On 20 October 2026, researchers posted a preprint on arXiv introducing SAGE-CP, a calibration method that guarantees prediction-set coverage within each computational route of adaptive vision-language model ensembles.
Adaptive ensembles reduce inference cost by routing easy inputs to early exits, but a single global conformal threshold can hide severe failures on individual routes. The preprint, which has not been peer reviewed, shows that marginal coverage allows a route’s conditional coverage to drop as low as max{0,1−α/p_i}—a gap SAGE-CP is designed to close.
Automatica reported earlier in October on work that derived oracle inequalities for mixture-of-experts models, quantifying the statistical cost of adaptive routing. The new preprint addresses a complementary problem: after the router is frozen on development data, it applies Mondrian split conformal prediction separately to early-exit and escalated groups, giving finite-sample coverage within each operational route.
Across 10 benchmarks, four frozen VLMs and five aligned splits, both global calibration and SAGE-CP achieved 0.900 marginal coverage at a 0.90 target. But SAGE-CP raised mean worst-route coverage from 0.623 to 0.887 and reduced the mean route-coverage gap from 0.324 to 0.025, using the same routed scores and 1.177 source calls per image. The stronger guarantee increased mean prediction-set size from 1.893 to 2.212, though sets remained smaller than those of the full ensemble (2.486) while requiring 70.6 percent fewer source calls.
The preprint also reports efficiency gains relative to full inference: batch-one latency fell by 58.8 to 63.4 percent and gross GPU energy by 55.7 to 57.5 percent on five datasets. Finer exit-depth groups, however, showed that rare routes need additional calibration data to maintain coverage.
The paper runs to 19 pages with two figures and 12 tables.