Object hallucination remains a major challenge in Multimodal Large Language Models (MLLMs), even though they have achieved strong performance across widely vision-language tasks. A key reason for object hallucination is limited fine-grained visual grounding provided by fixed-resolution visual encoding. In this paper, we propose Visual Inference Gating for Object Refinement (VIGOR), a training-free inference-time framework for binary object verification. VIGOR adopts a selective coarse-to-fine pipeline that improve visual grounding without the overhead of always-on multi-view inspection. It first uses the global yes–no margin to estimate whether the available evidence is sufficient. When the evidence is insufficient, VIGOR performs local refinement on uncertain cases using a small set of overlapping crops. Finally, VIGOR combines global and local evidence with an asymmetric decision rule. Experiments on three benchmark datasets across five MLLM backbones show that VIGOR consistently outperforms strong training-free baselines, recovering missed fine-grained evidence while maintaining robust performance on challenging visual reasoning.

Controlling the basic human values expressed by large language models (LLMs) has attracted growing attention in the mechanistic interpretability community. Recent work has made substantial attempts to steer such values through vector arithmetic or causal attribution over internal states. However, because values are complex social constructs, their mechanistic controllability still faces a major unresolved challenge: the auditability bottleneck, which makes it difficult to construct natural-language corpora that precisely capture value realization for mechanistic analysis. To address this challenge, we propose VALUEAUDITOR, a measurement-first framework that makes values behaviorally targetable before they are representationally steered. It operationalizes each value through grouped, theory-grounded response rubrics, derives graded supervision from same-scenario response contrasts, and performs inference-time steering through token-level sparse features that retain local cues of value expression. Experimental results show that VALUEAUDITOR significantly outperforms cutting-edge baseline methods on value steering benchmarks, while demonstrating robust control across values and scenarios with minimal trade-offs in response clarity. The code is publicly available at https://github.com/ZichenX/Steer-On-Rubrics

Recently, the rapid development of multilingual social media platforms (SNS) exacerbates new challenges in SNS content anomaly detection due to data islands and linguistic imbalance. While federated learning (FL) and parameter-efficient fine-tuning (PEFT) offer potential solutions in most cases, when every client is multilingual, existing solutions struggle with multilingual heterogeneity: 1) entangled language-specific knowledge during aggregation, 2) noise from minority languages, and 3) unstable cross-platform collaboration. Based on the asymmetric nature of LoRA, we propose MuLA-F, a multilingual Federated LoRA introducing SVD-based language-specific disentanglement of LoRA blocks and a local orthogonal tuning strategy. Evaluations across three SNS content anomaly detection tasks demonstrate MuLA-F’s superiority in multilingual performance while reducing multilingual knowledge conflicts and communication rounds.

The spread of social misinformation has been a global concern, particularly affecting non-native speaker users who are more susceptible to misinformation on foreign social media platforms. In light of this, this study focuses on mitigating the challenges faced by social misinformation detectors in quickly regaining capability after crossing linguistic borders, especially for non-native users with only monolingual social media histories. By integrating sentiment analysis as an auxiliary, less sensitive task, we transform the challenging cross-lingual transfer into a manageable multi-task framework. Then, we propose HierMoE-Adpt, a novel, cost-effective parameter efficient finetuning method based on hierarchical mixture-of-experts adaptation, to enhance cross-lingual social misinformation detection. HierMoE-Adpt includes a hierarchical routing strategy and an expert-mask mechanism, effectively merge knowledge about the understanding posts in new language and misinformation detection capabilities, contributing to recover the performance of personal misinformation detectors in sync with the dynamics of personal international travel.