This paper introduces Timbre, a reliability‑aware harmonic consensus framework for label‑free multimodal test‑time reinforcement learning. The method first builds a $2\times2$ grid that combines original and perturbed images with outputs from direct‑solving and reframing modes. Pseudo‑labels are then selected via weighted harmonic aggregation, while batch‑averaged, temporally smoothed shifts in pseudo‑label confidence serve as a heuristic reliability signal to adapt modality‑level weights.
The consensus is formulated as a weighted $\chi^2$‑barycenter problem; its unique minimizer is the normalized weighted harmonic aggregate, eliminating the need for correct‑answer invariance. A support‑dispersion decomposition characterizes when harmonic aggregation preserves or corrects arithmetic rankings, and a sensitivity analysis shows how modality reweighting alters pairwise ranking margins, offering interpretability for the heuristic updates.
Across six benchmarks and three scales of the Qwen3‑VL model, Timbre outperforms TTRV in 16 of 18 benchmark‑model settings on adaptation instances and in 17 of 18 settings on held‑out instances, using matched estimation and optimization rollout counts. The results demonstrate improved target‑task adaptation and effective transfer to unseen instances without further updates.
Review: Timbre mitigates the risk of misleading majority voting by leveraging reliability‑driven harmonic aggregation, delivering a theoretically grounded and empirically superior solution for multimodal reinforcement learning.