Kaixuan Ren

Yihao Ding, Kaixuan Ren, Jiabin Huang et al.

Document Question Answering (QA) presents a challenge in understanding visually-rich documents (VRD), particularly those dominated by lengthy textual content like research journal articles. Existing studies primarily focus on real-world documents with sparse text, while challenges persist in comprehending the hierarchical semantic relations among multiple pages to locate multimodal components. To address this gap, we propose PDF-MVQA, which is tailored for research journal articles, encompassing multiple pages and multimodal information retrieval. Unlike traditional machine reading comprehension (MRC) tasks, our approach aims to retrieve entire paragraphs containing answers or visually rich document entities like tables and figures. Our contributions include the introduction of a comprehensive PDF Document VQA dataset, allowing the examination of semantically hierarchical layout structures in text-dominant documents. We also present new VRD-QA frameworks designed to grasp textual contents and relations among document layouts simultaneously, extending page-level understanding to the entire multi-page document. Through this work, we aim to enhance the capabilities of existing vision-and-language models in handling challenges posed by text-dominant documents in VRD-QA.

Yiran Zhang, Mo Wang, Xiaoyang Li et al.

Despite impressive advances in large language models (LLMs), existing benchmarks often focus on single-turn or single-step tasks, failing to capture the kind of iterative reasoning required in real-world settings. To address this limitation, we introduce TurnBench, a novel benchmark that evaluates multi-turn, multi-step reasoning through an interactive code-breaking task inspired by a "Turing Machine Board Game." In each episode, a model must uncover hidden logical or arithmetic rules by making sequential guesses, receiving structured feedback, and integrating clues across multiple rounds. This dynamic setup requires models to reason over time, adapt based on past information, and maintain consistency across steps-capabilities underexplored in current benchmarks. TurnBench includes two modes: Classic, which tests standard reasoning, and Nightmare, which introduces increased complexity and requires robust inferential chains. To support fine-grained analysis, we provide ground-truth annotations for intermediate reasoning steps. Our evaluation of state-of-the-art LLMs reveals significant gaps: the best model achieves 81.5% accuracy in Classic mode, but performance drops to 17.8% in Nightmare mode. In contrast, human participants achieve 100% in both, underscoring the challenge TurnBench poses to current models. By incorporating feedback loops and hiding task rules, TurnBench reduces contamination risks and provides a rigorous testbed for diagnosing and advancing multi-step, multi-turn reasoning in LLMs.

Kaixuan Ren, Preslav Nakov, Usman Naseem

As vision-language models become increasingly capable, maintaining a balance between safety and usefulness remains a central challenge. Safety mechanisms, while essential, can backfire, causing over-refusal, where models decline benign requests out of excessive caution. Yet, no existing benchmark has systematically addressed over-refusal in the visual modality. This setting introduces unique challenges, such as dual-use cases where an instruction is harmless, but the accompanying image contains harmful content. Models frequently fail in such scenarios, either refusing too conservatively or completing tasks unsafely, which highlights the need for more fine-grained alignment. The ideal behavior is safe completion, i.e., fulfilling the benign parts of a request while explicitly warning about any potentially harmful elements. To address this, we present DUAL-Bench, the first multimodal benchmark focused on over-refusal and safe completion in VLMs. We evaluated 18 VLMs across 12 hazard categories, with focus on their robustness under semantics-preserving visual perturbations. The results reveal substantial room for improvement: GPT-5-Nano achieves 12.9% safe completion, GPT-5 models average 7.9%, and Qwen models only 3.9%. We hope that DUAL-Bench will foster the development of more nuanced alignment strategies that ensure models remain both safe and useful in complex multimodal settings.