across seven life-science benchmarks, a local decision model took over these judgements in 0.6 s at the median for an in-loop request on CPU at no request cost, with no measurable change in the agent's score on five of six benchmarks, whereas the general language model that drives the agent, asked the same questions, accounted for 8.0% to 35.2% of task cost. Every decision is recorded with its consequence. Retraining on labels from task construction, verifiers or execution raised decision accuracy on every question in each of three trained versions, and a new version enters service only through a promotion gate. Task-level results in this study use the public checkpoint. The same closed decisions serve wet-laboratory protocol troubleshooting and the design of a germline target enrichment panel. Like the regulator of an operon, which switches genes on and off without making their products, Operon-S1 steers the agent without doing its work, and makes the agent's judgements a layer that can be measured, replaced and trained.

Abstract: Retrieval-Augmented Generation (RAG) systems are increasingly used for biomedical Question Answering (QA). However, evaluating their effectiveness remains challenging due to the lack of reliable reference resources and the frequent misalignment between automatic evaluation metrics and human judgments. One practical question is how evaluation results should be used to guide development decisions, especially when they correlate weakly with human judgments and may even lead to different conclusions. We explore how to use a small amount of human preference signals to calibrate direct assessment evaluation metrics, and show that calibrated metrics consistently outperform baselines that either treat all metrics equally or rely on a single best metric across multiple datasets.

Abstract: Vision-Language Models (VLMs) have shown strong performance across a wide range of downstream vision tasks, thanks to the complementary information contained in the respective domains. Despite the performance gains, most of these approaches rely on aligning these domains using the cosine similarity metric, which fails to capture token-level structure and cross-modal interactions prior to the classification stage. This is especially critical in biomedical applications under federated constraints, where data sharing is restricted, labeled data is scarce at each site, and it differs widely across institutions, leading to substantial statistical heterogeneity. To overcome this issue, we propose FedSSMCoOp, a federated few-shot image classification framework that enables multimodal learning while preserving data privacy. With the help of the SSM-based Vision Mamba and Cross Mamba blocks, and by optimizing only the soft-prompt and communication-prompt updates in the federated setting, the framework prioritizes both computation and performance. Importantly, this eliminates the need to use an external Large Language Model (LLM) for feature alignment. The framework is further trained and evaluated on various biomedical image datasets, and its performance is assessed. The proposed framework delivers stable performance relative to the baselines and is, on average, 1.96 times lighter. The corresponding script will be made available soon.

Abstract: Repeatedly training language models on their own generated data may create a synthetic-data feedback loop in which errors and distributional biases are reintroduced into subsequent training datasets. This paper studies that process in biomedical question answering (QA) using PubMedQA and two Qwen2.5 model sizes, 0.5B and 3B parameters. The study compares a recursive synthetic-data condition, in which generation G(k+1) is trained on answers produced by G(k), against a Human-Control condition that repeatedly uses the original human training data. The study evaluates across four generations from G0-G3 with two random seeds (42 and 123) and a fixed evaluation set of 1,000 expert-labeled samples. The evaluation includes disease and chemical entity F1, context-supported rate, lexical and semantic similarity, answer length, repetition rate, and other evaluation metrics. The Recursive condition for both model sizes and both seeds showed larger declines than the Human-Control condition in disease entity F1, chemical entity F1, context-supported rate, ROUGE-L, and cosine similarity. Under the fixed no-repeat 3-gram decoding constraint, the main observed behavioral change was increased answer length, while the measured 3-gram repetition rate did not increase. The magnitude of the difference-in-change was larger for the 3B model than for the 0.5B model. This difference was particularly apparent in disease F1, context-supported rate, cosine similarity, and answer length. These results show domain-specific changes associated with using recursive synthetic-data training in biomedical QA, but do not establish clinical hallucination rates or universal model collapse.

Abstract: Automated biomedical evidence synthesis depends on retrieving published studies, but the biomedical literature is systematically skewed toward positive findings. Deeper retrieval can therefore make a system \emph{more} likely to falsely infer benefit when the true effect is null. We formalise this phenomenon as \emph{evidence drift} and prove that, under a standard publication-bias model, the false-positive probability on null-effect queries follows a strictly increasing large-sample envelope in retrieval depth, approaching one. Empirically, on a held-out test set of 140 Cochrane-derived queries, drift rises monotonically from 7.9\% to 15.7\% as the retrieval budget grows from 3 to 20 steps, and concentrates in the null-effect class. We present DACG-agent, a drift-aware causal-graph agent that incrementally builds a causal knowledge graph from PubMed abstracts and applies a two-layer stopping policy with complementary roles: a KL-divergence monitor that detects posterior convergence (the accuracy layer), and a Bradley--Terry process reward model (PRM) whose online decline detection halts retrieval once evidence quality peaks (the efficiency layer). Against full-budget retrieval, DACG-agent reduces evidence drift from 15.7\% to 6.4\% and improves null-effect accuracy by 21 percentage points (40.0\%$\to$61.4\%) while using 67\% fewer retrieval steps; overall accuracy rises from 61.4\% to 69.3\% (95\% CI 61--77). A simulation confirms the drift result transfers from the analysed vote-counting aggregator to the deployed noisy-OR one.