Recent reinforcement learning (RL) post-training approaches primarily optimize the final output policy using sparse outcome-level rewards, while largely overlooking predictive signals encoded in intermediate representations. In this paper, we introduce a new paradigm called on-policy internal self-distillation and propose the OISD framework, which improves reasoning by transferring on-policy predictive signals from the final layer to intermediate representations. During rollout and Group Relative Policy Optimization (GRPO) optimization, the final layer acts as both the policy and a detached internal teacher for selected intermediate layers, which are guided to align with it through two complementary mechanisms: logit alignment, which transfers high-level reasoning behaviors (how to think), and attention alignment, which enforces consistent attention patterns (where to look) from the final layer to the selected intermediate layer, both without requiring external privileged information. Our OISD, together with GRPO, employs signed advantage-weighted Jensen–Shannon alignment to distill informative intermediate representations while preserving policy consistency under a unified acting policy. Experimental results demonstrate the effectiveness of OISD, with substantial and consistent improvements over strong reasoning RL baselines across four mathematical reasoning tasks.

We present the system description of our team Zhegu in SemEval-2023 Task 9 Multilingual Tweet Intimacy Analysis. We propose \textbf{EPM} (\textbf{E}xponential \textbf{P}enalty \textbf{M}ean Squared Loss) for the purpose of enhancing the ability of learning difficult samples during the training process. Meanwhile, we also apply several methods (frozen Tuning & contrastive learning based on Language) on the XLM-R multilingual language model for fine-tuning and model ensemble. The results in our experiments provide strong faithful evidence of the effectiveness of our methods. Eventually, we achieved a Pearson score of 0.567 on the test set.

SMM4H-2022 (CITATION) Task 2 is to detect whether containing premise in the tweets of users about COVID-19 on the social medias or their stances for the claims. In this paper, we propose Tweet Claim Matching (TCM), which is a new pre-training task constructed by the tweets and claims similarly to Next Sentence Prediction (NSP). We first continue to pre-train the standard pre-trained language models on the labelled dataset and then fine-tune them for obtaining better performance. Compared with the solid baseline (CITATION), we achieve the absolute improvement of 7.9% in Task 2a and obtain the SOTA results.