Recent advancements in LLM pretraining have featured ever-expanding context windows to process longer sequences. However, our pilot study reveals that models pretrained with shorter context windows consistently outperform their long-context counterparts under a fixed token budget. This finding motivates us to explore an optimal context window scheduling strategy to better balance long-context capability with pretraining efficiency. To this end, we propose SkyLadder, a simple yet effective approach that implements a short-to-long context window transition. SkyLadder preserves strong standard benchmark performance, while matching or exceeding baseline results on long-context tasks. Through extensive experiments, we pre-train 1B-parameter models (up to 32K context) and 3B-parameter models (8K context) on 100B tokens, demonstrating that SkyLadder yields consistent gains of up to 3.7% on common benchmarks, while achieving up to 22% faster training speeds compared to baselines. The evolution of language models has been marked by a consistent expansion in context window sizes (Figure 1 left). While early models like GPT (Radford, 2018) and BERT (Kenton & Toutanova, 2019) were limited to context windows of 512 tokens, subsequent models have pushed these boundaries significantly. GPT-2 (Radford et al., 2019) doubled this capacity to 1024 tokens, and with the advent of Large Language Models (LLMs) exceeding 1B parameters, the progression continued: Llama (Touvron et al., 2023a) implemented a 2048-token window, Llama-2 (Touvron et al., 2023b) extended it to 4096, and Llama-3 (Dubey et al., 2024) further expanded to 8192 tokens. The push to expand the context window is motivated by the need for models to handle longer sequences during inference. The development is also driven by a widespread belief that models pretrained with longer context windows should perform comparably to, or even surpass, their shorter context counterparts, as extended windows reduce document truncation and preserve coherence (Ding et al., 2024). We question whether the common belief that larger context windows does actually improve performance.

Sailor2 is a family of cutting-edge multilingual language models for South-East Asian (SEA) languages, available in 1B, 8B, and 20B sizes to suit diverse applications. Building on Qwen2.5, Sailor2 undergoes continuous pre-training on 500B tokens (400B SEA-specific and 100B replay tokens) to support 13 SEA languages while retaining proficiency in Chinese and English. Sailor2-20B model achieves a 50-50 win rate against GPT-4o across SEA languages. We also deliver a comprehensive cookbook on how to develop the multilingual model in an efficient manner, including five key aspects: data curation, pre-training, post-training, model customization and evaluation. We hope that Sailor2 model (Apache 2.0 license) will drive language development in the SEA region, and Sailor2 cookbook will inspire researchers to build more inclusive LLMs for other under-served languages.

The Earth's weather system involves intricate weather data modalities and diverse weather understanding tasks, which hold significant value to human life. Existing data-driven models focus on single weather understanding tasks (e.g., weather forecasting). While these models have achieved promising results, they fail to tackle various complex tasks within a single and unified model. Moreover, the paradigm that relies on limited real observations for a single scenario hinders the model's performance upper bound. Inspired by the in-context learning paradigm from visual foundation models and large language models, in this paper, we introduce the first generalist weather generalist foundation model (WeatherGFM) to address weather understanding tasks in a unified manner. Specifically, we first unify the representation and definition for diverse weather understanding tasks. Subsequently, we design weather prompt formats to handle different weather data modalities, including single, multiple, and temporal modalities. Finally, we adopt a visual prompting question-answering paradigm for the training of unified weather understanding tasks. Extensive experiments indicate that our WeatherGFM can effectively handle up to ten weather understanding tasks, including weather forecasting, super-resolution, weather image translation, and post-processing. Modeling Earth weather systems involves a series of complex subprocesses that are intended to transform intricate Earth observation data into applications like weather forecasting (Chen et al., 2023a; Bi et al., 2023), downscaling (Chen et al., 2022), assimilation (Huang et al., 2024), retrieval (Liu et al., 2011), and bias correction (Gong et al., 2024). During the past decade, many data-driven machine learning methods have been investigated for various weather understanding tasks and delivering desirable performance on specific tasks. For example, recent studies using large-scale training data (e.g., ERA5 reanalysis data (Hersbach et al., 2020)) have exceeded the accuracy of conventional numerical weather forecasts.

DiffSR: Learning Radar Reflectivity Synthesis via Diffusion Model from Satellite Observations 1 st Xuming He Open Science Lab Shanghai AI Laboratory Shanghai, China hexuming773@gmail.com 2 nd Zhiwang Zhou Open Science Lab Shanghai AI Laboratory Shanghai, China zhouzhiwang@pjlab.org.cn 3 rd Wenlong Zhang null Open Science Lab Shanghai AI Laboratory Shanghai, China zhangwenlong@pjlab.org.cn 4 th Xiangyu Zhao Department of Computing Hong Kong Polytechnic University HongKong SAR, China 22123675r@connect.polyu.hk 5 th Hao Chen Open Science Lab Shanghai AI Laboratory Shanghai, China chenhao1@pjlab.org.cn Abstract --Weather radar data synthesis can fill in data for areas where ground observations are missing. Existing methods often employ reconstruction-based approaches with MSE loss to reconstruct radar data from satellite observation. However, such methods lead to over-smoothing, which hinders the generation of high-frequency details or high-value observation areas associated with convective weather . T o address this issue, we propose a two-stage diffusion-based method called DiffSR.

With the rapid development of blockchain technology, smart contract security has become a critical challenge. Existing smart contract vulnerability detection methods face three main issues: (1) Insufficient quality of datasets, lacking detailed explanations and precise vulnerability locations. (2) Limited adaptability of large language models (LLMs) to the smart contract domain, as most LLMs are pre-trained on general text data but minimal smart contract-specific data. (3) Lack of high-quality explanations for detected vulnerabilities, as existing methods focus solely on detection without clear explanations. These limitations hinder detection performance and make it harder for developers to understand and fix vulnerabilities quickly, potentially leading to severe financial losses. To address these problems, we propose Smart-LLaMA, an advanced detection method based on the LLaMA language model. First, we construct a comprehensive dataset covering four vulnerability types with labels, detailed explanations, and precise vulnerability locations. Second, we introduce Smart Contract-Specific Continual Pre-Training, using raw smart contract data to enable the LLM to learn smart contract syntax and semantics, enhancing their domain adaptability. Furthermore, we propose Explanation-Guided Fine-Tuning, which fine-tunes the LLM using paired vulnerable code and explanations, enabling both vulnerability detection and reasoned explanations. We evaluate explanation quality through LLM and human evaluation, focusing on Correctness, Completeness, and Conciseness. Experimental results show that Smart-LLaMA outperforms state-of-the-art baselines, with average improvements of 6.49% in F1 score and 3.78% in accuracy, while providing reliable explanations.

Generative models cover various application areas, including image, video and music synthesis, natural language processing, and molecular design, among many others. As digital generative models become larger, scalable inference in a fast and energy-efficient manner becomes a challenge. Here, we present optical generative models inspired by diffusion models, where a shallow and fast digital encoder first maps random noise into phase patterns that serve as optical generative seeds for a desired data distribution; a jointly-trained free-space-based reconfigurable decoder all-optically processes these generative seeds to create novel images (never seen before) following the target data distribution. Except for the illumination power and the random seed generation through a shallow encoder, these optical generative models do not consume computing power during the synthesis of novel images. We report the optical generation of monochrome and multi-color novel images of handwritten digits, fashion products, butterflies, and human faces, following the data distributions of MNIST, Fashion MNIST, Butterflies-100, and Celeb-A datasets, respectively, achieving an overall performance comparable to digital neural network-based generative models. To experimentally demonstrate optical generative models, we used visible light to generate, in a snapshot, novel images of handwritten digits and fashion products. These optical generative models might pave the way for energy-efficient, scalable and rapid inference tasks, further exploiting the potentials of optics and photonics for artificial intelligence-generated content.

While large language models (LLMs) have demonstrated remarkable abilities across various fields, hallucination remains a significant challenge. Recent studies have explored hallucinations through the lens of internal representations, proposing mechanisms to decipher LLMs' adherence to facts. However, these approaches often fail to generalize to out-of-distribution data, leading to concerns about whether internal representation patterns reflect fundamental factual awareness, or only overfit spurious correlations on the specific datasets. In this work, we investigate whether a universal truthfulness hyperplane that distinguishes the model's factually correct and incorrect outputs exists within the model. To this end, we scale up the number of training datasets and conduct an extensive evaluation -- we train the truthfulness hyperplane on a diverse collection of over 40 datasets and examine its cross-task, cross-domain, and in-domain generalization. Our results indicate that increasing the diversity of the training datasets significantly enhances the performance in all scenarios, while the volume of data samples plays a less critical role. This finding supports the optimistic hypothesis that a universal truthfulness hyperplane may indeed exist within the model, offering promising directions for future research.

Human priors play a crucial role in efficiently utilizing data in deep learning. However, with the development of large language models (LLMs), there is an increasing emphasis on scaling both model size and data volume, which often diminishes the importance of human priors in data construction. Influenced by these trends, existing Small Language Models (SLMs) mainly rely on web-scraped large-scale training data, neglecting the proper incorporation of human priors. This oversight limits the training efficiency of language models in resource-constrained settings. In this paper, we propose a principle to leverage human priors for data construction. This principle emphasizes achieving high-performance SLMs by training on a concise dataset that accommodates both semantic diversity and data quality consistency, while avoiding benchmark data leakage. Following this principle, we train an SLM named HARE-1.1B. Extensive experiments on large-scale benchmark datasets demonstrate that HARE-1.1B performs favorably against state-of-the-art SLMs, validating the effectiveness of the proposed principle. Additionally, this provides new insights into efficient language model training in resource-constrained environments from the view of human priors.

Large language models (LLMs) frequently hallucinate and produce factual errors, yet our understanding of why they make these errors remains limited. In this study, we delve into the underlying mechanisms of LLM hallucinations from the perspective of inner representations, and discover a salient pattern associated with hallucinations: correct generations tend to have sharper context activations in the hidden states of the in-context tokens, compared to the incorrect ones. Leveraging this insight, we propose an entropy-based metric to quantify the "sharpness" among the in-context hidden states and incorporate it into the decoding process to formulate a constrained decoding approach. Experiments on various knowledge-seeking and hallucination benchmarks demonstrate our approach's consistent effectiveness, for example, achieving up to an 8.6 point improvement on TruthfulQA. We believe this study can improve our understanding of hallucinations and serve as a practical solution for hallucination mitigation. Large language models (LLMs) have made remarkable advancements in recent years, with extensive applications across various domains (OpenAI, 2022; 2023; Kaddour et al., 2023). Despite these advances, LLMs still face notable challenges regarding factuality, which could critically undermine the trustworthiness and reliability of LLMs, as highlighted in recent studies (Chen et al., 2023; Ji et al., 2023; Wang et al., 2023). To address the factuality issue, many efforts have focused on retrieving external knowledge (Ram et al., 2023b; Yu et al., 2023; Jiang et al., 2023) for generation or fact-checking, as well as fine-tuning (Asai et al., 2023) and self-evaluation (Pan et al., 2023; Xiong et al., 2024). However, these methods often require high computational resources or high-quality knowledge bases, which may not be available for domain-specific cases. In contrast, we aim to tackle this challenge from the perspective of model's inner representations, investigating whether the hidden states contain information about hallucination.

As an efficient alternative to conventional full finetuning, parameter-efficient finetuning (PEFT) is becoming the prevailing method to adapt pretrained language models. In PEFT, a lightweight module is learned on each dataset while the underlying pretrained language model remains unchanged, resulting in multiple compact modules representing diverse skills when applied to various domains and tasks. In this paper, we propose to compose these parameter-efficient modules through linear arithmetic operations in the weight space, thereby integrating different module capabilities. Specifically, we first define addition and negation operators for the module, and then further compose these two basic operators to perform flexible arithmetic. Our approach requires \emph{no additional training} and enables highly flexible module composition. We apply different arithmetic operations to compose the parameter-efficient modules for (1) distribution generalization, (2) multi-tasking, (3) unlearning, and (4) domain transfer. Additionally, we extend our approach to detoxify Alpaca-LoRA, the latest instruction-tuned large language model based on LLaMA. Empirical results demonstrate that our approach produces new and effective parameter-efficient modules that significantly outperform existing ones across all settings.