Recent Video-Language Models (VLMs) achieve promising results on long-video understanding, but their performance still lags behind that achieved on tasks involving images or short videos. This has led to great interest in improving the long context modeling of VLMs by introducing novel modules and additional complexity. In this paper, we take a different approach: rather than fine-tuning VLMs with the limited data available, we attempt to maximize the performance of existing models. To this end, we propose a novel visual prompting strategy specifically designed for long-video understanding. By combining multiple frames as panels into one image, we effectively trade off spatial details for temporal resolution. Our approach is training-free, parameter-free, and model-agnostic, and can be seamlessly integrated into existing VLMs. Extensive experiments on five established benchmarks across a wide range of model architectures, sizes, and context windows confirm the consistency of our approach. For the TimeScope (Long) dataset, which has the longest videos, the accuracy for video question answering is improved by up to 19.4%. Overall, our method raises the bar for long video understanding models. We will make our code available upon acceptance. Connecting the reasoning capabilities of Large Language Models (LLMs) with other modalities has opened up the possibility to reason about and interact with multi-modal data (Yin et al., 2024; Yi et al., 2025).

Pipeline parallelism is widely used to train large language models (LLMs). However, increasing heterogeneity in model architectures exacerbates pipeline bubbles, thereby reducing training efficiency. Existing approaches overlook the co-optimization of model partition, model placement, and workload scheduling, resulting in limited efficiency improvement or even performance degradation. To respond, we propose AdaPtis, an LLM training system that supports adaptive pipeline parallelism. First, we develop a pipeline performance model to accurately estimate training throughput. Second, AdaPtis jointly optimizes model partition, model placement, and workload scheduling policies guided by this performance model. Third, we design a unified pipeline executor that efficiently supports the execution of diverse pipeline strategies. Extensive experiments show that AdaPtis achieves an average speedup of 1.42x (up to 2.14x) over Megatron-LM I-1F1B across various LLM architectures and scales.

Sparse Autoencoder (SAE) features have become essential tools for mechanistic interpretability research. SAE features are typically characterized by examining their activating examples, which are often "monosemantic" and align with human interpretable concepts. However, these examples don't reveal feature sensitivity: how reliably a feature activates on texts similar to its activating examples. In this work, we develop a scalable method to evaluate feature sensitivity. Our approach avoids the need to generate natural language descriptions for features; instead we use language models to generate text with the same semantic properties as a feature's activating examples. We then test whether the feature activates on these generated texts. We demonstrate that sensitivity measures a new facet of feature quality and find that many interpretable features have poor sensitivity. Human evaluation confirms that when features fail to activate on our generated text, that text genuinely resembles the original activating examples. Lastly, we study feature sensitivity at the SAE level and observe that average feature sensitivity declines with increasing SAE width across 7 SAE variants. Our work establishes feature sensitivity as a new dimension for evaluating both individual features and SAE architectures.

Ensuring that both new and experienced drivers master current traffic rules is critical to road safety. This paper evaluates Large Language Models (LLMs) on Romanian driving-law QA with explanation generation. We release a 1{,}208-question dataset (387 multimodal) and compare text-only and multimodal SOTA systems, then measure the impact of domain-specific fine-tuning for Llama 3.1-8B-Instruct and RoLlama 3.1-8B-Instruct. SOTA models perform well, but fine-tuned 8B models are competitive. Textual descriptions of images outperform direct visual input. Finally, an LLM-as-a-Judge assesses explanation quality, revealing self-preference bias. The study informs explainable QA for less-resourced languages.

Multimodal Large Language Models (MLLMs) show promising results for embodied agents in operating meaningfully in complex, human-centered environments. Yet, evaluating their capacity for nuanced, human-like reasoning and decision-making remains challenging. In this work, we introduce VIVA+, a cognitively grounded benchmark for evaluating the reasoning and decision-making of MLLMs in human-centered situations. VIVA+ consists of 1,317 real-world situations paired with 6,373 multiple-choice questions, targeting three core abilities for decision-making: (1) Foundational Situation Comprehension, (2) Context-Driven Action Justification, and (3) Reflective Reasoning. Together, these dimensions provide a systematic framework for assessing a model's ability to perceive, reason, and act in socially meaningful ways. We evaluate the latest commercial and open-source models on VIVA+, where we reveal distinct performance patterns and highlight significant challenges. We further explore targeted training and multi-step reasoning strategies, which yield consistent performance improvements. Finally, our in-depth analysis highlights current model limitations and provides actionable insights for advancing MLLMs toward more robust, context-aware, and socially adept decision-making in real-world settings.

Time series foundation models (TSFMs) offer strong zero-shot forecasting via large-scale pre-training, yet fine-tuning remains critical for boosting performance in domains with limited public data. With the growing number of TSFMs, efficiently identifying the best model for downstream fine-tuning becomes increasingly challenging. Leveraging the natural tabular structure formed by dataset meta-features, model characteristics, and fine-tuned performance, we employ tabular foundation models to serve as in-context learners. We establish a comprehensive benchmark for transferability estimation including 10 datasets, 10 foundation models, and 3 forecasting tasks. 's estimation demonstrates strong alignment with actual fine-tuned performance for previously unseen datasets, achieving a mean rank correlation of approximately 0.6 and a 30% improvement compared to using zero-shot performance as the transferability score. The emergence of time series foundation models (TSFMs) is reshaping the paradigm of time series forecasting (Liang et al., 2025) through their strong zero-shot capabilities. Although efficient and cost-effective, zero-shot inference often underperforms in out-of-distribution scenarios, particularly in domains with limited public data, such as healthcare (Gupta et al., 2024) and finance (Fu et al., 2024). Fine-tuning helps bridge the gap by transferring generalized knowledge from large-scale pre-training to specific, resource-limited downstream tasks (Li & Zhu, 2025). However, due to the inherent diversity of time series data, no single model consistently outperforms others in all scenarios (Brigato et al., 2025).

Fine-tuned Large Language Models (LLMs) encode rich task-specific features, but the form of these representations, especially within MLP layers, remains unclear. Empirical inspection of LoRA updates shows that new features concentrate in mid-layer MLPs, yet the scale of these layers obscures meaningful structure. Prior probing suggests that statistical priors may strengthen, split, or vanish across depth, motivating the need to study how neurons work together rather than in isolation. We introduce a mechanistic interpretability framework based on coalitional game theory, where neurons mimic agents in a hedonic game whose preferences capture their synergistic contributions to layer-local computations. Using top-responsive utilities and the PAC-Top-Cover algorithm, we extract stable coalitions of neurons: groups whose joint ablation has non-additive effects. We then track their transitions across layers as persistence, splitting, merging, or disappearance. Applied to LLaMA, Mistral, and Pythia rerankers fine-tuned on scalar IR tasks, our method finds coalitions with consistently higher synergy than clustering baselines. By revealing how neurons cooperate to encode features, hedonic coalitions uncover higher-order structure beyond disentanglement and yield computational units that are functionally important, interpretable, and predictive across domains.

Mixture-of-Experts (MoE) models have become the consensus approach for enabling parameter-efficient scaling and cost-effective deployment in large language models. However, existing scaling laws for dense models are inapplicable to MoE models, which stems from three critical challenges: the multiplicity of influencing factors, their intricate coupling relationships and the non-monotonic nature of their performance impacts. Specifically, we design 446 controlled experiments to characterize their marginal effects, ultimately constructing a comprehensive and precise joint MoE scaling law that considers all essential factors. Our results demonstrate that the optimal settings for G and S are independent of both the model architecture and data size. Our proposed MoE scaling law could function as an accurate and insightful guidance to facilitate future MoE model design and training. Large language models (LLMs) have been widely verified and utilized in our daily lives. It is impressive and lucky to discover that LLMs can continuously expand its ability boundaries with increasing model and training data sizes. The scaling laws of LLMs (Kaplan et al., 2020; Hoffmann et al., 2022; Sun et al., 2025), which could predict the model loss based on crucial factors (e.g., data/model sizes) before training, shed lights on the promising way of wisely selecting appropriate model structures and settings before experiments and continuously enhancing the ability of LLMs under given training budget or environment constraints. Recently, Mixture-of-Experts (MoE) becomes one of the mainstream structures broadly used in powerful industry-level LLMs (Dubey et al., 2024; Liu et al., 2024; Sun et al., 2024; Liu et al., 2025; Qwen Team et al., 2025; OpenAI et al., 2025).

Large Reasoning Models (LRMs) generate explicit reasoning traces alongside final answers, yet the extent to which these traces influence answer generation remains unclear. In this work, we conduct a three-stage investigation into the interplay between reasoning and answer generation in three distilled DeepSeek R1 models. First, through empirical evaluation, we demonstrate that including explicit reasoning consistently improves answer quality across diverse domains. Second, attention analysis reveals that answer tokens attend substantially to reasoning tokens, with certain mid-layer Reasoning-Focus Heads (RFHs) closely tracking the reasoning trajectory, including self-reflective cues. Third, we apply mechanistic interventions using activation patching to assess the dependence of answer tokens on reasoning activations. Our results show that perturbations to key reasoning tokens can reliably alter the final answers, confirming a directional and functional flow of information from reasoning to answer. These findings deepen our understanding of how LRMs leverage reasoning tokens for answer generation, highlighting the functional role of intermediate reasoning in shaping model outputs. Our data and code are publicly available at \href{https://aka.ms/R2A-code}{this URL}.

Enhancing the complex reasoning capabilities of Large Language Models (LLMs) attracts widespread attention. While reinforcement learning (RL) has shown superior performance for improving complex reasoning, its impact on cross-lingual generalization compared to Supervised Fine-Tuning (SFT) remains unexplored. We present the first systematic investigation into cross-lingual reasoning generalization of RL and SFT. Using Qwen2.5-3B-Base as our foundation model, we conduct experiments on diverse multilingual reasoning benchmarks, including math reasoning, commonsense reasoning, and scientific reasoning. Our investigation yields two significant findings: (1) Tuning with RL not only achieves higher accuracy but also demonstrates substantially stronger cross-lingual generalization capabilities compared to SFT. (2) RL training on non-English data yields better overall performance and generalization than training on English data, which is not observed with SFT. Furthermore, through comprehensive mechanistic analyses, we explore the underlying factors of RL's superiority and generalization across languages. Our results provide compelling evidence that RL enables the model with more robust reasoning strategies, offering crucial guidance for more equitable and effective multilingual reasoning.