Partial unemployment insurance enables claimants to keep part of their unemployment benefits when they work in low-earnings jobs. The reduction in current benefits leads to an increase in future benefits, so that forward-looking claimants are taxed according to a lower dynamic marginal tax rate than the static benefit-reduction rate. I develop a dynamic model of working claimants and use bunching in U.S. data at kinks of the benefit-reduction schedule to estimate the earnings elasticity to the net-of-tax rate. Consistent with forward-looking behaviors, I document that claimants expecting short spells bunch more. Counterfactual simulations suggest lowering the U.S. benefit reduction rate.