Despite their impressive capabilities, aligned large language models (LLMs) often generate outputs that lack diversity. What drives this consistency in the generation? We investigate this phenomenon through the lens of probability concentration in the model's output distribution. To quantify this concentration, we introduce the *Branching Factor* (BF)--a token-invariant measure of the effective number of plausible next steps during generation. Our empirical analysis reveals two key findings: (1) BF often decreases as generation progresses, suggesting that LLMs become more predictable as they generate. (2) alignment tuning substantially sharpens the model's output distribution from the outset, reducing BF by nearly an order of magnitude (e.g., from 12 to 1.2) relative to base models. This stark reduction helps explain why aligned models often appear less sensitive to decoding strategies. Building on this insight, we find this consistency has surprising implications for complex reasoning. Aligned Chain-of-Thought (CoT) models (e.g., DeepSeek-distilled models), for instance, leverage this effect; by generating longer reasoning chains, they push generation into later, more deterministic (lower BF) stages, resulting in more stable outputs. We hypothesize that alignment tuning does not fundamentally change a model's behavior, but instead steers it toward stylistic tokens (e.g., ``Sure'') that unlock low-entropy trajectories already present in the base model. This view is supported by nudging experiments, which show prompting base models with such tokens can similarly reduce BF. Together, our findings establish BF as a powerful diagnostic for understanding and controlling LLM outputs - clarifying how alignment reduces variability, how CoT promotes stable generations, and how base models can be steered away from diversity.

Recent developed graph-based methods for diagnosing brain disorders using functional connectivity highly rely on predefined brain atlases, but overlook the rich information embedded within atlases and the confounding effects of site and phenotype variability. To address these challenges, we propose a two-stage Brain-to-Population Graph Learning (B2P-GL) framework that integrates the semantic similarity of brain regions and condition-based population graph modeling. In the first stage, termed brain representation learning, we leverage brain atlas knowledge from GPT-4 to enrich the graph representation and refine the brain graph through an adaptive node reassignment graph attention network. In the second stage, termed population disorder diagnosis, phenotypic data is incorporated into population graph construction and feature fusion to mitigate confounding effects and enhance diagnosis performance. Experiments on the ABIDE I, ADHD-200, and Rest-meta-MDD datasets show that B2P-GL outperforms state-of-the-art methods in prediction accuracy while enhancing interpretability. Overall, our proposed framework offers a reliable and personalized approach to brain disorder diagnosis, advancing clinical applicability.

Large Language Models (LLMs) exhibit enhanced capabilities by Chain-of-Thought reasoning. However, the extended reasoning sequences introduce significant GPU memory overhead due to increased key-value (KV) cache. Existing KV cache compression methods mitigate memory bottlenecks but struggle in long reasoning tasks. In this paper, we analyze attention patterns in reasoning tasks and reveal a Token Importance Recurrence phenomenon: a large proportion of tokens regain high attention after multiple decoding steps, which is failed to capture by existing works and may lead to unpredictable eviction on such periodically critical tokens. To address this, we propose LazyEviction, an observation window-based lagged eviction framework retaining latent recurring tokens by prioritized eviction based on tokens' recurrence patterns. Extensive experiments demonstrate that LazyEviction reduces KV cache by 50%~70% while maintaining comparable accuracy, outperforming existing KV cache compression baselines. Our implementation code can be found at https://github.com/Halo-949/LazyEviction.

Across languages, numeral systems vary widely in how they construct and combine numbers. While humans consistently learn to navigate this diversity, large language models (LLMs) struggle with linguistic-mathematical puzzles involving cross-linguistic numeral systems, which humans can learn to solve successfully. We investigate why this task is difficult for LLMs through a series of experiments that untangle the linguistic and mathematical aspects of numbers in language. Our experiments establish that models cannot consistently solve such problems unless the mathematical operations in the problems are explicitly marked using known symbols ($+$, $\times$, etc., as in "twenty + three"). In further ablation studies, we probe how individual parameters of numeral construction and combination affect performance. While humans use their linguistic understanding of numbers to make inferences about the implicit compositional structure of numerals, LLMs seem to lack this notion of implicit numeral structure. We conclude that the ability to flexibly infer compositional rules from implicit patterns in human-scale data remains an open challenge for current reasoning models.

Human cognitive behavior arises from the interaction of specialized brain networks dedicated to distinct functions, such as language, logic, and social reasoning. Concretely, we partition the layers of a pretrained language model into four expert modules aligned with well-studied cognitive networks in the human brain. 's behavior can be dynamically steered at inference time by routing tokens to particular experts (e.g., favoring social over logical reasoning), enabling fine-grained control over outputs. Taken together, cognitively grounded functional specialization yields models that are both more humanlike and more human-interpretable. Neuroscience research suggests that distinct brain regions support language, reasoning, social cognition, and other cognitive functions (Saxe & Kanwisher, 2003; Kanwisher, 2010; Fedorenko et al., 2024). In contrast, the internal organization of Large Language Models (LLMs) is largely unstructured. While certain units or subnetworks show selective activation (Zhang et al., 2022; 2023; Bayazit et al., 2023; AlKhamissi et al., 2025a; Wang et al., 2025), such specialization is implicit and difficult to interpret or control. Motivated by this discrepancy, we propose a model architecture that explicitly incorporates specialization. On the machine learning (ML) side, such designs hold great potential for improving interpretability and controllability; on the cognitive science side, they provide a framework toward formulating testable computational hypotheses about how the relative contributions of different brain networks support complex behavior. The final training stage of this curriculum uses this now inductively-biased architecture to perform large-scale supervised finetuning.

Time series data in real-world applications such as healthcare, climate modeling, and finance are often irregular, multimodal, and messy, with varying sampling rates, asynchronous modalities, and pervasive missingness. However, existing benchmarks typically assume clean, regularly sampled, unimodal data, creating a significant gap between research and real-world deployment. We introduce Time-IMM, a dataset specifically designed to capture cause-driven irregularity in multimodal multivariate time series. Time-IMM represents nine distinct types of time series irregularity, categorized into trigger-based, constraint-based, and artifact-based mechanisms. Complementing the dataset, we introduce IMM-TSF, a benchmark library for forecasting on irregular multimodal time series, enabling asynchronous integration and realistic evaluation. IMM-TSF includes specialized fusion modules, including a timestamp-to-text fusion module and a multimodality fusion module, which support both recency-aware averaging and attention-based integration strategies. Empirical results demonstrate that explicitly modeling multimodality on irregular time series data leads to substantial gains in forecasting performance. Time-IMM and IMM-TSF provide a foundation for advancing time series analysis under real-world conditions. The dataset is publicly available at https://github.com/blacksnail789521/Time-IMM, and the benchmark library can be accessed at https://github.com/blacksnail789521/IMM-TSF. Project page: https://blacksnail789521.github.io/time-imm-project-page/

The rapid evolution of communication technologies has led to an explosion of standards, rendering traditional expert-dependent consultation methods inefficient and slow. To address this challenge, we propose \textbf{KG2QA}, a question answering (QA) framework for communication standards that integrates fine-tuned large language models (LLMs) with a domain-specific knowledge graph (KG) via a retrieval-augmented generation (RAG) pipeline. We construct a high-quality dataset of 6,587 QA pairs from ITU-T recommendations and fine-tune Qwen2.5-7B-Instruct, achieving significant performance gains: BLEU-4 increases from 18.86 to 66.90, outperforming both the base model and Llama-3-8B-Instruct. A structured KG containing 13,906 entities and 13,524 relations is built using LLM-assisted triple extraction based on a custom ontology. In our KG-RAG pipeline, the fine-tuned LLMs first retrieves relevant knowledge from KG, enabling more accurate and factually grounded responses. Evaluated by DeepSeek-V3 as a judge, the KG-enhanced system improves performance across five dimensions, with an average score increase of 2.26\%, demonstrating superior factual accuracy and relevance. Integrated with Web platform and API, KG2QA delivers an efficient and interactive user experience. Our code and data have been open-sourced https://github.com/luozhongze/KG2QA.

Pretrained molecular encoders have become indispensable in computational chemistry for tasks such as property prediction and molecular generation. However, the standard practice of relying solely on final-layer embeddings for downstream tasks may discard valuable information. In this work, we first analyze the information flow in five diverse molecular encoders and find that intermediate layers retain more general-purpose features, whereas the final-layer specializes and compresses information. We then perform an empirical layer-wise evaluation across 22 property prediction tasks. We find that using frozen embeddings from optimal intermediate layers improves downstream performance by an average of 5.4%, up to 28.6%, compared to the final-layer. Furthermore, finetuning encoders truncated at intermediate depths achieves even greater average improvements of 8.5%, with increases as high as 40.8%, obtaining new state-of-the-art results on several benchmarks. These findings highlight the importance of exploring the full representational depth of molecular encoders to achieve substantial performance improvements and computational efficiency. The code will be made publicly available.

Vision-language models (VLMs) excel in semantic tasks but falter at a core human capability: detecting hidden content in optical illusions or AI-generated images through perceptual adjustments like zooming. We introduce HC-Bench, a benchmark of 112 images with hidden text, objects, and illusions, revealing that leading VLMs achieve near-zero accuracy (0-5.36%)-even with explicit prompting. Humans resolve such ambiguities instinctively, yet VLMs fail due to an overreliance on high-level semantics. Strikingly, we propose SemVink (Semantic Visual Thinking) by simply scaling images to low resolutions (32-128 pixels), which unlocks >99% accuracy by eliminating redundant visual noise. This exposes a critical architectural flaw: VLMs prioritize abstract reasoning over low-level visual operations crucial for real-world robustness. Our work urges a shift toward hybrid models integrating multi-scale processing, bridging the gap between computational vision and human cognition for applications in medical imaging, security, and beyond.

The emergence of ChatGPT marked a transformative milestone for Artificial Intelligence (AI), showcasing the remarkable potential of Large Language Models (LLMs) to generate human-like text. This wave of innovation has revolutionized how we interact with technology, seamlessly integrating LLMs into everyday tasks such as vacation planning, email drafting, and content creation. While English-speaking users have significantly benefited from these advancements, the Arabic world faces distinct challenges in developing Arabic-specific LLMs. Arabic, one of the languages spoken most widely around the world, serves more than 422 million native speakers in 27 countries and is deeply rooted in a rich linguistic and cultural heritage. Developing Arabic LLMs (ALLMs) presents an unparalleled opportunity to bridge technological gaps and empower communities. The journey of ALLMs has been both fascinating and complex, evolving from rudimentary text processing systems to sophisticated AI-driven models. This article explores the trajectory of ALLMs, from their inception to the present day, highlighting the efforts to evaluate these models through benchmarks and public leaderboards. We also discuss the challenges and opportunities that ALLMs present for the Arab world.