The current study introduces a novel adaptation of speculative decoding, repurposed from generation to classification tasks. We propose a multi-model framework employing up to three lightweight worker models and a single, more robust judge model analogous to draft models and target model, respectively, in speculative decoding. The worker models, tasked with the bulk of the computation, independently predict discrete class labels for a given input. When majority worker models agree on a label, it is accepted as the final label, optimizing efficiency by bypassing the computationally expensive judge model. In cases of disagreement, the judge model intervenes to resolve the label. This approach minimizes redundant computation, leverages the redundancy of multiple workers for confidence, and confines the judge model's role to challenging cases, offering a practical balance of efficiency and accuracy. Our analysis suggests that smaller out of the box instruction/chat finetuned worker models with 3 billion parameters (hereafter, 3B) demonstrate a level of alignment with judge models comparable to that of larger finetuned worker models with 7 billion parameters (hereafter, 7B) across both simple and higher order reasoning tasks. The top performing 3B worker model pair achieve an agreement rate of approximately 80-83% for sentiment and around 50-80% for similar ticket when compared to judge models. Additionally, 3B worker models provide a speedup ranging from 2.8x to 9x relative to the judge models, while 7B worker model combinations achieve a speedup ranging from 1.28x to 0.28x

Large Vision-Language Models (LVLMs) typically follow a two-stage training paradigm-pretraining and supervised fine-tuning. Recently, preference optimization, derived from the language domain, has emerged as an effective post-training reinforcement strategy to enhance capabilities of LVLMs. However, constructing high-quality human-annotated preference data and developing robust reward models to mimic these preferences are both costly and challenging. Motivated by this observation, we propose Vision-R1, a novel vision-guided R1-like reinforcement learning algorithm for LVLMs that rewards models with definitive vision feedback. It only leverages curated instruction data, eliminating the need for specialized reward models and handcrafted preference datasets. We incorporate a criterion-driven reward function that further integrates multi-dimensional feedback to evaluate model completions comprehensively based on the vision task logic. Furthermore, we introduce a progressive rule refinement strategy that dynamically adjusts the reward criteria during training, enabling continuous model improvement and mitigating reward hacking. Extensive experiments on both in-distribution and out-of-distribution benchmarks demonstrate that fine-tuning the 7B LVLMs with Vision-R1 achieves consistent performance gains, with even up to 50% improvement and surpassing the state-of-the-art 10x size model.

Does the prior knowledge of the vision encoder constrain the capability boundary of Multi-modal Large Language Models (MLLMs)? While most existing research treats MLLMs as unified systems optimized through end-to-end training, the impact of vision encoder's prior knowledge is seldom investigated. In this work, we introduce a novel metric, $Rank_e$, to quantify the effect of the vision encoder's prior knowledge on MLLM performance. Our analysis reveals a positive correlation between prior knowledge and MLLM performance. Moreover, we find that domain-specific fine-tuning using solely end-to-end visual question answering (VQA) data is insufficient--particularly for entities with low inherent visual prior knowledge. To address this issue, we propose VisPRE (Vision Prior Remediation), a two-stage training framework that explicitly incorporates prior knowledge at the vision encoder level. Experimental results demonstrate that augmenting vision encoder's prior knowledge substantially boosts the visual understanding capabilities of MLLMs, offering a novel and effective strategy for improving performance, especially in scenarios involving uncommon visual entities.

Large Language Models (LLMs) have become increasingly vulnerable to jailbreak attacks that circumvent their safety mechanisms. While existing defense methods either suffer from adaptive attacks or require computationally expensive auxiliary models, we present STShield, a lightweight framework for real-time jailbroken judgement. STShield introduces a novel single-token sentinel mechanism that appends a binary safety indicator to the model's response sequence, leveraging the LLM's own alignment capabilities for detection. Our framework combines supervised fine-tuning on normal prompts with adversarial training using embedding-space perturbations, achieving robust detection while preserving model utility. Extensive experiments demonstrate that STShield successfully defends against various jailbreak attacks, while maintaining the model's performance on legitimate queries. Compared to existing approaches, STShield achieves superior defense performance with minimal computational overhead, making it a practical solution for real-world LLM deployment.

Misleading chart visualizations, which intentionally manipulate data representations to support specific claims, can distort perceptions and lead to incorrect conclusions. Despite decades of research, misleading visualizations remain a widespread and pressing issue. Recent advances in multimodal large language models (MLLMs) have demonstrated strong chart comprehension capabilities, yet no existing work has systematically evaluated their ability to detect and interpret misleading charts. This paper introduces the Misleading Chart Question Answering (Misleading ChartQA) Benchmark, a large-scale multimodal dataset designed to assess MLLMs in identifying and reasoning about misleading charts. It contains over 3,000 curated examples, covering 21 types of misleaders and 10 chart types. Each example includes standardized chart code, CSV data, and multiple-choice questions with labeled explanations, validated through multi-round MLLM checks and exhausted expert human review. We benchmark 16 state-of-the-art MLLMs on our dataset, revealing their limitations in identifying visually deceptive practices. We also propose a novel pipeline that detects and localizes misleaders, enhancing MLLMs' accuracy in misleading chart interpretation. Our work establishes a foundation for advancing MLLM-driven misleading chart comprehension. We publicly release the sample dataset to support further research in this critical area.

Islamophobia started evolving into a global phenomenon by attracting followers across the globe, particularly in Western societies. Thus, understanding Islamophobia's global spread and online dissemination is crucial. This paper performs a large-scale analysis of specialized, semi-coded Islamophobic terms such as (muzrat, pislam, mudslime, mohammedan, muzzies) floated on extremist social platforms, i.e., 4Chan, Gab, Telegram, etc. First, we use large language models (LLMs) to show their ability to understand these terms. Second, using Google Perspective API, we also find that Islamophobic text is more toxic compared to other kinds of hate speech. Finally, we use BERT topic modeling approach to extract different topics and Islamophobic discourse on these social platforms. Our findings indicate that LLMs understand these Out-Of-Vocabulary (OOV) slurs; however, measures are still required to control such discourse. Our topic modeling also indicates that Islamophobic text is found across various political, conspiratorial, and far-right movements and is particularly directed against Muslim immigrants. Taken altogether, we performed the first study on Islamophobic semi-coded terms and shed a global light on Islamophobia.

Producing high-quality segmentation masks for medical images is a fundamental challenge in biomedical image analysis. Recent research has explored large-scale supervised training to enable segmentation across various medical imaging modalities and unsupervised training to facilitate segmentation without dense annotations. However, constructing a model capable of segmenting diverse medical images in a zero-shot manner without any annotations remains a significant hurdle. This paper introduces the Attention Diffusion Zero-shot Unsupervised System (ADZUS), a novel approach that leverages self-attention diffusion models for zero-shot biomedical image segmentation. ADZUS harnesses the intrinsic capabilities of pre-trained diffusion models, utilizing their generative and discriminative potentials to segment medical images without requiring annotated training data or prior domain-specific knowledge. The ADZUS architecture is detailed, with its integration of self-attention mechanisms that facilitate context-aware and detail-sensitive segmentations being highlighted. Experimental results across various medical imaging datasets, including skin lesion segmentation, chest X-ray infection segmentation, and white blood cell segmentation, reveal that ADZUS achieves state-of-the-art performance. Notably, ADZUS reached Dice scores ranging from 88.7\% to 92.9\% and IoU scores from 66.3\% to 93.3\% across different segmentation tasks, demonstrating significant improvements in handling novel, unseen medical imagery. It is noteworthy that while ADZUS demonstrates high effectiveness, it demands substantial computational resources and extended processing times. The model's efficacy in zero-shot settings underscores its potential to reduce reliance on costly annotations and seamlessly adapt to new medical imaging tasks, thereby expanding the diagnostic capabilities of AI-driven medical imaging technologies.

Grading is a time-consuming and laborious task that educators must face. It is an important task since it provides feedback signals to learners, and it has been demonstrated that timely feedback improves the learning process. In recent years, the irruption of LLMs has shed light on the effectiveness of automatic grading. In this paper, we explore the performance of different LLMs and prompting techniques in automatically grading short-text answers to open-ended questions. Unlike most of the literature, our study focuses on a use case where the questions, answers, and prompts are all in Spanish. Experimental results comparing automatic scores to those of human-expert evaluators show good outcomes in terms of accuracy, precision and consistency for advanced LLMs, both open and proprietary. Results are notably sensitive to prompt styles, suggesting biases toward certain words or content in the prompt. However, the best combinations of models and prompt strategies, consistently surpasses an accuracy of 95% in a three-level grading task, which even rises up to more than 98% when the it is simplified to a binary right or wrong rating problem, which demonstrates the potential that LLMs have to implement this type of automation in education applications.

Language models have recently advanced into the realm of reasoning, yet it is through multimodal reasoning that we can fully unlock the potential to achieve more comprehensive, human-like cognitive capabilities. This survey provides a systematic overview of the recent multimodal reasoning approaches, categorizing them into two levels: language-centric multimodal reasoning and collaborative multimodal reasoning. The former encompasses one-pass visual perception and active visual perception, where vision primarily serves a supporting role in language reasoning. The latter involves action generation and state update within reasoning process, enabling a more dynamic interaction between modalities. Furthermore, we analyze the technical evolution of these methods, discuss their inherent challenges, and introduce key benchmark tasks and evaluation metrics for assessing multimodal reasoning performance. Finally, we provide insights into future research directions from the following two perspectives: (i) from visual-language reasoning to omnimodal reasoning and (ii) from multimodal reasoning to multimodal agents. This survey aims to provide a structured overview that will inspire further advancements in multimodal reasoning research.

Progress in scientific discovery is rarely the result of a single "Eureka" moment, but is rather the product of hundreds of scientists incrementally working together toward a common goal. While existing agent workflows are capable of producing research autonomously, they do so in isolation, without the ability to continuously improve upon prior research results. To address these challenges, we introduce AgentRxiv-a framework that lets LLM agent laboratories upload and retrieve reports from a shared preprint server in order to collaborate, share insights, and iteratively build on each other's research. We task agent laboratories to develop new reasoning and prompting techniques and find that agents with access to their prior research achieve higher performance improvements compared to agents operating in isolation (11.4% relative improvement over baseline on MATH-500). We find that the best performing strategy generalizes to benchmarks in other domains (improving on average by 3.3%). Multiple agent laboratories sharing research through AgentRxiv are able to work together towards a common goal, progressing more rapidly than isolated laboratories, achieving higher overall accuracy (13.7% relative improvement over baseline on MATH-500). These findings suggest that autonomous agents may play a role in designing future AI systems alongside humans. We hope that AgentRxiv allows agents to collaborate toward research goals and enables researchers to accelerate discovery.