YingQiao Lin

Jia Liu, ChangYi He, YingQiao Lin et al.

Recent advancements in Large Language Models have yielded significant improvements in complex reasoning tasks such as mathematics and programming. However, these models remain heavily dependent on annotated data and exhibit limited adaptability in unsupervised scenarios. To address these limitations, test-time reinforcement learning (TTRL) has been proposed, which enables self-optimization by leveraging model-generated pseudo-labels. Despite its promise, TTRL faces several key challenges, including high inference costs due to parallel rollouts and early-stage estimation bias that fosters overconfidence, reducing output diversity and causing performance plateaus. To address these challenges, we introduce an entropy-based mechanism to enhance the exploration-exploitation balance in test-time reinforcement learning through two strategies: Entropy-fork Tree Majority Rollout (ETMR) and Entropy-based Advantage Reshaping (EAR). Compared with the baseline, our approach enables Llama3.1-8B to achieve a 68 percent relative improvement in Pass at 1 metric on the AIME 2024 benchmark, while consuming only 60 percent of the rollout tokens budget. This highlights our method's ability to effectively optimize the trade-off between inference efficiency, diversity, and estimation robustness, thereby advancing unsupervised reinforcement learning for open-domain reasoning tasks.

Cheng Dai, Yingqiao Lin, Fan Li et al.

In Domain Generalization (DG) tasks, models are trained by using only training data from the source domains to achieve generalization on an unseen target domain, this will suffer from the distribution shift problem. So it's important to learn a classifier to focus on the common representation which can be used to classify on multi-domains, so that this classifier can achieve a high performance on an unseen target domain as well. With the success of cross attention in various cross-modal tasks, we find that cross attention is a powerful mechanism to align the features come from different distributions. So we design a model named CADG (cross attention for domain generalization), wherein cross attention plays a important role, to address distribution shift problem. Such design makes the classifier can be adopted on multi-domains, so the classifier will generalize well on an unseen domain. Experiments show that our proposed method achieves state-of-the-art performance on a variety of domain generalization benchmarks compared with other single model and can even achieve a better performance than some ensemble-based methods.