Few-step flow-map guidance provides an efficient way to steer generation without retraining, but its updates are defined in a continuous or relaxed state space. In categorical text generation, however, the final output is determined by simplex-constrained token distributions and hard decoding. Under low NFE, where each update induces a relatively large finite displacement, this continuous-to-discrete mismatch can make terminal reward guidance unreliable, causing overshoot, reward–quality conflict, saturation, or poor transfer from relaxed reward gains to decoded tokens.

We present Reliability-Calibrated Flow Guidance (RCFG), a sampling-time reliability-calibration framework for terminal reward pullback in few-step categorical flow maps. Rather than replacing the underlying guidance primitive, RCFG calibrates its magnitude and direction according to transition scale, reward saturation, local reward–quality conflict, and categorical endpoint structure. Experiments on LM1B and Text8 show more favorable aggregate control–likelihood operating points after calibration: on LM1B, RCFG maintains or improves verifier Reward and Target while reducing NLL and PPL relative to terminal reward-pullback guidance, while Text8 shows a modest aggregate improvement in lexical-control balance. Code and reproduction instructions are available at: https://github.com/ZionLau/RCFG