Gregory D. Lyng
Publications
Fast and Effective On-policy Distillation from Reasoning Prefixes
On-policy distillation (OPD), which samples trajectories from the student model and supervises them with a teacher at the token level, avoids relying solely on verifiable terminal rewards and can yield better generalization than off-policy distillation. However, OPD requires expensive on-the-fly sampling of the student policy during training, which substantially increases training cost, especially for long responses. Our initial analysis shows that, during OPD, training signals are often concentrated in the prefix of each output, and that even a short teacher-generated prefix can significantly help the student produce the correct answer. Motivated by these observations, we propose a simple yet effective modification of OPD: we apply the distillation objective only to prefixes of student-generated outputs and terminate each sampling early during distillation. Experiments on a suite of AI-for-Math and out-of-domain benchmarks show that on-policy prefix distillation matches the performance of full OPD while reducing training FLOP by 2x-47x.
POET: Protocol Optimization via Eligibility Tuning
Eligibility criteria (EC) are essential for clinical trial design, yet drafting them remains a time-intensive and cognitively demanding task for clinicians. Existing automated approaches often fall at two extremes either requiring highly structured inputs, such as predefined entities to generate specific criteria, or relying on end-to-end systems that produce full eligibility criteria from minimal input such as trial descriptions limiting their practical utility. In this work, we propose a guided generation framework that introduces interpretable semantic axes, such as Demographics, Laboratory Parameters, and Behavioral Factors, to steer EC generation. These axes, derived using large language models, offer a middle ground between specificity and usability, enabling clinicians to guide generation without specifying exact entities. In addition, we present a reusable rubric-based evaluation framework that assesses generated criteria along clinically meaningful dimensions. Our results show that our guided generation approach consistently outperforms unguided generation in both automatic, rubric-based and clinician evaluations, offering a practical and interpretable solution for AI-assisted trial design.
Health-SCORE: Towards Scalable Rubrics for Improving Health-LLMs
Rubrics are essential for evaluating open-ended LLM responses, especially in safety-critical domains such as healthcare. However, creating high-quality and domain-specific rubrics typically requires significant human expertise time and development cost, making rubric-based evaluation and training difficult to scale. In this work, we introduce Health-SCORE, a generalizable and scalable rubric-based training and evaluation framework that substantially reduces rubric development costs without sacrificing performance. We show that Health-SCORE provides two practical benefits beyond standalone evaluation: it can be used as a structured reward signal to guide reinforcement learning with safety-aware supervision, and it can be incorporated directly into prompts to improve response quality through in-context learning. Across open-ended healthcare tasks, Health-SCORE achieves evaluation quality comparable to human-created rubrics while significantly lowering development effort, making rubric-based evaluation and training more scalable.