We propose an Autoregressive (AR) Moving-average (MA) attention structure that can adapt to various linear attention mechanisms, enhancing their ability to capture long-range and local temporal patterns in time series. In this paper, we first demonstrate that, for the time series forecasting (TSF) task, the previously overlooked decoder-only autoregressive Transformer model can achieve results comparable to the best baselines when appropriate tokenization and training methods are applied. Moreover, inspired by the ARMA model from statistics and recent advances in linear attention, we introduce the full ARMA structure into existing autoregressive attention mechanisms. By using an indirect MA weight generation method, we incorporate the MA term while maintaining the time complexity and parameter size of the underlying efficient attention models. We further explore how indirect parameter generation can produce implicit MA weights that align with the modeling requirements for local temporal impacts. Experimental results show that incorporating the ARMA structure consistently improves the performance of various AR attentions on TSF tasks, achieving state-of-the-art results.

Large language models (LLMs) have shown remarkable advances in supporting long-context comprehension and processing tasks. However, scaling the generation inference of LLMs to such long contexts incurs significant additional computation load, and demands a substantial GPU memory footprint to maintain the key-value (KV) cache of transformer-based LLMs. Existing KV cache compression methods, such as quantization, face memory bottlenecks as context length increases, while static-sized caches, such as eviction, suffer from inefficient policies. These limitations restrict deployment on consumer-grade devices like a single Nvidia 4090 GPU. To overcome this, we propose Locret, a framework for long-context LLM inference that introduces retaining heads to evaluate the causal importance of KV cache units, allowing for more accurate eviction within a fixed cache size. Locret is fine-tuned on top of the frozen backbone LLM using a minimal amount of data from standard long-context SFT datasets. During inference, we evict low-importance cache units along with a chunked prefill pattern, significantly reducing peak GPU memory usage. We conduct an extensive empirical study to evaluate Locret, where the experimental results show that Locret outperforms the recent competitive approaches, including InfLLM, Quantization, SirLLM, and MInference, in terms of memory efficiency and the quality of generated contents -- Locret achieves over a 20x and 8x KV cache compression ratio compared to the full KV cache for Phi-3-mini-128K and Llama-3.1-8B-instruct. Additionally, Locret can be combined with other methods, such as quantization and token merging. To our knowledge, Locret is the first framework capable of deploying Llama-3.1-8B or similar models on a single Nvidia 4090 GPU, enabling 128K long-context inference without compromising generation quality, and requiring little additional system optimizations.

The continual evolution of autonomous driving technology requires car-following models that can adapt to diverse and dynamic traffic environments. Traditional learning-based models often suffer from performance degradation when encountering unseen traffic patterns due to a lack of continual learning capabilities. This paper proposes a novel car-following model based on continual learning that addresses this limitation. Our framework incorporates Elastic Weight Consolidation (EWC) and Memory Aware Synapses (MAS) techniques to mitigate catastrophic forgetting and enable the model to learn incrementally from new traffic data streams. We evaluate the performance of the proposed model on the Waymo and Lyft datasets which encompass various traffic scenarios. The results demonstrate that the continual learning techniques significantly outperform the baseline model, achieving 0\% collision rates across all traffic conditions. This research contributes to the advancement of autonomous driving technology by fostering the development of more robust and adaptable car-following models.

Evaluation is pivotal for refining Large Language Models (LLMs), pinpointing their capabilities, and guiding enhancements. The rapid development of LLMs calls for a lightweight and easy-to-use framework for swift evaluation deployment. However, considering various implementation details, developing a comprehensive evaluation platform is never easy. Existing platforms are often complex and poorly modularized, hindering seamless incorporation into research workflows. This paper introduces UltraEval, a user-friendly evaluation framework characterized by its lightweight nature, comprehensiveness, modularity, and efficiency. We identify and reimplement three core components of model evaluation (models, data, and metrics). The resulting composability allows for the free combination of different models, tasks, prompts, benchmarks, and metrics within a unified evaluation workflow. Additionally, UltraEval supports diverse models owing to a unified HTTP service and provides sufficient inference acceleration. UltraEval is now available for researchers publicly.

Activation sparsity refers to the existence of considerable weakly-contributed elements among activation outputs. As a prevalent property of the models using the ReLU activation function, activation sparsity has been proven a promising paradigm to boost model inference efficiency. Nevertheless, most large language models (LLMs) adopt activation functions without intrinsic activation sparsity (e.g., GELU and Swish). Some recent efforts have explored introducing ReLU or its variants as the substitutive activation function to help LLMs achieve activation sparsity and inference acceleration, but few can simultaneously obtain high sparsity and comparable model performance. This paper introduces a simple and effective sparsification method named "ProSparse" to push LLMs for higher activation sparsity while maintaining comparable performance. Specifically, after substituting the activation function of LLMs with ReLU, ProSparse adopts progressive sparsity regularization with a factor smoothly increasing along the multi-stage sine curves. This can enhance activation sparsity and mitigate performance degradation by avoiding radical shifts in activation distributions. With ProSparse, we obtain high sparsity of 89.32% for LLaMA2-7B, 88.80% for LLaMA2-13B, and 87.89% for end-size MiniCPM-1B, respectively, achieving comparable performance to their original Swish-activated versions. These present the most sparsely activated models among open-source LLaMA versions and competitive end-size models, considerably surpassing ReluLLaMA-7B (66.98%) and ReluLLaMA-13B (71.56%). Our inference acceleration experiments further demonstrate the significant practical acceleration potential of LLMs with higher activation sparsity, obtaining up to 4.52$\times$ inference speedup.

Fine-tuning pre-trained Vision-Language Models (VLMs) has shown remarkable capabilities in medical image and textual depiction synergy. Nevertheless, many pre-training datasets are restricted by patient privacy concerns, potentially containing noise that can adversely affect downstream performance. Moreover, the growing reliance on multi-modal generation exacerbates this issue because of its susceptibility to adversarial attacks. To investigate how VLMs trained on adversarial noisy data perform on downstream medical tasks, we first craft noisy upstream datasets using multi-modal adversarial attacks. Through our comprehensive analysis, we unveil that moderate noise enhances model robustness and transferability, but increasing noise levels negatively impact downstream task performance. To mitigate this issue, we propose rectify adversarial noise (RAN) framework, a recipe designed to effectively defend adversarial attacks and rectify the influence of upstream noise during fine-tuning.

In the realm of driving technologies, fully autonomous vehicles have not been widely adopted yet, making advanced driver assistance systems (ADAS) crucial for enhancing driving experiences. Adaptive Cruise Control (ACC) emerges as a pivotal component of ADAS. However, current ACC systems often employ fixed settings, failing to intuitively capture drivers' social preferences and leading to potential function disengagement. To overcome these limitations, we propose the Editable Behavior Generation (EBG) model, a data-driven car-following model that allows for adjusting driving discourtesy levels. The framework integrates diverse courtesy calculation methods into long short-term memory (LSTM) and Transformer architectures, offering a comprehensive approach to capture nuanced driving dynamics. By integrating various discourtesy values during the training process, our model generates realistic agent trajectories with different levels of courtesy in car-following behavior. Experimental results on the HighD and Waymo datasets showcase a reduction in Mean Squared Error (MSE) of spacing and MSE of speed compared to baselines, establishing style controllability. To the best of our knowledge, this work represents the first data-driven car-following model capable of dynamically adjusting discourtesy levels. Our model provides valuable insights for the development of ACC systems that take into account drivers' social preferences.

Reinforcement Learning (RL) plays a crucial role in advancing autonomous driving technologies by maximizing reward functions to achieve the optimal policy. However, crafting these reward functions has been a complex, manual process in many practices. To reduce this complexity, we introduce a novel framework that integrates Large Language Models (LLMs) with RL to improve reward function design in autonomous driving. This framework utilizes the coding capabilities of LLMs, proven in other areas, to generate and evolve reward functions for highway scenarios. The framework starts with instructing LLMs to create an initial reward function code based on the driving environment and task descriptions. This code is then refined through iterative cycles involving RL training and LLMs' reflection, which benefits from their ability to review and improve the output. We have also developed a specific prompt template to improve LLMs' understanding of complex driving simulations, ensuring the generation of effective and error-free code. Our experiments in a highway driving simulator across three traffic configurations show that our method surpasses expert handcrafted reward functions, achieving a 22% higher average success rate. This not only indicates safer driving but also suggests significant gains in development productivity.

Fine-tuning is a crucial process for adapting large language models (LLMs) to diverse applications. In certain scenarios, such as multi-tenant serving, deploying multiple LLMs becomes necessary to meet complex demands. Recent studies suggest decomposing a fine-tuned LLM into a base model and corresponding delta weights, which are then compressed using low-rank or low-bit approaches to reduce costs. In this work, we observe that existing low-rank and low-bit compression methods can significantly harm the model performance for task-specific fine-tuned LLMs (e.g., WizardMath for math problems). Motivated by the long-tail distribution of singular values in the delta weights, we propose a delta quantization approach using mixed-precision. This method employs higher-bit representation for singular vectors corresponding to larger singular values. We evaluate our approach on various fine-tuned LLMs, including math LLMs, code LLMs, chat LLMs, and even VLMs. Experimental results demonstrate that our approach performs comparably to full fine-tuned LLMs, surpassing both low-rank and low-bit baselines by a considerable margin. Additionally, we show that our method is compatible with various backbone LLMs, such as Llama-2, Llama-3, and Mistral, highlighting its generalizability.

Recent advancements have seen Large Language Models (LLMs) and Large Multimodal Models (LMMs) surpassing general human capabilities in various tasks, approaching the proficiency level of human experts across multiple domains. With traditional benchmarks becoming less challenging for these models, new rigorous challenges are essential to gauge their advanced abilities. In this work, we present OlympiadBench, an Olympiad-level bilingual multimodal scientific benchmark, featuring 8,476 problems from Olympiad-level mathematics and physics competitions, including the Chinese college entrance exam. Each problem is detailed with expert-level annotations for step-by-step reasoning. Evaluating top-tier models on OlympiadBench, we implement a comprehensive assessment methodology to accurately evaluate model responses. Notably, the best-performing model, GPT-4V, attains an average score of 17.97% on OlympiadBench, with a mere 10.74% in physics, highlighting the benchmark rigor and the intricacy of physical reasoning. Our analysis orienting GPT-4V points out prevalent issues with hallucinations, knowledge omissions, and logical fallacies. We hope that our challenging benchmark can serve as a valuable resource for helping future AGI research endeavors. The data and evaluation code are available at \url{https://github.com/OpenBMB/OlympiadBench}