Step-width modulation contributes to mediolateral balance during walking, and soft exosuits are promising for providing step-width guidance with limited interference to natural movement. In a previous study, a human-in-the-loop (HIL) framework based on Bayesian optimization (BO) was proposed for step-width guidance using a soft pneumatic exosuit driven by pneumatic artificial muscles (PAM-SPE). However, its effectiveness was evaluated only in a limited user study (N = 2), with analysis restricted to the single factor of assistance condition. In this study, we further evaluated PAM-SPE-based step-width guidance with more participants (N = 7) using a two-factor analysis of assistance condition and dynamically varying reference step width (mean ± SD: 0.251 ± 0.061 m). The assistance conditions were initial assistance using the initially obtained parameters, optimized assistance using the parameters obtained via BO, and the unassisted condition. Both assistance conditions significantly outperformed the unassisted condition, while no significant difference was found between the initial and optimized conditions. Post-hoc analysis suggested that assistance was effective for medium-to-large reference step widths, but not for small ones, likely because of physical interference between the PAM-SPE and participants’ legs. The BO process in our hip abduction/adduction assistance system also varied across participants, consistent with previous findings on participant-dependent HIL optimization.