Selective Refusal as Protective Calibration

F-01 · measurementVLM calibration on robot perception frames, measured on Modal A100.

Question
Which VLM should a safety-critical robot trust when it is uncertain?
Method
Six-pipeline benchmark, Qwen2.5-VL 3B, 7B and 72B plus InternVL2-2B, on Modal A100s over robot perception frames.
Result
Failure modes are model-family-specific: one family refuses under uncertainty, another confabulates on identical inputs. The refusing family reaches ~2× precision and 5–10× recall on engaged frames.
Implication
Refusal is calibration, not weakness. Pick model families by failure mode, not leaderboard rank.
precision ~2× recall 5–10×
Precision and recall on engaged frames. Grey is the confabulating family, blue the refusing family.