Machine unlearning aims to remove the influence of training data without retraining a model from scratch. Evaluation of vision-language models relies on forget accuracy (FA), retain accuracy (RA), membership inference, and similarity to a retain-only reference, but evaluation pipelines are rarely tested. We propose a validity-gated unlearning audit (VGUA) with three checks: the forget set must be memorized without duplicate leakage into the retain set, the scorer must measure correctness, and the training code must implement the intended procedure. Applying these checks reveals that random MMBench splits leak duplicate questions, a lenient scorer hides memorization, and two coding errors remove the signal while training completes. After correcting these problems, we evaluate 234 unlearned models using six methods, three learning rates, and three backbones. Each model is compared with a retain-only reference trained using low-rank adaptation (LoRA). The results depend on update strength. Among models with RA at or above 0.70, lower FA generally indicates closer agreement with the retrained FA-RA operating point. Aggressive updates overshoot this point, collapse RA, and reverse method rankings. Relative to the original model, 78% of healthy models move closer to the retrained model's scores, whereas 94% move farther from it in activation space. Membership-inference advantage does not reliably track retraining alignment. Under option-shuffled training, unlearned models answer 25% of reordered forgotten ScienceQA questions, compared with 6% for the retain-only reference. These results show that multimodal unlearning conclusions depend on pipeline validity, update strength, and the comparison criterion.
Closed-form fractional programming (FP), weighted MMSE (WMMSE), and the 2026 surrogate-enhanced methods SEFP and XMMSE all maximize the weighted sum rate by iteratively maximizing a surrogate linear in the interference power. We show that this family admits a unique best member: the interference-conjugate transform (ICT), which replaces each rate term $\log(1+a/b)$ by $\max_{κ\ge0}[Ω(κa)-κb]$, where $Ω$ is the closed-form partial convex conjugate of the rate in the interference variable and the optimal multiplier is the interference price $a/(b(a+b))$. We prove that, for any fixed price, $Ω(κ\,\cdot)$ is the pointwise largest own-signal function compatible with a valid interference-linear lower bound, and that tightness at an anchor forces the price. Hence all four benchmark surrogates, which share this price and differ only in their own-signal term, are dominated pointwise by ICT. At SINR $γ$, the benchmarks' own-signal curvature exceeds the true curvature by factors behaving as $γ/2$ (FP/WMMSE) and $(\logγ)/2$ (SEFP/XMMSE), against $1+1/(γ(γ+2))$ for ICT, and this ordering carries over to the certified ascent per iteration and the asymptotic linear convergence rate. ICT yields closed-form updates at the benchmarks' per-iteration cost for single-antenna power control, uplink joint scheduling and power control, and multi-cell MISO beamforming, and the first solver-free algorithm for sum-least-$q$th-percentile rate maximization, via a $K$-dimensional convex dual with closed-form gradient and Hessian and a certified duality gap. In uplink scheduling, ICT improves proportional-fairness utility over FP by +7.03 and over SEFP by +3.35 (60/60 drops) using 18.7 iterations per slot vs. 49.7 and 38.0; in MISO beamforming with $M=8$, it reaches 99.9% of its final value in a median of 9 iterations vs. 16 (XMMSE) and 1160 (FP/WMMSE).