TraceGain: Marginal Answer Utility for Document VLM Reveal-or-Stop Control

At a visited document state, stopping and acquiring another high-resolution region are mutually exclusive actions, yet region saliency and answer confidence score them in incompatible units. TraceGain turns this choice into a measurable control problem by assigning every legal zoom and shift its realized one-step change in official answer score and placing stop at the exact zero of that scale. Exhaustive counterfactual expansion of student-visited states supplies the targets; a centered dense teacher shapes only the offline regression target, and a 1.35M-parameter head batches all legal actions at inference. On DocVQA, InfographicVQA, and ChartQA, raw predictions attain Spearman $\rho =0.41$–$0.47$, Utility ECE $=0.051$–$0.062$, and one-step stop AUPRC $=0.78$–$0.84$. Against DirectPolicy matched for inputs, capacity, teacher privilege, roll-ins, and budget, fixed-cache point estimates show 57% more stop errors and 65% more regret for direct classification. Matched endpoint gains span 0.3–1.1 points with LLaVA and 0.3–1.0 with Qwen2.5-VL; in the DocVQA operating study, TraceGain reaches 81.7 ANLS at 968 reported pre-encoding tokens and 438 ms, compared with 82.4 at 2,016 tokens and 847 ms for Full-Res. A common marginal-utility coordinate thus connects action selection, stopping, calibration, and compute allocation for short-horizon document inference.