Text detoxification is increasingly used in content moderation and dialogue systems, where systems must produce safe rewrites while revealing what was changed and why. Existing approaches typically trade off interpretability and end-to-end optimization. Monolithic rewriting methods treat detoxification as a single opaque generation step, whereas edit-based pipelines make edits explicit but are trained in separate stages, preventing joint optimization. We propose STEER, a framework that models detoxification as a structured edit chain formulation and optimizes the entire editing process end-to-end with Group Relative Policy Optimization (GRPO). To make this formulation trainable, we construct synthetic edit-chain supervision using a multi-agent pipeline, perform supervised fine-tuning (SFT) to teach the model the structured editing format, and then apply GRPO with a veto-aligned multiplicative reward and granular format scoring. Experiments on ParaDetox and two out-of-domain benchmarks show that STEER achieves the highest aggregate J score among the compared systems on ParaDetox and remains competitive under out-of-domain evaluation.

Text detoxification should produce safer rewrites while retaining identity mentions from the original text. However, classifier-guided methods can remove identity descriptors alongside genuinely harmful content when toxicity classifiers associate identity mentions with harmfulness and guidance relies on absolute scores. We propose CIDER (Contrastive Identity-preserving Detoxification via Explicit Reference), a default-relative, margin-gated selective-guidance method for masked diffusion language models. At each selected masked position, CIDER evaluates every candidate in the admissible top-k set relative to the model-default token in the same partially decoded state and commits a non-default candidate only when the best candidate lowers the one-step proxy score by more than a preset margin. On the primary LLaDA-8B CivilComments×Jigsaw setting, CIDER improves retention by +0.06 to +0.16 over the interpolated absolute-score frontier at matched residual-toxicity levels. At the reported aggressive endpoints, the retention advantage recurs across four data pairings on LLaDA-8B and three bases on CivilComments×Jigsaw. Controlled ablations indicate that the selection rule is an important contributor to the improvement.

Hate speech classifiers often misclassify counter speech, including condemnation and quotation of hateful content, as hateful because they fail to distinguish propositional content from speaker stance. We identify this issue as content stance conflation and address it as a representation learning problem. We learn stance invariant representations of the same proposition by constructing contrastive pairs that share content but differ in stance. We implement this idea through factorized annotation, stance controlled augmentation, and Stance Contrastive Learning (SCL). On the Implicit Hate Corpus, stance variant augmentation reduces counter speech false positives by 60%, and SCL further improves cross domain F1 by up to 6.5 points on top of the augmented signal. On SBIC, where use mention conflation is less prominent in the training labels, SCL maintains a low counter speech false positive rate. Semantic similarity pairing on the same augmented data yields no cross domain gain over the model trained without contrastive learning (0.623 vs. 0.625), which identifies the pair construction signal rather than contrastive learning itself as the determining factor.

With the explosive growth of short-video data on industrial video-sharing platforms such as TikTok and YouTube, text-video retrieval techniques have become increasingly important. Most existing works for text-video retrieval focus on designing informative representation learning methods and delicate matching mechanisms, which leverage the content information of queries and videos themselves (i.e., textual information of queries and multimodal information of videos). However, real-world scenarios often involve brief, ambiguous queries and low-quality videos, making content-based retrieval less effective. In order to accommodate various search requirements and enhance user satisfaction, this study introduces a novel Text-video Retrieval method via Watch-time-aware Heterogeneous Graph Contrastive Learning (termed ORANGE). This approach aims to learn informative embeddings for queries and videos by leveraging both content information and the abundant relational information present in video-search scenarios. Specifically, we first construct a heterogeneous information graph where nodes represent domain objects (e.g., query, video, tag) and edges represent rich relations among these objects. Afterwards, a meta-path-guided heterogeneous graph attention encoder with the awareness of video watch time is devised to encode various semantic aspects of query and video nodes. To train our model, we introduce a meta-path-wise contrastive learning paradigm that facilitates capturing dependencies across multiple semantic relations, thereby enhancing the obtained embeddings. Finally, when deployed online, for new queries non-existent in the constructed graph, a bert-based query encoder distilled from our ORANGE is employed. Offline experiments conducted on a real-world dataset demonstrate the effectiveness of our ORANGE. Moreover, it has been implemented in the matching stage of an industrial online video-search service, where it exhibited statistically significant improvements over the online baseline in an A/B test.