Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites

Detoxifying offensive language while preserving the speaker's original intent is a challenging yet critical goal for improving the quality of online interactions. Although large language models (LLMs) show promise in rewriting toxic content, they often default to overly polite rewrites, distorting the emotional tone and communicative intent. This problem is especially acute in Chinese, where toxicity often arises implicitly through emojis, homophones, or discourse context. We present ToxiRewriteCN, the first Chinese detoxification dataset explicitly designed to preserve sentiment polarity. The dataset comprises 1,556 carefully annotated triplets, each containing a toxic sentence, a sentiment-aligned non-toxic rewrite, and labeled toxic spans. It covers five real-world scenarios: standard expressions, emoji-induced and homophonic toxicity, as well as single-turn and multi-turn dialogues. We evaluate 17 LLMs, including commercial and open-source models with variant architectures, across four dimensions: detoxification accuracy, fluency, content preservation, and sentiment polarity. Results show that while commercial and MoE models perform best overall, all models struggle to balance safety with emotional fidelity in more subtle or context-heavy settings such as emoji, homophone, and dialogue-based inputs. We release ToxiRewriteCN to support future research on controllable, sentiment-aware detoxification for Chinese.

CogSteer: Cognition-Inspired Selective Layer Intervention for Efficient Semantic Steering in Large Language Models

Despite their impressive capabilities, large language models (LLMs) often lack interpretability and can generate toxic content. While using LLMs as foundation models and applying semantic steering methods are widely practiced, we believe that efficient methods should be based on a thorough understanding of LLM behavior. To this end, we propose using eye movement measures to interpret LLM behavior across layers. We find that LLMs exhibit patterns similar to human gaze across layers and different layers function differently. Inspired by these findings, we introduce a heuristic steering layer selection and apply it to layer intervention methods via fine-tuning and inference. Using language toxification and detoxification as test beds, we demonstrate that our proposed CogSteer methods achieve better results in terms of toxicity scores while efficiently saving 97% of the computational resources and 60% of the training time. Our model-agnostic approach can be adopted into various LLMs, contributing to their interpretability and promoting trustworthiness for safe deployment.

Mitigating Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding

Large Vision-Language Models (LVLMs) are increasingly adept at generating contextually detailed and coherent responses from visual inputs. However, their application in multimodal decision-making and open-ended generation is hindered by a notable rate of hallucinations, where generated text inaccurately represents the visual contents. To address this issue, this paper introduces the Instruction Contrastive Decoding (ICD) method, a novel approach designed to reduce hallucinations during LVLM inference. Our method is inspired by our observation that what we call disturbance instructions significantly exacerbate hallucinations in multimodal fusion modules. ICD contrasts distributions from standard and instruction disturbance, thereby increasing alignment uncertainty and effectively subtracting hallucinated concepts from the original distribution. Through comprehensive experiments on discriminative benchmarks (POPE and MME) and a generative benchmark (LLaVa-Bench), we demonstrate that ICD significantly mitigates both object-level and attribute-level hallucinations. Moreover, our method not only addresses hallucinations but also significantly enhances the general perception and recognition capabilities of LVLMs.