Yuxiang Huang

Weilin Zhao, Yuxiang Huang, Xu Han et al. · tsinghua

Recently, there has been a demand to deploy Large Language Models (LLMs) on personal devices such as laptops and smartphones. These LLMs have different model variants when handling different tasks. However, personal devices have limited resources and require reduced storage overhead. To address this, there are two key methods available: the first is model compression, which compresses LLMs into smaller sizes; the second is LoRA, which can transfer an LLM to other tasks with very few parameters, avoiding the storage of multiple model variants in multi-task scenarios by only preserving LoRAs. However, our experiments show that directly combining these two methods yields sub-optimal performance. Considering that the open-source community has already contributed many LoRAs to LLMs, we propose to adapt these existing LoRAs from the LLMs to their compressed version and introduce a Compression-Aware LoRA (CA-LoRA) framework. We incorporate knowledge inheritance and recovery strategies to recover the lost knowledge caused by model compression. Experiment results demonstrate that CA-LoRA outperforms the vanilla LoRA methods applied to a compressed LLM and achieves comparable performance to the non-compressed LLM with existing LoRA modules. The source code of CA-LoRA is available at https://github.com/thunlp/CA-LoRA.

Yujia Qin, Shengding Hu, Yankai Lin et al. · tsinghua

Humans possess an extraordinary ability to create and utilize tools, allowing them to overcome physical limitations and explore new frontiers. With the advent of foundation models, AI systems have the potential to be equally adept in tool use as humans. This paradigm, i.e., tool learning with foundation models, combines the strengths of specialized tools and foundation models to achieve enhanced accuracy, efficiency, and automation in problem-solving. Despite its immense potential, there is still a lack of a comprehensive understanding of key challenges, opportunities, and future endeavors in this field. To this end, we present a systematic investigation of tool learning in this paper. We first introduce the background of tool learning, including its cognitive origins, the paradigm shift of foundation models, and the complementary roles of tools and models. Then we recapitulate existing tool learning research into tool-augmented and tool-oriented learning. We formulate a general tool learning framework: starting from understanding the user instruction, models should learn to decompose a complex task into several subtasks, dynamically adjust their plan through reasoning, and effectively conquer each sub-task by selecting appropriate tools. We also discuss how to train models for improved tool-use capabilities and facilitate the generalization in tool learning. Considering the lack of a systematic tool learning evaluation in prior works, we experiment with 18 representative tools and show the potential of current foundation models in skillfully utilizing tools. Finally, we discuss several open problems that require further investigation for tool learning. In general, we hope this paper could inspire future research in integrating tools with foundation models.

Yuxiang Huang, Mingye Li, Xu Han et al.

The efficiency of long-video inference remains a critical bottleneck, mainly due to the dense computation in the prefill stage of Large Multimodal Models (LMMs). Existing methods either compress visual embeddings or apply sparse attention on a single GPU, yielding limited acceleration or degraded performance and restricting LMMs from handling longer, more complex videos. To overcome these issues, we propose Spava, a sequence-parallel framework with optimized attention that accelerates long-video inference across multiple GPUs. By distributing approximate attention, Spava reduces computation and increases parallelism, enabling efficient processing of more visual embeddings without compression and thereby improving task performance. System-level optimizations, such as load balancing and fused forward passes, further unleash the potential of Spava, delivering speedups of 12.72x, 1.70x, and 1.18x over FlashAttn, ZigZagRing, and APB, without notable performance loss. Code available at https://github.com/thunlp/APB

Shengding Hu, Yuge Tu, Xu Han et al. · tsinghua

The burgeoning interest in developing Large Language Models (LLMs) with up to trillion parameters has been met with concerns regarding resource efficiency and practical expense, particularly given the immense cost of experimentation. This scenario underscores the importance of exploring the potential of Small Language Models (SLMs) as a resource-efficient alternative. In this context, we introduce MiniCPM, specifically the 1.2B and 2.4B non-embedding parameter variants, not only excel in their respective categories but also demonstrate capabilities on par with 7B-13B LLMs. While focusing on SLMs, our approach exhibits scalability in both model and data dimensions for future LLM research. Regarding model scaling, we employ extensive model wind tunnel experiments for stable and optimal scaling. For data scaling, we introduce a Warmup-Stable-Decay (WSD) learning rate scheduler (LRS), conducive to continuous training and domain adaptation. We present an in-depth analysis of the intriguing training dynamics that occurred in the WSD LRS. With WSD LRS, we are now able to efficiently study data-model scaling law without extensive retraining experiments on both axes of model and data, from which we derive the much higher compute optimal data-model ratio than Chinchilla Optimal. Additionally, we introduce MiniCPM family, including MiniCPM-DPO, MiniCPM-MoE and MiniCPM-128K, whose excellent performance further cementing MiniCPM's foundation in diverse SLM applications. MiniCPM models are available publicly at https://github.com/OpenBMB/MiniCPM .

Fanxing Li, Shengyang Wang, Yuxiang Huang et al.

Obstacle avoidance is a fundamental vision-based task essential for enabling quadrotors to perform advanced applications. When planning the trajectory, existing approaches both on optimization and learning typically regard quadrotor as a point-mass model, giving path or velocity commands then tracking the commands by outer-loop controller. However, at high speeds, planned trajectories sometimes become dynamically infeasible in actual flight, which beyond the capacity of controller. In this paper, we propose a novel end-to-end policy that directly maps depth images to low-level bodyrate commands by reinforcement learning via differentiable simulation. The high-fidelity simulation in training after parameter identification significantly reduces all the gaps between training, simulation and real world. Analytical process by differentiable simulation provides accurate gradient to ensure efficiently training the low-level policy without expert guidance. The policy employs a lightweight and the most simple inference pipeline that runs without explicit mapping, backbone networks, primitives, recurrent structures, or backend controllers, nor curriculum or privileged guidance. By inferring low-level command directly to the hardware controller, the method enables full flight envelope control and avoids the dynamic-infeasible issue.Experimental results demonstrate that the proposed approach achieves the highest success rate and the lowest jerk among state-of-the-art baselines across multiple benchmarks. The policy also exhibits strong generalization, successfully deploying zero-shot in unseen, outdoor environments while reaching speeds of up to 7.5m/s as well as stably flying in the super-dense forest. This work is released at https://github.com/Fanxing-LI/avoidance.

Weilin Zhao, Yuxiang Huang, Xu Han et al. · tsinghua

Speculative decoding is a widely used method that accelerates the generation process of large language models (LLMs) with no compromise in model performance. It achieves this goal by using an existing smaller model for drafting and then employing the target LLM to verify the draft in a low-cost parallel manner. Under such a drafting-verification framework, drafting efficiency has become a bottleneck in the final speedup of speculative decoding. Therefore, generating longer drafts at less cost can lead to better decoding speedup. To achieve this, we introduce Ouroboros, which can generate draft phrases to parallelize the drafting process and meanwhile lengthen drafts in a training-free manner. The experimental results on various typical text generation tasks show that Ouroboros can achieve speedups of up to $2.8\times$ over speculative decoding and $3.9\times$ over vanilla decoding, without fine-tuning draft and target models. The source code of Ouroboros is available at https://github.com/thunlp/Ouroboros.

Zhenshan Tan, Chenhan Lu, Yuxiang Huang et al.

Image Manipulation Localization (IML) aims to identify edited regions in an image. However, with the increasing use of modern image editing and generative models, many manipulations no longer exhibit obvious low-level artifacts. Instead, they often involve subtle but meaning-altering edits to an object's attributes, state, or relationships while remaining highly consistent with the surrounding content. This makes conventional IML methods less effective because they mainly rely on artifact detection rather than semantic sensitivity. To address this issue, we introduce Semantic Manipulation Localization (SML), a new task that focuses on localizing subtle semantic edits that significantly change image interpretation. We further construct a dedicated fine-grained benchmark for SML using a semantics-driven manipulation pipeline with pixel-level annotations. Based on this task, we propose TRACE (Targeted Reasoning of Attributed Cognitive Edits), an end-to-end framework that models semantic sensitivity through three progressively coupled components: semantic anchoring, semantic perturbation sensing, and semantic-constrained reasoning. Specifically, TRACE first identifies semantically meaningful regions that support image understanding, then injects perturbation-sensitive frequency cues to capture subtle edits under strong visual consistency, and finally verifies candidate regions through joint reasoning over semantic content and semantic scope. Extensive experiments show that TRACE consistently outperforms existing IML methods on our benchmark and produces more complete, compact, and semantically coherent localization results. These results demonstrate the necessity of moving beyond artifact-based localization and provide a new direction for image forensics in complex semantic editing scenarios.

MiniCPM Team, Chaojun Xiao, Yuxuan Li et al. · tencent-ai, tsinghua

This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and UltraChat v2, a comprehensive supervised fine-tuning dataset. These datasets enable satisfactory model performance to be achieved using just 8 trillion training tokens. Regarding training algorithms, we propose ModelTunnel v2 for efficient pre-training strategy search, and improve existing post-training methods by introducing chunk-wise rollout for load-balanced reinforcement learning and data-efficient tenary LLM, BitCPM. Regarding inference systems, we propose CPM.cu that integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding. To meet diverse on-device requirements, MiniCPM4 is available in two versions, with 0.5B and 8B parameters, respectively. Furthermore, we construct a hybrid reasoning model, MiniCPM4.1, which can be used in both deep reasoning mode and non-reasoning mode. Evaluation results demonstrate that MiniCPM4 and MiniCPM4.1 outperform similar-sized open-source models across benchmarks, with the 8B variants showing significant speed improvements on long sequence understanding and generation.

Yuxiang Huang, Nuno M. T. Gonçalves, Federico Alvetreti et al.

Current hierarchical attention methods, such as NSA and InfLLMv2, select the top-k relevant key-value (KV) blocks based on coarse attention scores and subsequently apply fine-grained softmax attention on the selected tokens. However, the top-k operation assumes the number of relevant tokens for any query is fixed and it precludes the gradient flow between the sparse and dense stages. In this work, we propose DashAttention (Differentiable and Adaptive Sparse Hierarchical Attention), which leverages the adaptively sparse $α$-entmax transformation to select a variable number of blocks according to the current query in the first stage. This in turn provides a prior for the second-stage softmax attention, keeping the entire hierarchy fully differentiable. Contrary to other hierarchical attention methods, we show that DashAttention is non-dispersive, translating to better long-context modeling ability. Experiments with large language models (LLMs) show that DashAttention achieves comparable accuracy as full attention with 75% sparsity and a better Pareto frontier than NSA and InfLLMv2, especially in high-sparsity regimes. We also provide an efficient, GPU-aware implementation of DashAttention in Triton, which achieves a speedup of up to over FlashAttention-3 at inference time. Overall, DashAttention offers a cost-effective strategy to model long contexts.

Weilin Zhao, Tengyu Pan, Xu Han et al. · tsinghua

Speculative sampling has emerged as an important technique for accelerating the auto-regressive generation process of large language models (LLMs) by utilizing a draft-then-verify mechanism to produce multiple tokens per forward pass. While state-of-the-art speculative sampling methods use only a single layer and a language modeling (LM) head as the draft model to achieve impressive layer compression, their efficiency gains are substantially reduced for large-vocabulary LLMs, such as Llama-3-8B with a vocabulary of 128k tokens. To address this, we present FR-Spec, a frequency-ranked speculative sampling framework that optimizes draft candidate selection through vocabulary space compression. By constraining the draft search to a frequency-prioritized token subset, our method reduces LM Head computation overhead by 75% while ensuring the equivalence of the final output distribution. Experiments across multiple datasets demonstrate an average of 1.12$\times$ speedup over the state-of-the-art speculative sampling method EAGLE-2. Code available at https://github.com/thunlp/FR-Spec.

Weilin Zhao, Zihan Zhou, Zhou Su et al. · tsinghua

Long-sequence processing is a critical capability for modern large language models. However, the self-attention mechanism in the standard Transformer architecture faces severe computational and memory bottlenecks when processing long sequences. While trainable sparse attention methods offer a promising solution, existing approaches such as NSA introduce excessive extra parameters and disrupt the conventional \textit{pretrain-on-short, finetune-on-long} workflow, resulting in slow convergence and difficulty in acceleration. To overcome these limitations, we introduce dense-sparse switchable attention framework, termed as InfLLM-V2. InfLLM-V2 is a trainable sparse attention that seamlessly adapts models from short to long sequences. Specifically, InfLLM-V2 reuses dense attention parameters through parameter-free architecture modification, maintaining consistency between short and long sequence processing. Additionally, InfLLM-V2 ensures computational efficiency across all sequence lengths, by using dense attention for short inputs and smoothly transitioning to sparse attention for long sequences. To achieve practical acceleration, we further introduce an efficient implementation of InfLLM-V2 that significantly reduces the computational overhead. Our experiments on long-context understanding and chain-of-thought reasoning demonstrate that InfLLM-V2 is 4$\times$ faster than dense attention while retaining 98.1% and 99.7% of the performance, respectively. Based on the InfLLM-V2 framework, we have trained and open-sourced MiniCPM4.1 (https://huggingface.co/openbmb/MiniCPM4.1-8B), a hybrid reasoning model, providing a reproducible implementation for the research community.

Tianyu Yu, Zefan Wang, Chongyi Wang et al. · tsinghua

Multimodal Large Language Models (MLLMs) are undergoing rapid progress and represent the frontier of AI development. However, their training and inference efficiency have emerged as a core bottleneck in making MLLMs more accessible and scalable. To address the challenges, we present MiniCPM-V 4.5, an 8B parameter model designed for high efficiency and strong performance. We introduce three core improvements in model architecture, data strategy and training method: a unified 3D-Resampler model architecture for highly compact encoding over images and videos, a unified learning paradigm for document knowledge and text recognition without heavy data engineering, and a hybrid reinforcement learning strategy for proficiency in both short and long reasoning modes. Comprehensive experimental results in OpenCompass evaluation show that MiniCPM-V 4.5 surpasses widely used proprietary models such as GPT-4o-latest, and significantly larger open-source models such as Qwen2.5-VL 72B. Notably, the strong performance is achieved with remarkable efficiency. For example, on the widely adopted VideoMME benchmark, MiniCPM-V 4.5 achieves state-of-the-art performance among models under 30B size, using just 46.7\% GPU memory cost and 8.7\% inference time of Qwen2.5-VL 7B.

Yuxiang Huang, John Zelek

Identifying and segmenting moving objects from a moving monocular camera is difficult when there is unknown camera motion, different types of object motions and complex scene structures. To tackle these challenges, we take advantage of two popular branches of monocular motion segmentation approaches: point trajectory based and optical flow based methods, by synergistically fusing these two highly complementary motion cues at object level. By doing this, we are able to model various complex object motions in different scene structures at once, which has not been achieved by existing methods. We first obtain object-specific point trajectories and optical flow mask for each common object in the video, by leveraging the recent foundational models in object recognition, segmentation and tracking. We then construct two robust affinity matrices representing the pairwise object motion affinities throughout the whole video using epipolar geometry and the motion information provided by optical flow. Finally, co-regularized multi-view spectral clustering is used to fuse the two affinity matrices and obtain the final clustering. Our method shows state-of-the-art performance on the KT3DMoSeg dataset, which contains complex motions and scene structures. Being able to identify moving objects allows us to remove them for map building when using visual SLAM or SFM.

Yuxiang Huang, Mingye Li, Xu Han et al. · tsinghua

While long-context inference is crucial for advancing large language model (LLM) applications, its prefill speed remains a significant bottleneck. Current approaches, including sequence parallelism strategies and compute reduction through approximate attention mechanisms, still fall short of delivering optimal inference efficiency. This hinders scaling the inputs to longer sequences and processing long-context queries in a timely manner. To address this, we introduce APB, an efficient long-context inference framework that leverages multi-host approximate attention to enhance prefill speed by reducing compute and enhancing parallelism simultaneously. APB introduces a communication mechanism for essential key-value pairs within a sequence parallelism framework, enabling a faster inference speed while maintaining task performance. We implement APB by incorporating a tailored FlashAttn kernel alongside optimized distribution strategies, supporting diverse models and parallelism configurations. APB achieves speedups of up to 9.2x, 4.2x, and 1.6x compared with FlashAttn, RingAttn, and StarAttn, respectively, without any observable task performance degradation. We provide the implementation and experiment code of APB in https://github.com/thunlp/APB.

Adel ElZemity, Budi Arief, Shujun Li et al.

Bare-metal operational technology (OT) devices -- especially the microcontrollers running Modbus/TCP and CoAP at the base of industrial control systems -- have remained outside the reach of autonomous security attacks. Prior autonomous pentesting studies target Linux and web systems, whose shells and filesystems are familiar to LLM agents. Bare-metal OT has neither, so agents must reason directly over protocol fields and parser semantics. This requires new action-space designs and runtime controls, and opens new research questions about protocol-level exploit reasoning and its deployment envelope. We present APIOT (Autonomous Purple-teaming for Industrial OT), the first large language model (LLM) framework demonstrating an autonomous attack and remediation of bare-metal OT devices, achieving the full discovery -> exploitation -> patching -> verification cycle without step-by-step human intervention. We implemented and evaluated this framework on Zephyr RTOS firmware across heterogeneous industrial IoT (IIoT) topologies. Through 290 experiment runs spanning five frontier LLMs, three network topologies, two impairment levels, and guided versus unguided conditions, APIOT achieved a mission success rate of 90.0% on the full attack-remediation cycle. We found that the runtime governance layer (which we call an overseer) is a critical engineering variable: without it, agents exhibit systematic degenerate patterns, including repetition loops, missing crash verification, and reconnaissance deadlocks. Together, these findings carry two implications beyond our testbed. Attacker expertise is no longer the binding constraint on bare-metal OT exploitation, and defender threat models must now assume LLM-augmented adversaries capable of executing autonomous discovery-through-remediation cycles against industrial firmware.

Yuxiang Huang, John Zelek

Motion segmentation is a fundamental problem in computer vision and is crucial in various applications such as robotics, autonomous driving and action recognition. Recently, spectral clustering based methods have shown impressive results on motion segmentation in dynamic environments. These methods perform spectral clustering on motion affinity matrices to cluster objects or point trajectories in the scene into different motion groups. However, existing methods often need the number of motions present in the scene to be known, which significantly reduces their practicality. In this paper, we propose a unified model selection technique to automatically infer the number of motion groups for spectral clustering based motion segmentation methods by combining different existing model selection techniques together. We evaluate our method on the KT3DMoSeg dataset and achieve competitve results comparing to the baseline where the number of clusters is given as ground truth information.

Yuxiang Huang, Yuhao Chen, John Zelek

Detecting and segmenting moving objects from a moving monocular camera is challenging in the presence of unknown camera motion, diverse object motions and complex scene structures. Most existing methods rely on a single motion cue to perform motion segmentation, which is usually insufficient when facing different complex environments. While a few recent deep learning based methods are able to combine multiple motion cues to achieve improved accuracy, they depend heavily on vast datasets and extensive annotations, making them less adaptable to new scenarios. To address these limitations, we propose a novel monocular dense segmentation method that achieves state-of-the-art motion segmentation results in a zero-shot manner. The proposed method synergestically combines the strengths of deep learning and geometric model fusion methods by performing geometric model fusion on object proposals. Experiments show that our method achieves competitive results on several motion segmentation datasets and even surpasses some state-of-the-art supervised methods on certain benchmarks, while not being trained on any data. We also present an ablation study to show the effectiveness of combining different geometric models together for motion segmentation, highlighting the value of our geometric model fusion strategy.

Yuxiang Huang, Chaojun Xiao, Xu Han et al.

Trainable sparse attention has emerged as a promising solution to address the decoding efficiency bottleneck of LLMs in long-context processing, significantly saving memory accesses while minimally impacting task performance. However, existing sparse attention methods leave a crucial limitation unresolved: the size of the key-value (KV) cache remains unreduced, which constrains on-GPU batch sizes and throttles decoding throughput, especially in large-scale batched inference. In this paper, we show that trainable sparse attention naturally exhibits strong locality in token selection across adjacent decoding steps, thereby enabling KV cache offloading without altering the underlying attention computation. However, the inherent locality remains insufficient to achieve efficient offloading, as the transfer of selected KV pairs between the CPU and GPU continues to dominate the overall decoding cost. Building on this insight, we present NOSA, a trainable sparse attention framework designed to natively support KV cache offloading. NOSA introduces explicit locality constraints by decomposing token selection into query-aware and query-agnostic components, thereby reducing KV transfers while preserving the same attention computation as used during training. We pretrain a 1B-parameter model with NOSA and conduct extensive benchmarks, showing that it preserves near-lossless performance while achieving up to a 2.3x improvement in decoding throughput compared with the vanilla trainable sparse attention baseline (InfLLM-V2).

Yuxiang Huang, Yuhao Chen, John Zelek

Motion segmentation from a single moving camera presents a significant challenge in the field of computer vision. This challenge is compounded by the unknown camera movements and the lack of depth information of the scene. While deep learning has shown impressive capabilities in addressing these issues, supervised models require extensive training on massive annotated datasets, and unsupervised models also require training on large volumes of unannotated data, presenting significant barriers for both. In contrast, traditional methods based on optical flow do not require training data, however, they often fail to capture object-level information, leading to over-segmentation or under-segmentation. In addition, they also struggle in complex scenes with substantial depth variations and non-rigid motion, due to the overreliance of optical flow. To overcome these challenges, we propose an innovative hybrid approach that leverages the advantages of both deep learning methods and traditional optical flow based methods to perform dense motion segmentation without requiring any training. Our method initiates by automatically generating object proposals for each frame using foundation models. These proposals are then clustered into distinct motion groups using both optical flow and relative depth maps as motion cues. The integration of depth maps derived from state-of-the-art monocular depth estimation models significantly enhances the motion cues provided by optical flow, particularly in handling motion parallax issues. Our method is evaluated on the DAVIS-Moving and YTVOS-Moving datasets, and the results demonstrate that our method outperforms the best unsupervised method and closely matches with the state-of-theart supervised methods.

Ying Liu, Liucheng Guo, Valeri A. Makarov et al.

Automated hand gesture recognition has been a focus of the AI community for decades. Traditionally, work in this domain revolved largely around scenarios assuming the availability of the flow of images of the user hands. This has partly been due to the prevalence of camera-based devices and the wide availability of image data. However, there is growing demand for gesture recognition technology that can be implemented on low-power devices using limited sensor data instead of high-dimensional inputs like hand images. In this work, we demonstrate a hand gesture recognition system and method that uses signals from capacitive sensors embedded into the etee hand controller. The controller generates real-time signals from each of the wearer five fingers. We use a machine learning technique to analyse the time series signals and identify three features that can represent 5 fingers within 500 ms. The analysis is composed of a two stage training strategy, including dimension reduction through principal component analysis and classification with K nearest neighbour. Remarkably, we found that this combination showed a level of performance which was comparable to more advanced methods such as supervised variational autoencoder. The base system can also be equipped with the capability to learn from occasional errors by providing it with an additional adaptive error correction mechanism. The results showed that the error corrector improve the classification performance in the base system without compromising its performance. The system requires no more than 1 ms of computing time per input sample, and is smaller than deep neural networks, demonstrating the feasibility of agile gesture recognition systems based on this technology.