Locret: Enhancing Eviction in Long-Context LLM Inference with Trained Retaining Heads

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.

Focus Entirety and Perceive Environment for Arbitrary-Shaped Text Detection

Due to the diversity of scene text in aspects such as font, color, shape, and size, accurately and efficiently detecting text is still a formidable challenge. Among the various detection approaches, segmentation-based approaches have emerged as prominent contenders owing to their flexible pixel-level predictions. However, these methods typically model text instances in a bottom-up manner, which is highly susceptible to noise. In addition, the prediction of pixels is isolated without introducing pixel-feature interaction, which also influences the detection performance. To alleviate these problems, we propose a multi-information level arbitrary-shaped text detector consisting of a focus entirety module (FEM) and a perceive environment module (PEM). The former extracts instance-level features and adopts a top-down scheme to model texts to reduce the influence of noises. Specifically, it assigns consistent entirety information to pixels within the same instance to improve their cohesion. In addition, it emphasizes the scale information, enabling the model to distinguish varying scale texts effectively. The latter extracts region-level information and encourages the model to focus on the distribution of positive samples in the vicinity of a pixel, which perceives environment information. It treats the kernel pixels as positive samples and helps the model differentiate text and kernel features. Extensive experiments demonstrate the FEM's ability to efficiently support the model in handling different scale texts and confirm the PEM can assist in perceiving pixels more accurately by focusing on pixel vicinities. Comparisons show the proposed model outperforms existing state-of-the-art approaches on four public datasets.

Spotlight Text Detector: Spotlight on Candidate Regions Like a Camera

The irregular contour representation is one of the tough challenges in scene text detection. Although segmentation-based methods have achieved significant progress with the help of flexible pixel prediction, the overlap of geographically close texts hinders detecting them separately. To alleviate this problem, some shrink-based methods predict text kernels and expand them to restructure texts. However, the text kernel is an artificial object with incomplete semantic features that are prone to incorrect or missing detection. In addition, different from the general objects, the geometry features (aspect ratio, scale, and shape) of scene texts vary significantly, which makes it difficult to detect them accurately. To consider the above problems, we propose an effective spotlight text detector (STD), which consists of a spotlight calibration module (SCM) and a multivariate information extraction module (MIEM). The former concentrates efforts on the candidate kernel, like a camera focus on the target. It obtains candidate features through a mapping filter and calibrates them precisely to eliminate some false positive samples. The latter designs different shape schemes to explore multiple geometric features for scene texts. It helps extract various spatial relationships to improve the model's ability to recognize kernel regions. Ablation studies prove the effectiveness of the designed SCM and MIEM. Extensive experiments verify that our STD is superior to existing state-of-the-art methods on various datasets, including ICDAR2015, CTW1500, MSRA-TD500, and Total-Text.

Enabling Real-Time Conversations with Minimal Training Costs

Large language models (LLMs) have demonstrated the ability to improve human efficiency through conversational interactions. Conventional LLM-powered dialogue systems, operating on a turn-based paradigm, preclude real-time interaction during response generation. To address this limitation, researchers have proposed duplex models. These models can dynamically adapt to user input, facilitating real-time interactive feedback. However, these methods typically require substantial computational resources to acquire the ability. To reduce overhead, this paper presents a new duplex decoding approach that enhances LLMs with duplex ability, requiring minimal additional training. Specifically, our method employs parallel decoding of queries and responses in conversations, effectively implementing a channel-division-multiplexing decoding strategy. Experimental results indicate that our proposed method significantly enhances the naturalness and human-likeness of user-AI interactions with minimal training costs.

From Words to Wheels: Automated Style-Customized Policy Generation for Autonomous Driving

Autonomous driving technology has witnessed rapid advancements, with foundation models improving interactivity and user experiences. However, current autonomous vehicles (AVs) face significant limitations in delivering command-based driving styles. Most existing methods either rely on predefined driving styles that require expert input or use data-driven techniques like Inverse Reinforcement Learning to extract styles from driving data. These approaches, though effective in some cases, face challenges: difficulty obtaining specific driving data for style matching (e.g., in Robotaxis), inability to align driving style metrics with user preferences, and limitations to pre-existing styles, restricting customization and generalization to new commands. This paper introduces Words2Wheels, a framework that automatically generates customized driving policies based on natural language user commands. Words2Wheels employs a Style-Customized Reward Function to generate a Style-Customized Driving Policy without relying on prior driving data. By leveraging large language models and a Driving Style Database, the framework efficiently retrieves, adapts, and generalizes driving styles. A Statistical Evaluation module ensures alignment with user preferences. Experimental results demonstrate that Words2Wheels outperforms existing methods in accuracy, generalization, and adaptability, offering a novel solution for customized AV driving behavior. Code and demo available at https://yokhon.github.io/Words2Wheels/.

From MOOC to MAIC: Reshaping Online Teaching and Learning through LLM-driven Agents

Since the first instances of online education, where courses were uploaded to accessible and shared online platforms, this form of scaling the dissemination of human knowledge to reach a broader audience has sparked extensive discussion and widespread adoption. Recognizing that personalized learning still holds significant potential for improvement, new AI technologies have been continuously integrated into this learning format, resulting in a variety of educational AI applications such as educational recommendation and intelligent tutoring. The emergence of intelligence in large language models (LLMs) has allowed for these educational enhancements to be built upon a unified foundational model, enabling deeper integration. In this context, we propose MAIC (Massive AI-empowered Course), a new form of online education that leverages LLM-driven multi-agent systems to construct an AI-augmented classroom, balancing scalability with adaptivity. Beyond exploring the conceptual framework and technical innovations, we conduct preliminary experiments at Tsinghua University, one of China's leading universities. Drawing from over 100,000 learning records of more than 500 students, we obtain a series of valuable observations and initial analyses. This project will continue to evolve, ultimately aiming to establish a comprehensive open platform that supports and unifies research, technology, and applications in exploring the possibilities of online education in the era of large model AI. We envision this platform as a collaborative hub, bringing together educators, researchers, and innovators to collectively explore the future of AI-driven online education.

Configurable Foundation Models: Building LLMs from a Modular Perspective

Advancements in LLMs have recently unveiled challenges tied to computational efficiency and continual scalability due to their requirements of huge parameters, making the applications and evolution of these models on devices with limited computation resources and scenarios requiring various abilities increasingly cumbersome. Inspired by modularity within the human brain, there is a growing tendency to decompose LLMs into numerous functional modules, allowing for inference with part of modules and dynamic assembly of modules to tackle complex tasks, such as mixture-of-experts. To highlight the inherent efficiency and composability of the modular approach, we coin the term brick to represent each functional module, designating the modularized structure as configurable foundation models. In this paper, we offer a comprehensive overview and investigation of the construction, utilization, and limitation of configurable foundation models. We first formalize modules into emergent bricks - functional neuron partitions that emerge during the pre-training phase, and customized bricks - bricks constructed via additional post-training to improve the capabilities and knowledge of LLMs. Based on diverse functional bricks, we further present four brick-oriented operations: retrieval and routing, merging, updating, and growing. These operations allow for dynamic configuration of LLMs based on instructions to handle complex tasks. To verify our perspective, we conduct an empirical analysis on widely-used LLMs. We find that the FFN layers follow modular patterns with functional specialization of neurons and functional neuron partitions. Finally, we highlight several open issues and directions for future research. Overall, this paper aims to offer a fresh modular perspective on existing LLM research and inspire the future creation of more efficient and scalable foundational models.

Multi-Modal Multi-Granularity Tokenizer for Chu Bamboo Slip Scripts

This study presents a multi-modal multi-granularity tokenizer specifically designed for analyzing ancient Chinese scripts, focusing on the Chu bamboo slip (CBS) script used during the Spring and Autumn and Warring States period (771-256 BCE) in Ancient China. Considering the complex hierarchical structure of ancient Chinese scripts, where a single character may be a combination of multiple sub-characters, our tokenizer first adopts character detection to locate character boundaries, and then conducts character recognition at both the character and sub-character levels. Moreover, to support the academic community, we have also assembled the first large-scale dataset of CBSs with over 100K annotated character image scans. On the part-of-speech tagging task built on our dataset, using our tokenizer gives a 5.5% relative improvement in F1-score compared to mainstream sub-word tokenizers. Our work not only aids in further investigations of the specific script but also has the potential to advance research on other forms of ancient Chinese scripts.

RIDE: Boosting 3D Object Detection for LiDAR Point Clouds via Rotation-Invariant Analysis

The rotation robustness property has drawn much attention to point cloud analysis, whereas it still poses a critical challenge in 3D object detection. When subjected to arbitrary rotation, most existing detectors fail to produce expected outputs due to the poor rotation robustness. In this paper, we present RIDE, a pioneering exploration of Rotation-Invariance for the 3D LiDAR-point-based object DEtector, with the key idea of designing rotation-invariant features from LiDAR scenes and then effectively incorporating them into existing 3D detectors. Specifically, we design a bi-feature extractor that extracts (i) object-aware features though sensitive to rotation but preserve geometry well, and (ii) rotation-invariant features, which lose geometric information to a certain extent but are robust to rotation. These two kinds of features complement each other to decode 3D proposals that are robust to arbitrary rotations. Particularly, our RIDE is compatible and easy to plug into the existing one-stage and two-stage 3D detectors, and boosts both detection performance and rotation robustness. Extensive experiments on the standard benchmarks showcase that the mean average precision (mAP) and rotation robustness can be significantly boosted by integrating with our RIDE, with +5.6% mAP and 53% rotation robustness improvement on KITTI, +5.1% and 28% improvement correspondingly on nuScenes. The code will be available soon.

More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding

Enabling Large Language Models (LLMs) to comprehend the 3D physical world remains a significant challenge. Due to the lack of large-scale 3D-text pair datasets, the success of LLMs has yet to be replicated in 3D understanding. In this paper, we rethink this issue and propose a new task: 3D Data-Efficient Point-Language Understanding. The goal is to enable LLMs to achieve robust 3D object understanding with minimal 3D point cloud and text data pairs. To address this task, we introduce GreenPLM, which leverages more text data to compensate for the lack of 3D data. First, inspired by using CLIP to align images and text, we utilize a pre-trained point cloud-text encoder to map the 3D point cloud space to the text space. This mapping leaves us to seamlessly connect the text space with LLMs. Once the point-text-LLM connection is established, we further enhance text-LLM alignment by expanding the intermediate text space, thereby reducing the reliance on 3D point cloud data. Specifically, we generate 6M free-text descriptions of 3D objects, and design a three-stage training strategy to help LLMs better explore the intrinsic connections between different modalities. To achieve efficient modality alignment, we design a zero-parameter cross-attention module for token pooling. Extensive experimental results show that GreenPLM requires only 12% of the 3D training data used by existing state-of-the-art models to achieve superior 3D understanding. Remarkably, GreenPLM also achieves competitive performance using text-only data. The code and weights are available at: https://github.com/TangYuan96/GreenPLM.