Large-scale content moderation requires models that can apply complex, evolving policies under limited expert supervision. We present an empirical study of reinforcement learning (RL) for policy-grounded moderation across three production-derived tasks. Using established RL and inference primitives, we develop three integrity-specific designs: a policy-grounded full-trace reward, rationale-marginalized Monte Carlo scoring, and policy-structured reflection. Each design addresses an observed failure mode, including reasoning-length collapse and polarized confidence estimates. Across the three tasks, RL-Only systems trained on 200–836 labeled examples remain competitive with supervised fine-tuning (SFT) systems trained on thousands of examples, corresponding to up to approximately one order of magnitude greater label efficiency at the evaluated operating points. Performance generally improves with additional data and training compute before saturating, although larger rollout counts exhibit variability. The strongest SFT→RL systems attain point-estimate performance comparable to policy-trained full-time employee (FTE) reviewers across the three tasks.

Multilingual biomedical entity linking (MBEL) aims to map language-specific mentions in the biomedical text to standardized concepts in a multilingual knowledge base (KB) such as Unified Medical Language System (UMLS). In this paper, we propose Con2GEN, a prompt-based controllable contrastive generation framework for MBEL, which summarizes multidimensional information of the UMLS concept mentioned in biomedical text into a natural sentence following a predefined template. Instead of tackling the MBEL problem with a discriminative classifier, we formulate it as a sequence-to-sequence generation task, which better exploits the shared dependencies between source mentions and target entities. Moreover, Con2GEN matches against UMLS concepts in as many languages and types as possible, hence facilitating cross-information disambiguation. Extensive experiments show that our model achieves promising performance improvements compared with several state-of-the-art techniques on the XL-BEL and the Mantra GSC datasets spanning 12 typologically diverse languages.

In community question answering (cQA), the quality of answers are determined by the matching degree between question-answer pairs and the correlation among the answers. In this paper, we show that the dependency between the answer quality labels also plays a pivotal role. To validate the effectiveness of label dependency, we propose two neural network-based models, with different combination modes of Convolutional Neural Net-works, Long Short Term Memory and Conditional Random Fields. Extensive experi-ments are taken on the dataset released by the SemEval-2015 cQA shared task. The first model is a stacked ensemble of the networks. It achieves 58.96% on macro averaged F1, which improves the state-of-the-art neural network-based method by 2.82% and outper-forms the Top-1 system in the shared task by 1.77%. The second is a simple attention-based model whose input is the connection of the question and its corresponding answers. It produces promising results with 58.29% on overall F1 and gains the best performance on the Good and Bad categories.