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.

This paper focuses on dealing with data scarcity in spoken question answering (QA) using automatic question-answer generation and a carefully selected fine-tuning strategy that leverages limited annotated data (paragraphs and question-answer pairs). Spoken QA is a challenging task due to using spoken documents, i.e., erroneous automatic speech recognition (ASR) transcriptions, and the scarcity of spoken QA data. We propose a framework for utilizing limited annotated data effectively to improve spoken QA performance. To deal with data scarcity, we train a question-answer generation model with annotated data and then produce large amounts of question-answer pairs from unannotated data (paragraphs). Our experiments demonstrate that incorporating limited annotated data and the automatically generated data through a carefully selected fine-tuning strategy leads to 5.5% relative F1 gain over the model trained only with annotated data. Moreover, the proposed framework is also effective in high ASR errors.

This paper describes a framework to automatically generate a spoken question answering (QA) dataset. The framework consists of a question generation (QG) module to generate questions automatically from given text documents, a text-to-speech (TTS) module to convert the text documents into spoken form and an automatic speech recognition (ASR) module to transcribe the spoken content. The final dataset contains question-answer pairs for both the reference text and ASR transcriptions as well as the audio files corresponding to each reference text. For QG and ASR systems we used pre-trained multilingual encoder-decoder transformer models and fine-tuned these models using a limited amount of manually generated QA data and TTS-based speech data, respectively. As a proof of concept, we investigated the proposed framework for Turkish and generated the Turkish Question Answering (TurQuAse) dataset using Wikipedia articles. Manual evaluation of the automatically generated question- answer pairs and QA performance evaluation with state of-the-art models on TurQuAse show that the proposed framework is efficient for automatically generating spoken QA datasets. To the best of our knowledge, TurQuAse is the first publicly available spoken question answering dataset for Turkish. The proposed framework can be easily extended to other languages where a limited amount of QA data is available.