Shudong Liu

Zihao Zhou, Shudong Liu, Maizhen Ning et al.

Exceptional mathematical reasoning ability is one of the key features that demonstrate the power of large language models (LLMs). How to comprehensively define and evaluate the mathematical abilities of LLMs, and even reflect the user experience in real-world scenarios, has emerged as a critical issue. Current benchmarks predominantly concentrate on problem-solving capabilities, presenting a substantial risk of model overfitting and fails to accurately measure the genuine mathematical reasoning abilities. In this paper, we argue that if a model really understands a problem, it should be robustly applied across a diverse array of tasks. To this end, we introduce MathCheck, a well-designed checklist for testing task generalization and reasoning robustness, as well as an automatic tool to generate checklists efficiently. MathCheck includes multiple mathematical reasoning tasks and robustness tests to facilitate a comprehensive evaluation of both mathematical reasoning ability and behavior testing. Utilizing MathCheck, we develop MathCheck-GSM and MathCheck-GEO to assess math textual reasoning and multi-modal reasoning abilities, respectively, serving as upgraded versions of benchmarks including GSM8k, GeoQA, UniGeo, and Geometry3K. We adopt MathCheck-GSM and MathCheck-GEO to evaluate 26 LLMs and 17 MLLMs. Our results demonstrate that while frontier LLMs like GPT-4o continue to excel in various abilities on the checklist, many other model families exhibit a significant decline. Further experiments indicate that, compared to traditional math benchmarks, MathCheck better reflects true mathematical abilities and represents mathematical intelligence more linearly, thereby supporting our design. Using MathCheck, we can efficiently conduct informative behavior analysis to deeply investigate models. Finally, we show that our checklist paradigm can easily extend to other reasoning tasks.

Zihan Ma, Dongsheng Zhu, Shudong Liu et al.

Current safety evaluations for LLM-driven agents primarily focus on atomic harms, failing to address sophisticated threats where malicious intent is concealed or diluted within complex tasks. We address this gap with a two-dimensional analysis of agent safety brittleness under the orthogonal pressures of intent concealment and task complexity. To enable this, we introduce OASIS (Orthogonal Agent Safety Inquiry Suite), a hierarchical benchmark with fine-grained annotations and a high-fidelity simulation sandbox. Our findings reveal two critical phenomena: safety alignment degrades sharply and predictably as intent becomes obscured, and a "Complexity Paradox" emerges, where agents seem safer on harder tasks only due to capability limitations. By releasing OASIS and its simulation environment, we provide a principled foundation for probing and strengthening agent safety in these overlooked dimensions.

Dingling Zhang, He Zhu, Jincheng Ren et al.

Deep Research Agents (DRAs) aim to automatically produce analyst-level reports through iterative information retrieval and synthesis. However, most existing DRAs were validated on question-answering benchmarks, while research on generating comprehensive reports remains overlooked. Worse, current benchmarks for report synthesis suffer from task complexity and subjective metrics -- this fails to reflect user demands and limits the practical utility of generated reports. To address these gaps, we present Fine-grained DEepResearch bench (FINDER), an enhanced benchmark consisting of 100 human-curated research tasks with 419 structured checklist items that standardize report structure, analytical depth, and factual grounding. Based on approximately 1,000 reports produced by mainstream DRAs, we further propose Deep rEsearch Failure Taxonomy (DEFT), the first failure taxonomy for deep research agents. DEFT contains 14 fine-grained failure modes across reasoning, retrieval, and generation, and is built upon grounded theory with human-LLM co-annotating and inter-annotator reliability validation. Our experimental findings reveal that current DRAs struggle not with task comprehension but with evidence integration, verification, and reasoning-resilient planning.

Shudong Liu, Yiqiao Jin, Cheng Li et al.

Vision-language models (VLMs) have advanced human-AI interaction but struggle with cultural understanding, often misinterpreting symbols, gestures, and artifacts due to biases in predominantly Western-centric training data. In this paper, we construct CultureVerse, a large-scale multimodal benchmark covering 19, 682 cultural concepts, 188 countries/regions, 15 cultural concepts, and 3 question types, with the aim of characterizing and improving VLMs' multicultural understanding capabilities. Then, we propose CultureVLM, a series of VLMs fine-tuned on our dataset to achieve significant performance improvement in cultural understanding. Our evaluation of 16 models reveals significant disparities, with a stronger performance in Western concepts and weaker results in African and Asian contexts. Fine-tuning on our CultureVerse enhances cultural perception, demonstrating cross-cultural, cross-continent, and cross-dataset generalization without sacrificing performance on models' general VLM benchmarks. We further present insights on cultural generalization and forgetting. We hope that this work could lay the foundation for more equitable and culturally aware multimodal AI systems.

Tingyuan Zhu, Shudong Liu, Yidong Wang et al. · pku

Constructing high-quality Supervised Fine-Tuning (SFT) datasets is critical for the training of large language models (LLMs). Recent studies have shown that using data from a specific source, Ruozhiba, a Chinese website where users ask "silly" questions to better understand certain topics, can lead to better fine-tuning performance. This paper aims to explore some hidden factors: the potential interpretations of its success and a large-scale evaluation of the performance. First, we leverage GPT-4 to analyze the successful cases of Ruozhiba questions from the perspective of education, psychology, and cognitive science, deriving a set of explanatory rules. Then, we construct fine-tuning datasets by applying these rules to the MMLU training set. Surprisingly, our results indicate that rules can significantly improve model performance in certain tasks, while potentially diminishing performance on others. For example, SFT data generated following the "Counterintuitive Thinking" rule can achieve approximately a 5% improvement on the "Global Facts" task, whereas the "Blurring the Conceptual Boundaries" rule leads to a performance drop of 6.14% on the "Econometrics" task. In addition, for specific tasks, different rules tend to have a consistent impact on model performance. This suggests that the differences between the extracted rules are not as significant, and the effectiveness of the rules is relatively consistent across tasks. Our research highlights the importance of considering task diversity and rule applicability when constructing SFT datasets to achieve more comprehensive performance improvements.

Chi Seng Cheang, Hou Pong Chan, Derek F. Wong et al.

Recent pre-trained language models (PLMs) achieve promising results in existing abstractive summarization datasets. However, existing summarization benchmarks overlap in time with the standard pre-training corpora and finetuning datasets. Hence, the strong performance of PLMs may rely on the parametric knowledge that is memorized during pre-training and fine-tuning. Moreover, the knowledge memorized by PLMs may quickly become outdated, which affects the generalization performance of PLMs on future data. In this work, we propose TempoSum, a novel benchmark that contains data samples from 2010 to 2022, to understand the temporal generalization ability of abstractive summarization models. Through extensive human evaluation, we show that parametric knowledge stored in summarization models significantly affects the faithfulness of the generated summaries on future data. Moreover, existing faithfulness enhancement methods cannot reliably improve the faithfulness of summarization models on future data. Finally, we discuss several recommendations to the research community on how to evaluate and improve the temporal generalization capability of text summarization models.