Are Reasoning Models More Prone to Hallucination?

Recently evolved large reasoning models (LRMs) show powerful performance in solving complex tasks with long chain-of-thought (CoT) reasoning capability. As these LRMs are mostly developed by post-training on formal reasoning tasks, whether they generalize the reasoning capability to help reduce hallucination in fact-seeking tasks remains unclear and debated. For instance, DeepSeek-R1 reports increased performance on SimpleQA, a fact-seeking benchmark, while OpenAI-o3 observes even severer hallucination. This discrepancy naturally raises the following research question: Are reasoning models more prone to hallucination? This paper addresses the question from three perspectives. (1) We first conduct a holistic evaluation for the hallucination in LRMs. Our analysis reveals that LRMs undergo a full post-training pipeline with cold start supervised fine-tuning (SFT) and verifiable reward RL generally alleviate their hallucination. In contrast, both distillation alone and RL training without cold start fine-tuning introduce more nuanced hallucinations. (2) To explore why different post-training pipelines alters the impact on hallucination in LRMs, we conduct behavior analysis. We characterize two critical cognitive behaviors that directly affect the factuality of a LRM: Flaw Repetition, where the surface-level reasoning attempts repeatedly follow the same underlying flawed logic, and Think-Answer Mismatch, where the final answer fails to faithfully match the previous CoT process. (3) Further, we investigate the mechanism behind the hallucination of LRMs from the perspective of model uncertainty. We find that increased hallucination of LRMs is usually associated with the misalignment between model uncertainty and factual accuracy. Our work provides an initial understanding of the hallucination in LRMs.

Humanity's Last Exam

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

How Do Your Code LLMs Perform? Empowering Code Instruction Tuning with High-Quality Data

Recently, there has been a growing interest in studying how to construct better code instruction tuning data. However, we observe Code models trained with these datasets exhibit high performance on HumanEval but perform worse on other benchmarks such as LiveCodeBench. Upon further investigation, we find that many datasets suffer from severe data leakage. After cleaning up most of the leaked data, some well-known high-quality datasets perform poorly. This discovery reveals a new challenge: identifying which dataset genuinely qualify as high-quality code instruction data. To address this, we propose an efficient code data pruning strategy for selecting good samples. Our approach is based on three dimensions: instruction complexity, response quality, and instruction diversity. Based on our selected data, we present XCoder, a family of models finetuned from LLaMA3. Our experiments show XCoder achieves new state-of-the-art performance using fewer training data, which verify the effectiveness of our data strategy. Moreover, we perform a comprehensive analysis on the data composition and find existing code datasets have different characteristics according to their construction methods, which provide new insights for future code LLMs. Our models and dataset are released in https://github.com/banksy23/XCoder

Applications and Advances of Artificial Intelligence in Music Generation:A Review

In recent years, artificial intelligence (AI) has made significant progress in the field of music generation, driving innovation in music creation and applications. This paper provides a systematic review of the latest research advancements in AI music generation, covering key technologies, models, datasets, evaluation methods, and their practical applications across various fields. The main contributions of this review include: (1) presenting a comprehensive summary framework that systematically categorizes and compares different technological approaches, including symbolic generation, audio generation, and hybrid models, helping readers better understand the full spectrum of technologies in the field; (2) offering an extensive survey of current literature, covering emerging topics such as multimodal datasets and emotion expression evaluation, providing a broad reference for related research; (3) conducting a detailed analysis of the practical impact of AI music generation in various application domains, particularly in real-time interaction and interdisciplinary applications, offering new perspectives and insights; (4) summarizing the existing challenges and limitations of music quality evaluation methods and proposing potential future research directions, aiming to promote the standardization and broader adoption of evaluation techniques. Through these innovative summaries and analyses, this paper serves as a comprehensive reference tool for researchers and practitioners in AI music generation, while also outlining future directions for the field.

SwinSF: Image Reconstruction from Spatial-Temporal Spike Streams

The spike camera, with its high temporal resolution, low latency, and high dynamic range, addresses high-speed imaging challenges like motion blur. It captures photons at each pixel independently, creating binary spike streams rich in temporal information but challenging for image reconstruction. Current algorithms, both traditional and deep learning-based, still need to be improved in the utilization of the rich temporal detail and the restoration of the details of the reconstructed image. To overcome this, we introduce Swin Spikeformer (SwinSF), a novel model for dynamic scene reconstruction from spike streams. SwinSF is composed of Spike Feature Extraction, Spatial-Temporal Feature Extraction, and Final Reconstruction Module. It combines shifted window self-attention and proposed temporal spike attention, ensuring a comprehensive feature extraction that encapsulates both spatial and temporal dynamics, leading to a more robust and accurate reconstruction of spike streams. Furthermore, we build a new synthesized dataset for spike image reconstruction which matches the resolution of the latest spike camera, ensuring its relevance and applicability to the latest developments in spike camera imaging. Experimental results demonstrate that the proposed network SwinSF sets a new benchmark, achieving state-of-the-art performance across a series of datasets, including both real-world and synthesized data across various resolutions. Our codes and proposed dataset will be available soon.

Important Attribute Identification in Knowledge Graph

The knowledge graph(KG) composed of entities with their descriptions and attributes, and relationship between entities, is finding more and more application scenarios in various natural language processing tasks. In a typical knowledge graph like Wikidata, entities usually have a large number of attributes, but it is difficult to know which ones are important. The importance of attributes can be a valuable piece of information in various applications spanning from information retrieval to natural language generation. In this paper, we propose a general method of using external user generated text data to evaluate the relative importance of an entity's attributes. To be more specific, we use the word/sub-word embedding techniques to match the external textual data back to entities' attribute name and values and rank the attributes by their matching cohesiveness. To our best knowledge, this is the first work of applying vector based semantic matching to important attribute identification, and our method outperforms the previous traditional methods. We also apply the outcome of the detected important attributes to a language generation task; compared with previous generated text, the new method generates much more customized and informative messages.