让下一篇重要论文，不再错过。

Test-time reinforcement learning (TTRL) adapts pretrained language models to unlabeled target-domain data using self-generated pseudo-labels, whose reliability remains poorly understood. We analyze majority-vote supervision for TTRL with online Direct Preference Optimization (DPO) and decompose its error into population-majority bias and prompt-dependent finite-rollout variation. Because population-majority bias is generally unidentifiable without external supervision, we focus on mitigating the downstream effects of finite-rollout variation. We propose \emph{Prompt-Conditional Distributionally Robust DPO (PC-Dr.DPO)}, which conditions within-prompt robust reweighting on observed vote reliability to make updates from sampling-sensitive prompts more conservative. We further introduce a cache-based mechanism to reduce self-supervised forgetting during online adaptation. Across six mathematical and knowledge-intensive reasoning benchmarks and five model backbones, PC-Dr.DPO matches or outperforms the strongest TTRL baseline in 24 of 30 settings.

A real photograph is often blurry, grainy, compressed, scratched, and low-resolution at the same time. Each defect has its own severity and its own place in the image. The corrections for these defects conflict with each other. Exist- ing models describe the whole image with one description, so they cannot say where each correction belongs. To solve this problem, we present Wonder 4, a mixture-of-experts (MoE) model for image quality enhancement. It corrects several defects of a real photograph at the same time, each in the region where it occurs. Every block of the model holds a mixture of N = 13 experts. A router, trained with the model, sends every image patch to k = 5 of them. No degradation classifier, mask, prompt, detector, or task label is used. Another issue is that standard MoE training, taken from language models, pushes every expert toward equal use. Equal use applies every correction everywhere, even where it is not needed. We replace this rule with a bias that only sets a minimum share of use. Across all patches, an expert used less than this minimum gets higher router scores until it is selected again. An expert above the minimum is not changed, so no expert is forced to be used as often as the others. The bias needs no extra loss. The method adds no extra module and no extra processing step. We evaluate Wonder 4 on DIV2K and Flickr2K, against five published methods and three open-weight pipelines. It reaches the best LPIPS, DISTS, and NIQE on all three test settings (DIV2K x4, DIV2K x8, and Flickr2K x4).