Expense reimbursement internal control is a high-stakes corporate finance task that combines multimodal, visually rich evidence (e.g., receipts and approval forms), firm-specific policy text, and multi-step compliance reasoning. No existing benchmark evaluates large multimodal models (LMMs) in this domain. We introduce ReimburseBench, the first benchmark targeting full-workflow multimodal audit reasoning in corporate internal control. It comprises 151 expert-constructed reimbursement workflows, each packaging an application, receipts and supporting documents, and a firm-specific policy, evaluated by a process-level rubric grounded in twelve auditing logic points. Benchmarking nine frontier and locally deployable systems, we find the strongest attains only 72.5%, and no model exceeds 47.7% on hard workflows. Beyond this aggregate score, risk-aware, process-level evaluation surfaces three structural findings: (i) a persistent gap between binary admissibility judgment and precise quantitative computation; (ii) a systematic lenient bias in monetary predictions that poses deployment risk; and (iii) a performance pattern consistent with compositional-generalization difficulties under realistic workflow load. We provide ReimburseBench as a testbed for advancing LMMs toward trustworthy financial systems. Data and code: https://github.com/ikerzhan/ReimburseBench.

Large Language Model (LLM)-based automatic Multi-Agent Systems (MAS) generation has become a crucial frontier for tackling complex tasks. However, existing methods face a dilemma between model capability and experience retention. Inference-time MAS leverages frozen frontier LLMs but repeats identical searches without learning from past experience. Conversely, Training-time MAS internalizes experience via gradient updates but is constrained by the low capability ceiling of smaller models, and is hard to scale to large frontier LLMs. To bridge this gap, we propose Skill-MAS, a novel third path that decouples experience retention from parametric updates by conceptualizing the high-level orchestration capability as an evolvable Meta-Skill. Skill-MAS refines this architectural knowledge through a closed optimization loop: (1) Multi-Trajectory Rollout samples a behavioral distribution for each task under the current Meta-Skill; and (2) Selective Reflection adaptively selects priority tasks and applies hierarchical contrastive analysis to distill systemic experience into generalizable, strategy-level principles. Extensive experiments across four complex benchmarks and four distinct LLMs demonstrate that Skill-MAS not only achieves remarkable performance gains but also maintains a favorable cost-performance trade-off. Further analysis reveals that the evolved Meta-Skills are highly robust and exhibit strong transferability across unseen tasks and different LLMs.

Sarcasm detection is a crucial yet challenging task in natural language processing. Existing methods primarily rely on supervised learning or prompt engineering, which often struggle to capture the complex reasoning process required for effective sarcasm detection. This paper proposes a novel approach that decomposes sarcasm detection into three fundamental dimensions: language, context, and emotion, meticulously modeling the sarcasm reasoning process. To enhance the quality of reasoning, we employ reinforcement learning algorithms and design customized reward models for each dimension. We utilize five widely used sarcasm detection datasets and annotate the sarcasm reasoning process from these three dimensions to improve the performance of the reward models. Experiments demonstrate that our method outperforms state-of-the-art baseline methods in most cases. Additionally, we observe the central role of emotional contrast in sarcasm detection. Our research provides empirical insights into the mechanism of sarcasm, emphasizing that emotional contrast is at its core, supported by linguistic and contextual cues.

This paper describes our system used in the SemEval-2024 Task 9 on two sub-tasks, BRAINTEASER: A Novel Task Defying Common Sense. In this work, we developed a system SHTL, which means simulate human thinking capabilities by Large Language Model (LLM). Our approach bifurcates into two main components: Common Sense Reasoning and Rationalize Defying Common Sense. To mitigate the hallucinations of LLM, we implemented a strategy that combines Retrieval-augmented Generation (RAG) with the the Self-Adaptive In-Context Learning (SAICL), thereby sufficiently leveraging the powerful language ability of LLM. The effectiveness of our method has been validated by its performance on the test set, with an average performance on two subtasks that is 30.1 higher than ChatGPT setting zero-shot and only 0.8 lower than that of humans.

Yonkoma Manga, characterized by its four-panel structure, presents unique challenges due to its rich contextual information and strong sequential features. To address the limitations of current multimodal large language models (MLLMs) in understanding this type of data, we create a novel dataset named YManga from the Internet. After filtering out low-quality content, we collect a dataset of 1,015 yonkoma strips, containing 10,150 human annotations. We then define three challenging tasks for this dataset: panel sequence detection, generation of the author’s creative intention, and description generation for masked panels. These tasks progressively introduce the complexity of understanding and utilizing such image-text data. To the best of our knowledge, YManga is the first dataset specifically designed for yonkoma manga strips understanding. Extensive experiments conducted on this dataset reveal significant challenges faced by current multimodal large language models. Our results show a substantial performance gap between models and humans across all three tasks.

Large language models (LLMs) have revolutionized the domain of natural language processing because of their excellent performance on various tasks. Despite their impressive capabilities, LLMs also have the potential to generate texts that pose risks of misuse. Consequently, detecting LLM-generated text has become increasingly important.Previous LLM-generated text detection methods use semantic features, which are stored in the last layer. This leads to methods that overfit the training set domain and exhibit shortcomings in generalization. Therefore, We argue that utilizing intrinsic features rather than semantic features for detection results in better performance.In this work, we design Text Fluoroscopy, a black-box method with better generalizability for detecting LLM-generated text by mining the intrinsic features of the text to be detected. Our method captures the text’s intrinsic features by identifying the layer with the largest distribution difference from the last and first layers when projected to the vocabulary space.Our method achieves 7.36% and 2.84% average improvement in detection performance compared to the baselines in detecting texts from different domains generated by GPT-4 and Claude3, respectively.

The conversational machine reading comprehension (CMRC) task aims to answer questions in conversations, which has been a hot research topic in recent years because of its wide applications. However, existing CMRC benchmarks in which each conversation is assigned a static passage are inconsistent with real scenarios. Thus, model’s comprehension ability towards real scenarios are hard to evaluate reasonably. To this end, we propose the first Chinese CMRC benchmark Orca and further provide zero-shot/few-shot settings to evaluate model’s generalization ability towards diverse domains. We collect 831 hot-topic driven conversations with 4,742 turns in total. Each turn of a conversation is assigned with a response-related passage, aiming to evaluate model’s comprehension ability more reasonably. The topics of conversations are collected from social media platform and cover 33 domains, trying to be consistent with real scenarios. Importantly, answers in Orca are all well-annotated natural responses rather than the specific spans or short phrase in previous datasets. Besides, we implement three strong baselines to tackle the challenge in Orca. The results indicate the great challenge of our CMRC benchmark.