Songbo Yang

Geng Liu, Li Feng, Carlo Alberto Bono et al.

Recent research has highlighted that assigning specific personas to large language models (LLMs) can significantly increase harmful content generation. However, limited attention has been given to persona-driven toxicity in non-Western contexts, particularly in Chinese-based LLMs. In this paper, we perform a large-scale, cross-model analysis of refusal behavior and persona-driven toxicity amplification across four Chinese LLMs, leveraging a comprehensive dataset of over 1,400,000 generated texts. We identify significant disparities in persona-driven refusal behavior, including systematic gender differences in refusal triggering across the evaluated Chinese LLMs. Furthermore, we provide quantitative evidence of persona-driven toxicity amplification with respect to model default baselines. We show that this amplification--whose magnitude varies substantially across models--is driven by interactions across several factors, involving persona conditioning, prompting strategy, target social group, and model-specific safety mechanisms. Leveraging model-specific regression analyses, we systematically characterize how persona categories, target social groups, and prompt templates independently and jointly shape both refusal behavior and output toxicity. As a complementary case study, we further explore an iterative, evaluator-guided mitigation strategy based on model feedback with an external LLM evaluator, demonstrating that highly toxic outputs can be substantially reduced without costly model retraining. Overall, our findings highlight the importance of culturally contextualized safety evaluations for Chinese-language LLMs and provide a structured framework for assessing persona-induced risks and exploratory mitigation strategies in LLM-generated content.

Sihang Jia, Shuliang Liu, Songbo Yang et al.

Multimodal Large Language Models frequently suffer from inference hallucinations, partially stemming from language priors dominating visual evidence. Existing training-free mitigation methods either perturb the visual representation and deviate from the natural image distribution, or enforce intrusive manipulations that compromise the model's inherent generative fluency. We introduce a novel perspective that multimodal hallucination manifests as the hypersensitivity of visual grounding to textual phrasing during the decoding phase. Building on this insight, we propose Decoding by Perturbation (DeP), a training-free framework mitigating prior-induced hallucinations via controlled textual interventions. DeP employs a dynamic probe applying multi-level textual perturbations to elicit latent language priors. Leveraging attention variance, it enhances stable evidence regions while suppressing suspicious noise in the feature space. Furthermore, it constructs an interpretable prior drift direction using logits statistics to counteract probability biases from textual co-occurrences. Extensive experiments confirm DeP effectively reduces hallucinations and achieves superior performance across multiple benchmarks.

Shuliang Liu, Songbo Yang, Dong Fang et al.

Object hallucination critically undermines the reliability of Multimodal Large Language Models, often stemming from a fundamental failure in cognitive introspection, where models blindly trust linguistic priors over specific visual evidence. Existing mitigations remain limited: contrastive decoding approaches operate superficially without rectifying internal semantic misalignments, while current latent steering methods rely on static vectors that lack instance-specific precision. We introduce Vision-Language Introspection (VLI), a training-free inference framework that simulates a metacognitive self-correction process. VLI first performs Attributive Introspection to diagnose hallucination risks via probabilistic conflict detection and localize the causal visual anchors. It then employs Interpretable Bi-Causal Steering to actively modulate the inference process, dynamically isolating visual evidence from background noise while neutralizing blind confidence through adaptive calibration. VLI achieves state-of-the-art performance on advanced models, reducing object hallucination rates by 12.67% on MMHal-Bench and improving accuracy by 5.8% on POPE.