Evaluating the factuality of long-form generations from Large Language Models (LLMs) remains challenging due to efficiency bottlenecks and reliability concerns. Prior efforts attempt this by decomposing text into claims, searching for evidence, and verifying claims, but suffer from critical drawbacks: (1) inefficiency due to overcomplicated pipeline components, and (2) ineffectiveness stemming from inaccurate claim sets and insufficient evidence. To address these limitations, we propose \textbf{FaStfact}, an evaluation framework that achieves the highest alignment with human evaluation and time/token efficiency among existing baselines. FaStfact first employs chunk-level claim extraction integrated with confidence-based pre-verification, significantly reducing the time and token cost while ensuring reliability. For searching and verification, it collects document-level evidence from crawled web-pages and selectively retrieves it during verification. Extensive experiments based on an annotated benchmark \textbf{FaStfact-Bench} demonstrate the reliability of FaStfact in both efficiently and effectively evaluating long-form factuality. Code, benchmark data, and annotation interface tool are available at https://github.com/Yingjia-Wan/FaStfact.

Large-scale Speech Language Models (SpeechLMs) have enabled voice assistants capable of understanding natural spoken queries and performing complex tasks. However, existing speech benchmarks primarily focus on isolated capabilities such as transcription, or question-answering, and do not systematically evaluate agentic scenarios encompassing multilingual and cultural understanding, as well as adversarial robustness. To address this, we introduce VoiceAgentBench, a comprehensive benchmark designed to evaluate SpeechLMs in realistic spoken agentic settings. It comprises over 5,500 synthetic spoken queries, including dialogues grounded in Indian context, covering single-tool invocations, multi-tool workflows, multi-turn interactions, and safety evaluations. The benchmark supports English, Hindi, and 5 other Indian languages, reflecting real-world linguistic and cultural diversity. We simulate speaker variability using a novel sampling algorithm that selects audios for TTS voice conversion based on its speaker embeddings, maximizing acoustic and speaker diversity. Our evaluation measures tool selection accuracy, structural consistency, and the correctness of tool invocations, including adversarial robustness. Our experiments reveal significant gaps in contextual tool orchestration tasks, Indic generalization, and adversarial robustness, exposing critical limitations of current SpeechLMs.

We introduce Quokka, the first systematic scaling law for diffusion language models (DLMs), encompassing both compute-constrained and data-constrained regimes, and studying the key modeling and optimization designs. Quokka is a good friend of Chinchilla and provides wider scopes. We hope the results would bring short-term practical guidance in DLMs training and long-term inspirations for the whole AI community.

While table understanding increasingly relies on pixel-only settings where tables are processed as visual representations, current benchmarks predominantly use synthetic renderings that lack the complexity and visual diversity of real-world tables. Additionally, existing visual table understanding (VTU) datasets offer fixed examples with single visualizations and pre-defined instructions, providing no access to underlying serialized data for reformulation. Each example includes paired image-HTML representations, comprehensive metadata, and provenance information linking back to the source datasets. The field of table understanding focuses on techniques for representing and interpreting tabular data to support a wide range of practical tasks such as question answering, summarization, and information extraction. Research in this area has traditionally representated tables as structured text, encoding their content and layout through linearized or graph-based representations (see Figure 1b; Herzig et al. 2020; Zhang et al. 2020; Liu et al. 2022). While this unimodal view remains effective in certain domains, many tables found in documents and webpages contain irregular structures, rely on visual formatting (e.g., merged cells, background colors, font variations), or embed multimodal elements such as images (see Figure 1a). Advances in Vision-Language Models (VLMs; Radford et al. 2021; Liu et al. 2023) have provided impetus for treating tables as images, eschewing the step of rendering them as text sequences (like Markdown or HTML). The conceptual simplicity of this approach, coupled with improved performance on several tabular tasks (Alonso et al., 2024; Zhou et al., 2025) has driven significant research interest (Zheng et al., 2024b; Su et al., 2024; Jiang et al., 2025) in Visual T able Understanding (also known as Multimodal T able Understanding). Visual representations of tables are not only merely convenient but in many cases necessary, particularly for VLM agents that interact with the world exclusively through pixels (e.g., on a screen) and must interpret tables directly in their visual form (Deng et al., 2023; Zheng et al., 2024a; Lu et al., 2024). Despite the growing relevance of VTU, there are few resources that support training models directly on image-based representations of tables. Existing benchmarks like MMTab (Zheng et al., 2024b) consist of web tables (e.g., from Wikipedia which is a common source for many tabular datasets), that are serialized and subsequently rendered as synthetic images (see Figures 1b,c). As a result, models trained on such data face a train-test mismatch, since the visual patterns learned from serialized renderings do not generalize well to naturally occurring tables failing to capture critical visual cues like subtle ruling lines, intricate merged cell layouts, background colors, font variations, or embedded images that are inherent to real-world table comprehension (compare Figure 1a and 1c).

Detecting hateful content is a challenging and important problem. Automated tools, like machine-learning models, can help, but they require continuous training to adapt to the ever-changing landscape of social media. In this work, we evaluate eight open-source LLMs' capability to detect antisemitic content, specifically leveraging in-context definition. We also study how LLMs understand and explain their decisions given a moderation policy as a guideline. First, we explore various prompting techniques and design a new CoT-like prompt, Guided-CoT, and find that injecting domain-specific thoughts increases performance and utility. Guided-CoT handles the in-context policy well, improving performance and utility by reducing refusals across all evaluated models, regardless of decoding configuration, model size, or reasoning capability. Notably, Llama 3.1 70B outperforms fine-tuned GPT-3.5. Additionally, we examine LLM errors and introduce metrics to quantify semantic divergence in model-generated rationales, revealing notable differences and paradoxical behaviors among LLMs. Our experiments highlight the differences observed across LLMs' utility, explainability, and reliability. Code and resources available at: https://github.com/idramalab/quantify-llm-explanations

Attention mechanisms have achieved significant empirical success in multiple fields, but their underlying optimization objectives remain unclear yet. Moreover, the quadratic complexity of self-attention has become increasingly prohibitive. Although interpretability and efficiency are two mutually reinforcing pursuits, prior work typically investigates them separately. In this paper, we propose a unified optimization objective that derives inherently interpretable and efficient attention mechanisms through algorithm unrolling. Precisely, we construct a gradient step of the proposed objective with a set of forward-pass operations of our \emph{Contract-and-Broadcast Self-Attention} (CBSA), which compresses input tokens towards low-dimensional structures by contracting a few representatives of them. This novel mechanism can not only scale linearly by fixing the number of representatives, but also covers the instantiations of varied attention mechanisms when using different sets of representatives. We conduct extensive experiments to demonstrate comparable performance and superior advantages over black-box attention mechanisms on visual tasks. Our work sheds light on the integration of interpretability and efficiency, as well as the unified formula of attention mechanisms.

The increasing reliance on generative AI models is rapidly increasing the volume of synthetic data, with some projections suggesting that most available new data for training could be machine-generated by 2030 Gartner, Inc. (2022). This shift to a mainly synthetic content presents a critical challenge: repeated training in synthetic data leads to a phenomenon known as model collapse, where model performance degrades over generations of training, eventually rendering the models ineffective. While the causes of model collapse are increasingly understood, effective mitigation strategies remain scarce. We address this challenge by leveraging a key insight: auto-regressive models tend to generate text sequences to which they assign high confidence (i.e., high log-likelihood). Based on this observation, we introduce the Truncated-Cross-Entropy (TCE) loss function. Our experiments demonstrate that models trained with TCE not only learn effectively but also exhibit significantly increased resilience, tolerating over 2.3 more synthetic data before the onset of collapse. In addition, we provide an open-source benchmark for collapse dynamics in mixed-data settings. Our results demonstrate that confidence-aware training objectives can substantially delay collapse onset, offering a practical and generalizable tool for model robustness under synthetic-data exposure. Generative models have become the foundation for modern AI applications in several modalities, including text, image, code, and audio. Large Language Models (LLMs) such as ChatGPT (OpenAI et al., 2024), LLaMA (Grattafiori et al., 2024) and Gemma (Team et al., 2025), as well as image generators DALL-E (Ramesh et al., 2021) and Imagen (Saharia et al., 2022), all rely on large datasets scraped from the Web. As these models are continuously updated to reflect recent knowledge and linguistic patterns, the need for ever larger and frequently refreshed training corpora has grown substantially. However, this demand is colliding with a shift in the data landscape: synthetic content is increasingly populating the Internet, contaminating the very datasets used for model training. This shift raises fundamental concerns.

In this paper, we propose AnalogSeeker, an effort toward an open-source foundation language model for analog circuit design, with the aim of integrating domain knowledge and giving design assistance. To overcome the scarcity of data in this field, we employ a corpus collection strategy based on the domain knowledge framework of analog circuits. High-quality, accessible textbooks across relevant subfields are systematically curated and cleaned into a textual domain corpus. To address the complexity of knowledge of analog circuits, we introduce a granular domain knowledge distillation method. Raw, unlabeled domain corpus is decomposed into typical, granular learning nodes, where a multi-agent framework distills implicit knowledge embedded in unstructured text into question-answer data pairs with detailed reasoning processes, yielding a fine-grained, learnable dataset for fine-tuning. To address the unexplored challenges in training analog circuit foundation models, we explore and share our training methods through both theoretical analysis and experimental validation. We finally establish a fine-tuning-centric training paradigm, customizing and implementing a neighborhood self-constrained supervised fine-tuning algorithm. This approach enhances training outcomes by constraining the perturbation magnitude between the model's output distributions before and after training. In practice, we train the Qwen2.5-32B-Instruct model to obtain AnalogSeeker, which achieves 85.04% accuracy on AMSBench-TQA, the analog circuit knowledge evaluation benchmark, with a 15.67% point improvement over the original model and is competitive with mainstream commercial models. Furthermore, AnalogSeeker also shows effectiveness in the downstream operational amplifier design task. AnalogSeeker is open-sourced at https://huggingface.co/analogllm/analogseeker for research use.

Abstract--During reasoning in vision-language models (VLMs), false positive (FP) reasoning occurs when a model produces the correct answer but follows an incorrect reasoning path, resulting in undermined reasoning reliability. Existing approaches mainly rely on prompt engineering, knowledge distillation or reinforcement learning to improve reasoning reliability, both of which require large amounts of high-quality data and thus limit practical applicability. Few approaches have focused on directly detecting and correcting FPs. T o address these issues, we propose ViFP, a framework for Visual False Positive Detection to Enhance Reasoning Reliability in VLMs. ViFP builds effective reasoning paths through multi-turn QA and dynamically analyzes the consistency of the reasoning path to identify potential FPs. It also introduces a targeted reasoning chain correction mechanism to modify FP reasoning, thereby improving logical consistency and accuracy. Finally, we introduce a reliability evaluation metric--V oC, which integrates answer accuracy and the FP rate, providing a quantitative tool to assess whether a VLM not only answers correctly but also reasons reliably. Our experiments on closed-source VLMs show that ViFP consistently improves performance across three datasets: A-OKVQA, OK-VQA, and FVQA. On A-OKVQA, ViFP improves accuracy by up to 5.4%, surpassing the previous state-of-the-art by 4.3%, and significantly reduces the number of FPs, validating its benefits in enhancing reasoning reliability.

Large language models (LLMs) excel at generating code from natural language (NL) descriptions. However, the plain textual descriptions are inherently ambiguous and often fail to capture complex requirements like intricate system behaviors, conditional logic, and architectural constraints; implicit data dependencies in service-oriented architectures are difficult to infer and handle correctly. To bridge this gap, we propose a novel step-by-step code generation framework named UML2Dep by leveraging unambiguous formal specifications of complex requirements. First, we introduce an enhanced Unified Modeling Language (UML) sequence diagram tailored for service-oriented architectures. This diagram extends traditional visual syntax by integrating decision tables and API specifications, explicitly formalizing structural relationships and business logic flows in service interactions to rigorously eliminate linguistic ambiguity. Second, recognizing the critical role of data flow, we introduce a dedicated data dependency inference (DDI) task. DDI systematically constructs an explicit data dependency graph prior to actual code synthesis. To ensure reliability, we formalize DDI as a constrained mathematical reasoning task through novel prompting strategies, aligning with LLMs' excellent mathematical strengths. Additional static parsing and dependency pruning further reduce context complexity and cognitive load associated with intricate specifications, thereby enhancing reasoning accuracy and efficiency.