Researchers Advance Reinforcement Learning Methods for Clinical Decision-Making and AI Agent Training
Researchers have published multiple studies exploring reinforcement learning (RL) applications in partially verifiable environments, ranging from clinical decision-making to AI agent training.
According to arxiv.org, a new study formulates oncology treatment planning as an interactive RL problem under partial observability. The research uses GRPO and GDPO to optimize Llama-3.1-8B-Instruct, with the GDPO-trained policy increasing “Mean@3 scalar clinical reward by $79.60%$ over base Llama and exceeds GPT-5.2 by $18.82%$ under the same interactive environment, while reducing query count during training.”
In tool orchestration research, arxiv.org introduces ProgAgent, a training system using Programmatic Tool Calling (PTC) where tools are invoked through executable code rather than accumulating in model context. According to the source, “on LOCA-bench, ProgAgent-14B achieves a 30.7% pass rate at a 64k-token environment state, where these baselines score zero,” using only 215 tasks across four environment scales.
For video understanding, arxiv.org presents EventLens, an event-structure reinforcement learning framework that trains models to recover temporal structure through tasks including “event segmentation, progression discrimination, and multi-event relation reasoning.” The system uses “task-specialized GRPO training without human preference labels or learned reward models.”
Additionally, arxiv.org published a comprehensive survey on Flow Matching in reinforcement learning, noting it “has recently emerged as a principled and efficient generative modeling framework” that can “substantially reduce inference latency” compared to diffusion-based approaches. The survey was published in Transactions on Machine Learning Research (TMLR) in 2026.