Timbre
Timbre is a reliability-aware harmonic consensus framework for multimodal test-time reinforcement learning. It filters pseudo-labels through a 2x2 view grid and weighted harmonic aggregation, and adaptively adjusts modality weights using confidence drift, mitigating the reinforcement of systematic errors by majority voting.
Timeline (last 90 days)
Timbre is a reliability-aware harmonic consensus framework for multi-view aggregation in multimodal test-time reinforcement learning.
Timbre constructs a 2×2 grid of views: the original image and the perturbed image are each paired with two text modes, direct solving and reformulation.
Timbre uses weighted harmonic aggregation to select pseudo-labels rather than simple averaging.
The consensus problem is formalized as a weighted χ² barycenter problem, whose unique minimum is exactly the normalized weighted harmonic aggregation, and it does not require the "correct-answer invariance" assumption.
Timbre uses the batch-averaged, time-smoothed change in pseudo-label confidence as a heuristic reliability signal, lowering the weight of a modality when its confidence systematically drifts.
Timbre's weight update rule has no theoretical optimality guarantee, and the authors position it as a heuristic method.
On six benchmarks and three Qwen3-VL model scales, under matched numbers of estimation and optimization rollouts, Timbre outperforms TTRV on 16 of 18 benchmark-model settings on adaptation instances and on 17 on held-out instances.
Timbre's experiments are concentrated on the Qwen3-VL series of models, and generalization across model families has not been verified.
Timbre's reliability signal relies on batch statistics, and batch size and distribution shift may affect the stability of weight updates.
All Facts (9)
Timbre is a reliability-aware harmonic consensus framework for multi-view aggregation in multimodal test-time reinforcement learning.
50%UnverifiedTimbre constructs a 2×2 grid of views: the original image and the perturbed image are each paired with two text modes, direct solving and reformulation.
50%UnverifiedTimbre uses weighted harmonic aggregation to select pseudo-labels rather than simple averaging.
50%UnverifiedThe consensus problem is formalized as a weighted χ² barycenter problem, whose unique minimum is exactly the normalized weighted harmonic aggregation, and it does not require the "correct-answer invariance" assumption.
50%UnverifiedTimbre uses the batch-averaged, time-smoothed change in pseudo-label confidence as a heuristic reliability signal, lowering the weight of a modality when its confidence systematically drifts.
50%UnverifiedTimbre's weight update rule has no theoretical optimality guarantee, and the authors position it as a heuristic method.
50%UnverifiedOn six benchmarks and three Qwen3-VL model scales, under matched numbers of estimation and optimization rollouts, Timbre outperforms TTRV on 16 of 18 benchmark-model settings on adaptation instances and on 17 on held-out instances.
50%UnverifiedTimbre's experiments are concentrated on the Qwen3-VL series of models, and generalization across model families has not been verified.
50%UnverifiedTimbre's reliability signal relies on batch statistics, and batch size and distribution shift may affect the stability of weight updates.
50%