Wang and colleagues' October 21 preprint explores language-guided refinement after automated proton planning.

The big change

  • What changed: Candidate plans become evidence for interpreting further instructions.
  • Why it matters: Medical physicists can inspect the trade-offs behind proposed adjustments.
  • What to watch: Whether those adjustments withstand physical uncertainty.

From instructions to recalculated doses

Proximal Policy Optimization (PPO) generates candidate plans. Adapted Qwen3-14B retrieves objectives and metrics, selects a reference trade-off, and proposes API changes for human confirmation. Inverse optimization and Monte Carlo calculation produce the revised dose.

The retrospective study used 21 cases, with 42 training samples and eight test samples spanning five patients. The authors report improvements over base Qwen3-14B, with mixed trade-offs; patient separation between splits is unclear. Reinforcement learning rewards LLM-judged agreement with manually revised responses.

The framework omits robust optimization. Clinical outcomes and measured savings remain unestablished.

Why robustness is a separate question

In Liu and colleagues' 2012 study, robust optimization accounts for uncertainty about patient positioning and how far protons travel. Their method evaluates several possible dose distributions during optimization, incorporating the worst calculated doses for target and surrounding tissue into its objective.

That work also explains why comparing nominal dose distributions alone can mislead: plans should be evaluated under comparable uncertainty assumptions. Its own approximation treats uncertainties separately and does not represent anatomical deformation. An objective can describe a desirable dose while uncertainty changes the dose actually delivered.

BIG CHANGE assessed the papers; it did not test the system.