TL;DR
OpenGray is a new benchmark designed to evaluate the decision-making logic of LLM-based radiotherapy planning agents, moving beyond simple dose-quality metrics. The study reveals that while these agents can achieve high aggregate accuracy, they frequently fail to recognize anatomically impossible requests, highlighting a critical need for explicit "escalation" mechanisms in clinical AI workflows.
What Problem Does This Paper Address?
In radiotherapy, planning agents are increasingly tasked not just with optimizing radiation beams, but with managing the entire workflow: interpreting physician notes, selecting objectives, and deciding when a plan is ready for clinical use. Traditionally, these systems are evaluated by the quality of the final dose distribution.
However, dose alone is a "silent" metric. It doesn't tell you if an agent blindly optimized an impossible request (e.g., two contradictory dose goals) or if it correctly identified a flaw in the prescription and asked for human intervention. This "decision gap" is dangerous; a poorly designed agent might submit a plan that misses critical safety goals simply because the initial request was flawed, rather than the optimization being inadequate. The authors address this by creating OpenGray, a benchmark where the "correct" decision (submit vs. escalate) is known in advance.
Methods
The researchers developed OpenGray by augmenting the existing OpenKBP-Opt head-and-neck dataset with specific "tasks with known answers." These tasks include:
- Routine planning: Standard cases with achievable goals.
- Conflicting notes: Physician instructions that contradict formal dose goals.
- Missing structures: Requests for goals on anatomy not present in the scan.
- Contradictions: Mathematically or anatomically impossible requests (e.g., requiring a high dose in a region where a hard safety cap is set lower).
- Feasible controls: Difficult but proven-achievable tasks to ensure the agent doesn't over-escalate.
They tested four system architectures, ranging from a simple rule-based controller to a full LLM agent, using both a high-end commercial model (Terra, based on GPT-5.6-terra) and a smaller local model (Ministral 3 8B).
Finally, they performed a retrospective audit of SAGE, a clinical LLM agent, to see if these decision-making patterns held true in real-world planning logs.
Datasets
- OpenKBP-Opt: The primary foundation, consisting of 100 head-and-neck cases with pre-computed dose-influence matrices.
- OpenGray: The new benchmark introduced in this paper, which defines specific decision-based tasks on top of the OpenKBP-Opt data.
- SAGE Planning Logs: 39 stereotactic radiosurgery (SRS) runs from a single institution used for the retrospective audit.
Key Insights and Ideas
- Aggregate Scores are Deceptive: Systems can look highly capable on paper while failing catastrophically on specific edge cases. High overall accuracy was often driven by easy tasks, masking a near-total failure to handle complex anatomical contradictions.
- The Spatial Reasoning Bottleneck: LLMs were excellent at catching contradictions written in text (physician notes) but struggled significantly with "overlap contradictions" that required arithmetic reasoning about anatomical volumes and shared voxels.
- The "Action Gap": The audit of the clinical agent SAGE showed that the LLM often correctly identified that a goal was failing, but because it lacked an explicit "stop and escalate" command, it simply kept trying (and failing) until it hit an attempt limit.
- Rule-Based Safeguards Work: Simple arithmetic checks added to the LLM systems were highly effective at catching the contradictions that the LLMs missed, suggesting a hybrid approach is currently necessary for safety.
Key Results
- Terra-based agents achieved high aggregate decision accuracy (up to 95.4%), but this dropped sharply on "overlap contradictions," where they escalated only 25–50% of impossible tasks.
- Ministral 3 8B (the local model) struggled across the board, with decision accuracies between 31.7% and 43.5%, often failing due to technical formatting errors.
- Even on feasible controls (tasks known to be possible), all systems submitted plans that missed hard goals in 24% of cases, showing that optimization strategies still have room for improvement.
Limitations and Future Work
The authors acknowledge several limitations:
- Scope: The benchmark currently uses only one head-and-neck dataset and two types of contradictions.
- Optimization Constraints: The plans were optimized without machine-specific deliverability constraints, which might change what is considered "feasible."
- LLM Selection: Only two LLMs were tested in the reserved evaluation.
Future work involves expanding OpenGray with more diverse contradiction types and "withholding" certain conflict types from developers to prevent systems from being over-tuned to the benchmark. The authors strongly recommend that all future radiotherapy agents be given an explicit "escalate" action and be evaluated on their decisions alongside their dosimetric output.
Summary updated