Large language models (LLMs) deliver strong performance but are difficult to deploy due to high memory and compute costs. While pruning reduces these demands, most methods ignore activation sparsity observed at runtime. We reinterpret activation sparsity as dynamic structured weight sparsity and propose DuoGPT, a unified framework that constructs dual-sparse (spMspV) workloads by combining unstructured weight pruning with activation sparsity. To preserve accuracy, we extend the Optimal Brain Compression (OBC) framework with activation-aware calibration and introduce output residuals from the dense model as correction terms. We further optimize the solution for efficient GPU execution, enabling scalability to billion-parameter LLMs. Evaluations on LLaMA-2 and LLaMA-3 show that DuoGPT outperforms state-of-the-art structured pruning methods by up to 9.17% accuracy at an iso-speedup of 1.39$\times$ compared to the baseline dense model. Code is available at Github.

Large Language Models (LLMs) hold promise in automating data analysis tasks, yet open-source models face significant limitations in these kinds of reasoning-intensive scenarios. In this work, we investigate strategies to enhance the data analysis capabilities of open-source LLMs. By curating a seed dataset of diverse, realistic scenarios, we evaluate model behavior across three core dimensions: data understanding, code generation, and strategic planning. Our analysis reveals three key findings: (1) Strategic planning quality serves as the primary determinant of model performance; (2) Interaction design and task complexity significantly influence reasoning capabilities; (3) Data quality demonstrates a greater impact than diversity in achieving optimal performance. We leverage these insights to develop a data synthesis methodology, demonstrating significant improvements in open-source LLMs' analytical reasoning capabilities. Code is available at https://github.com/zjunlp/DataMind.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs), while its mechanisms are not yet well understood. In this work, we undertake a pioneering exploration of RLVR through the novel perspective of token entropy patterns, comprehensively analyzing how different tokens influence reasoning performance. By examining token entropy patterns in Chain-of-Thought (CoT) reasoning, we observe that only a small fraction of tokens exhibit high entropy, and these tokens act as critical forks that steer the model toward diverse reasoning pathways. Furthermore, studying how entropy patterns evolve during RLVR training reveals that RLVR largely adheres to the base model's entropy patterns, primarily adjusting the entropy of high-entropy tokens. These findings highlight the significance of high-entropy tokens (i.e., forking tokens) to RLVR. We ultimately improve RLVR by restricting policy gradient updates to forking tokens and uncover a finding even beyond the 80/20 rule: utilizing only 20% of the tokens while maintaining performance comparable to full-gradient updates on the Qwen3-8B base model and significantly surpassing full-gradient updates on the Qwen3-32B (+11.04 on AIME'25 and +7.71 on AIME'24) and Qwen3-14B (+4.79 on AIME'25 and +5.21 on AIME'24) base models, highlighting a strong scaling trend. In contrast, training exclusively on the 80% lowest-entropy tokens leads to a marked decline in performance. These findings indicate that the efficacy of RLVR primarily arises from optimizing the high-entropy tokens that decide reasoning directions. Collectively, our results highlight the potential to understand RLVR through a token-entropy perspective and optimize RLVR by leveraging high-entropy minority tokens to further improve LLM reasoning.

Vision-Language Models (VLMs) have shown remarkable progress in visual understanding in recent years. Yet, they still lag behind human capabilities in specific visual tasks such as counting or relational reasoning. To understand the underlying limitations, we adopt methodologies from cognitive science, analyzing VLM performance along core cognitive axes: Perception, Attention, and Memory. Using a suite of tasks targeting these abilities, we evaluate state-of-the-art VLMs, including GPT-4o. Our analysis reveals distinct cognitive profiles: while advanced models approach ceiling performance on some tasks (e.g. category identification), a significant gap persists, particularly in tasks requiring spatial understanding or selective attention. Investigating the source of these failures and potential methods for improvement, we employ a vision-text decoupling analysis, finding that models struggling with direct visual reasoning show marked improvement when reasoning over their own generated text captions. These experiments reveal a strong need for improved VLM Chain-of-Thought (CoT) abilities, even in models that consistently exceed human performance. Furthermore, we demonstrate the potential of targeted fine-tuning on composite visual reasoning tasks and show that fine-tuning smaller VLMs substantially improves core cognitive abilities. While this improvement does not translate to large enhancements on challenging, out-of-distribution benchmarks, we show broadly that VLM performance on our datasets strongly correlates with performance on these other benchmarks. Our work provides a detailed analysis of VLM cognitive strengths and weaknesses and identifies key bottlenecks in simultaneous perception and reasoning while also providing an effective and simple solution.

Abstract--In the era of rapid generative AI development, interactions with large language models (LLMs) pose increasing risks of misuse. Prior research has primarily focused on attacks using template-based prompts and optimization-oriented methods, while overlooking the fact that LLMs possess strong unconstrained deceptive capabilities to attack other LLMs. This paper introduces a novel jailbreaking method inspired by the Chain-of-Thought mechanism. The attacker employs mission transfer to conceal harmful user intent within dialogue and generates a progressive chain of lure questions without relying on predefined templates, enabling successful jailbreaks. T o further improve the attack's strength, we incorporate a helper LLM model that performs randomized narrative optimization over multi-turn interactions, enhancing the attack performance while preserving alignment with the original intent. We also propose a toxicity-based framework using third-party LLMs to evaluate harmful content and its alignment with malicious intent. Extensive experiments demonstrate that our method consistently achieves high attack success rates and elevated toxicity scores across diverse types of LLMs under black-box API settings. These findings reveal the intrinsic potential of LLMs to perform unrestricted attacks in the absence of robust alignment constraints. Our approach offers data-driven insights to inform the design of future alignment mechanisms. Finally, we propose two concrete defense strategies to support the development of safer generative models. Rapid advancement of large language models (LLMs) [1], [2] has greatly improved work efficiency, but has also introduced critical security risks [3]. One essential concern of LLM is jailbreak attacks [4], wherein attackers may craft adversarial prompts to bypass the model's safeguards and lead to harmful or unintentional results. Such attacks compromise the reliability of LLMs and may facilitate misinformation, privacy breaches, or other malicious uses [5].

Often, multilingual language models are trained with the objective to map semantically similar content (in different languages) in the same latent space. In this paper, we show a nuance in this training objective, and find that by changing the language of the input query, we can improve the question answering ability of language models. Our contributions are two-fold. First, we introduce the term Language Specific Knowledge (LSK) to denote queries that are best answered in an "expert language" for a given LLM, thereby enhancing its question-answering ability. We introduce the problem of language selection -- for some queries, language models can perform better when queried in languages other than English, sometimes even better in low-resource languages -- and the goal is to select the optimal language for the query. Second, we introduce simple to strong baselines to test this problem. Additionally, as a first-pass solution to this novel problem, we design LSKExtractor to benchmark the language-specific knowledge present in a language model and then exploit it during inference. To test our framework, we employ three datasets that contain knowledge about both cultural and social behavioral norms. Overall, LSKExtractor achieves up to 10% relative improvement across datasets, and is competitive against strong baselines, while being feasible in real-world settings. Broadly, our research contributes to the open-source development (https://github.com/agarwalishika/LSKExtractor/tree/main) of language models that are inclusive and more aligned with the cultural and linguistic contexts in which they are deployed.

Current research on bias in language models (LMs) predominantly focuses on data quality, with significantly less attention paid to model architecture and temporal influences of data. Even more critically, few studies systematically investigate the origins of bias. We propose a methodology grounded in comparative behavioral theory to interpret the complex interaction between training data and model architecture in bias propagation during language modeling. Building on recent work that relates transformers to n-gram LMs, we evaluate how data, model design choices, and temporal dynamics affect bias propagation. Our findings reveal that: (1) n-gram LMs are highly sensitive to context window size in bias propagation, while transformers demonstrate architectural robustness; (2) the temporal provenance of training data significantly affects bias; and (3) different model architectures respond differentially to controlled bias injection, with certain biases (e.g. sexual orientation) being disproportionately amplified. As language models become ubiquitous, our findings highlight the need for a holistic approach -- tracing bias to its origins across both data and model dimensions, not just symptoms, to mitigate harm.

Efficient quantum state preparation remains a central challenge in first-principles quantum simulations of dynamics in quantum field theories, where the Hilbert space is intrinsically infinite-dimensional. Here, we introduce a large language model (LLM)-assisted framework for quantum-circuit design that systematically scales state-preparation circuits to large lattice volumes. Applied to a 1+1d XY spin chain, the LLM autonomously discovers a compact 4-parameter circuit that captures boundary-induced symmetry breaking with sub-percent energy deviation, enabling successful validation on the \texttt{Zuchongzhi} quantum processor. Guided by this insight, we extend the framework to 2+1d quantum field theories, where scalable variational ansätze have remained elusive. For a scalar field theory, the search yields a symmetry-preserving, 3-parameter shallow-depth ansatz whose optimized parameters converge to size-independent constants for lattices $n \ge 4$, providing, to our knowledge, the first scalable ansatz for this class of 2+1d models. Our results establish a practical route toward AI-assisted, human-guided discovery in quantum simulation.

Integration of Large Language Models (LLMs) with Learning Management Systems (LMSs) can enhance task automation and accessibility in education. However, hallucination where LLMs generate inaccurate or misleading information remains a challenge. This study introduces the Dynamic Course Content Integration (DCCI) mechanism, which dynamically retrieves course content from Canvas LMS and structures it within an LLM's context window via prompt engineering, enabling the LLM-powered assistant, Ask ME, to deliver context-aware, curriculum-aligned responses while mitigating hallucinations. A mixed-methods pilot study grounded in Self-Determination Theory (autonomy, competence) and the Technology Acceptance Model (perceived usefulness, ease of use) evaluated DCCI's effectiveness with 120 first-year programming students at Eötvös Loránd University. The course focused on foundational programming patterns in C#, including writing program specifications. We analyzed 14,746 logged interactions and a post-course survey completed by 101 students. User satisfaction was measured via a 5-point Likert scale (turn-level ratings), while the survey assessed usability, engagement, and ethical concerns. Results indicated high satisfaction (mean 4.65/5) and strong recognition of Ask ME's ability to provide timely, contextually relevant answers to administrative and course-related queries. 78.06% agreed that Ask ME's Canvas integration reduced platform switching, improving usability, engagement, comprehension, and topic exploration. Many students reported reduced hesitation to ask questions and increased motivation for self-directed learning, though concerns about over-reliance on AI and reduced student-teacher interaction emerged. This study demonstrates that DCCI enhances LLM reliability, student satisfaction, and engagement in AI-driven educational automation, while highlighting the importance of balancing

Weight-only post-training quantization (PTQ) compresses the weights of Large Language Models (LLMs) into low-precision representations to reduce memory footprint and accelerate inference. However, the presence of outliers in weights and activations often leads to large quantization errors and severe accuracy degradation, especially in recent reasoning LLMs where errors accumulate across long chains of thought. Existing PTQ methods either fail to sufficiently suppress outliers or introduce significant overhead during inference. In this paper, we propose Pairwise Rotation Quantization (ParoQuant), a weight-only PTQ method that combines hardware-efficient and optimizable independent Givens rotations with channel-wise scaling to even out the magnitude across channels and narrow the dynamic range within each quantization group. We further co-design the inference kernel to fully exploit GPU parallelism and keep the rotations and scaling lightweight at runtime. ParoQuant achieves an average 2.4% accuracy improvement over AWQ on reasoning tasks with less than 10% overhead. This paves the way for more efficient and accurate deployment of reasoning LLMs.