We present Nanbeige4-3B, a family of small-scale but high-performing language models. Pretrained on 23T high-quality tokens and finetuned on over 30 million diverse instructions, we extend the boundary of the scaling law for small language models. In pre-training, we design a Fine-Grained Warmup-Stable-Decay (FG-WSD) training scheduler, which progressively refines data mixtures across stages to boost model performance. In post-training, to improve the quality of the SFT data, we design a joint mechanism that integrates deliberative generation refinement and chain-of-thought reconstruction, yielding substantial gains on complex tasks. Following SFT, we employ our flagship reasoning model to distill Nanbeige4-3B through our proposed Dual Preference Distillation (DPD) method, which leads to further performance gains. Finally, a multi-stage reinforcement learning phase was applied, leveraging verifiable rewards and preference modeling to strengthen abilities on both reasoning and human alignment. Extensive evaluations show that Nanbeige4-3B not only significantly outperforms models of comparable parameter scale but also rivals much larger models across a wide range of benchmarks. The model checkpoints are available at https://huggingface.co/Nanbeige.

We present a culturally-grounded multimodal dataset of 1,060 traditional recipes crowdsourced from rural communities across remote regions of Eastern India, spanning 10 endangered languages. These recipes, rich in linguistic and cultural nuance, were collected using a mobile interface designed for contributors with low digital literacy. Endangered Language Recipes (ELR)-1000 -- captures not only culinary practices but also the socio-cultural context embedded in indigenous food traditions. We evaluate the performance of several state-of-the-art large language models (LLMs) on translating these recipes into English and find the following: despite the models' capabilities, they struggle with low-resource, culturally-specific language. However, we observe that providing targeted context -- including background information about the languages, translation examples, and guidelines for cultural preservation -- leads to significant improvements in translation quality. Our results underscore the need for benchmarks that cater to underrepresented languages and domains to advance equitable and culturally-aware language technologies. As part of this work, we release the ELR-1000 dataset to the NLP community, hoping it motivates the development of language technologies for endangered languages.

Accurate prediction of enzyme kinetic parameters is crucial for drug discovery, metabolic engineering, and synthetic biology applications. Current computational approaches face limitations in capturing complex enzyme-substrate interactions and often focus on single parameters while neglecting the joint prediction of catalytic turnover numbers (Kcat) and Michaelis-Menten constants (Km). We present EnzyCLIP, a novel dual-encoder framework that leverages contrastive learning and cross-attention mechanisms to predict enzyme kinetic parameters from protein sequences and substrate molecular structures. Our approach integrates ESM-2 protein language model embeddings with ChemBERTa chemical representations through a CLIP-inspired architecture enhanced with bidirectional cross-attention for dynamic enzyme-substrate interaction modeling. EnzyCLIP combines InfoNCE contrastive loss with Huber regression loss to learn aligned multimodal representations while predicting log10-transformed kinetic parameters. The model is trained on the CatPred-DB database containing 23,151 Kcat and 41,174 Km experimentally validated measurements, and achieved competitive performance with R2 scores of 0.593 for Kcat and 0.607 for Km prediction. XGBoost ensemble methods applied to the learned embeddings further improved Km prediction (R2 = 0.61) while maintaining robust Kcat performance.

Recent research has attempted to associate preference optimization (PO) performance with the underlying preference datasets. In this work, our observation is that the differences between the preferred response $y^+$ and dispreferred response $y^-$ influence what LLMs can learn, which may not match the desirable differences to learn. Therefore, we use distance and reward margin to quantify these differences, and combine them to get Distance Calibrated Reward Margin (DCRM), a metric that measures the quality of a response pair for PO. Intuitively, DCRM encourages minimal noisy differences and maximal desired differences. With this, we study 3 types of commonly used preference datasets, classified along two axes: the source of the responses and the preference labeling function. We establish a general correlation between higher DCRM of the training set and better learning outcome. Inspired by this, we propose a best-of-$N^2$ pairing method that selects response pairs with the highest DCRM. Empirically, in various settings, our method produces training datasets that can further improve models' performance on AlpacaEval, MT-Bench, and Arena-Hard over the existing training sets.

Recent advances in large language models (LLMs) highlight the importance of training data structure and quality in shaping reasoning behavior. However, most existing approaches focus on transforming data formats while neglecting the internal reasoning complexity of training samples, leaving the reasoning potential of data under-explored and underutilized. In this work, we posit that LLM logical reasoning performance is jointly constrained by the potential of the training data and the cognitive capacity of the model. To make this relationship measurable, we introduce Data Reasoning Intensity (DRI), a novel metric that quantifies the latent logical reasoning complexity of samples by decomposing and aggregating their logical structures. This allows us to analyze how well current LLMs utilize logical reasoning signals and identify performance gaps relative to data potential. Based on this insight, we introduce a re-cognizing optimization strategy that systematically enhances the logical reasoning intensity of training data. Rather than increasing data volume, our method re-optimizes existing samples to better align with the LLM's logical reasoning boundary. Extensive experiments show that our approach significantly improves performance and generalization over data-centric strategies. We further validate our method under a reinforcement learning framework. Our results indicate that prioritizing reasoning complexity in data rather than sheer scale or superficial form is essential to realizing LLMs' full cognitive potential.

Direct Preference Optimization (DPO) has been widely used for aligning language models with human preferences in a supervised manner. However, several key questions remain unresolved: the rationale behind its log-ratio reward, how the statistical structure of preference datasets shapes its training dynamics, and how those dynamics impact downstream capabilities. We approach these questions from a Bayesian perspective, interpreting the goal of preference optimization as learning the differential information required to update a reference policy into a target policy. To formalize this view, we introduce the Differential Information Distribution (DID), defined as the distribution over samples that carry the Bayesian evidence required to update policies. We introduce three complementary insights by viewing preference optimization through the DID. First, we find that DPO's log-ratio reward is uniquely justified when preferences encode the Differential Information needed to update a reference policy into the target policy. Second, we discuss how commonly observed training dynamics in DPO, including changes in log-likelihood and policy exploration, stem from a power-law DID relationship. Finally, we analyze how training dynamics influence downstream performance using the entropy of DID, a principled measure of uncertainty in the learned information. We observe that learning high-entropy DID improves open-ended instruction-following, while low-entropy DID benefits knowledge-intensive QA. Taken together, our results show that DPO's reward design, training dynamics, and downstream capabilities all emerge as natural consequences of learning Differential Information, offering both a principled theoretical foundation and practical guidance for preference-based alignment.

Long chain-of-thought (CoT) supervision has become a common strategy to enhance reasoning in language models. While effective for large models, we identify a phenomenon we call Long CoT Degradation, in which small language models (SLMs; <=3B parameters) trained on limited long CoT data experience significant performance deterioration. Through extensive experiments on the Qwen2.5, LLaMA3 and Gemma3 families, we demonstrate that this degradation is widespread across SLMs. In some settings, models trained on only 8k long CoT examples lose up to 75% of their original performance before fine-tuning. Strikingly, we further observe that for some particularly small models, even training on 220k long CoT examples fails to recover or surpass their original performance prior to fine-tuning. Our analysis attributes this effect to error accumulation: while longer responses increase the capacity for multi-step reasoning, they also amplify the risk of compounding mistakes. Furthermore, we find that Long CoT Degradation may negatively impacts downstream reinforcement learning (RL), although this can be alleviated by sufficiently scaled supervised fine-tuning (SFT). Our findings challenge common assumptions about the benefits of long CoT training for SLMs and offer practical guidance for building more effective small-scale reasoning models.

Generative propaganda is the use of generative artificial intelligence (AI) to shape public opinion. To characterize its use in real-world settings, we conducted interviews with defenders (e.g., factcheckers, journalists, officials) in Taiwan and creators (e.g., influencers, political consultants, advertisers) as well as defenders in India, centering two places characterized by high levels of online propaganda. The term "deepfakes", we find, exerts outsized discursive power in shaping defenders' expectations of misuse and, in turn, the interventions that are prioritized. To better characterize the space of generative propaganda, we develop a taxonomy that distinguishes between obvious versus hidden and promotional versus derogatory use. Deception was neither the main driver nor the main impact vector of AI's use; instead, Indian creators sought to persuade rather than to deceive, often making AI's use obvious in order to reduce legal and reputational risks, while Taiwan's defenders saw deception as a subset of broader efforts to distort the prevalence of strategic narratives online. AI was useful and used, however, in producing efficiency gains in communicating across languages and modes, and in evading human and algorithmic detection. Security researchers should reconsider threat models to clearly differentiate deepfakes from promotional and obvious uses, to complement and bolster the social factors that constrain misuse by internal actors, and to counter efficiency gains globally.

Heterogeneous Vehicular Networks (HetVNets) play a key role by stacking different communication technologies such as sub-6GHz, mm-wave and DSRC to meet diverse connectivity needs of 5G/B5G vehicular networks. HetVNet helps address the humongous user demands-but maintaining a steady connection in a highly mobile, real-world conditions remain a challenge. Though there has been ample of studies on beam prediction models a dedicated solution for HetVNets is sparsely explored. Hence, it is the need of the hour to develop a reliable beam prediction solution, specifically for HetVNets. This paper introduces a lightweight deep learning-based solution termed-"CAR-BRAINet" which consists of convolutional neural networks with a powerful multi-head attention (MHA) mechanism. Existing literature on beam prediction is largely studied under a limited, idealised vehicular scenario, often overlooking the real-time complexities and intricacies of vehicular networks. Therefore, this study aims to mimic the complexities of a real-time driving scenario by incorporating key factors such as prominent MAC protocols-3GPP-C-V2X and IEEE 802.11BD, the effect of Doppler shifts under high velocity and varying distance and SNR levels into three high-quality dynamic datasets pertaining to urban, rural and highway vehicular networks. CAR-BRAINet performs effectively across all the vehicular scenarios, demonstrating precise beam prediction with minimal beam overhead and a steady improvement of 17.9422% on the spectral efficiency over the existing methods. Thus, this study justifies the effectiveness of CAR-BRAINet in complex HetVNets, offering promising performance without relying on the location angle and antenna dimensions of the mobile users, and thereby reducing the redundant sensor-latency.

Keeping pace with the rapid growth of academia literature presents a significant challenge for researchers, funding bodies, and academic societies. To address the time-consuming manual effort required for scholarly discovery, we present a novel, fully automated system that transitions from data discovery to direct action. Our pipeline demonstrates how a specialized AI agent, 'Agent-E', can be tasked with identifying papers from specific geographic regions within conference proceedings and then executing a Robotic Process Automation (RPA) to complete a predefined action, such as submitting a nomination form. We validated our system on 586 papers from five different conferences, where it successfully identified every target paper with a recall of 100% and a near perfect accuracy of 99.4%. This demonstration highlights the potential of task-oriented AI agents to not only filter information but also to actively participate in and accelerate the workflows of the academic community.