Zihan Jiang

Xiaomi LLM-Core Team, Dong Zhang, Gang Wang et al.

Existing audio language models typically rely on task-specific fine-tuning to accomplish particular audio tasks. In contrast, humans are able to generalize to new audio tasks with only a few examples or simple instructions. GPT-3 has shown that scaling next-token prediction pretraining enables strong generalization capabilities in text, and we believe this paradigm is equally applicable to the audio domain. By scaling MiMo-Audio's pretraining data to over one hundred million of hours, we observe the emergence of few-shot learning capabilities across a diverse set of audio tasks. We develop a systematic evaluation of these capabilities and find that MiMo-Audio-7B-Base achieves SOTA performance on both speech intelligence and audio understanding benchmarks among open-source models. Beyond standard metrics, MiMo-Audio-7B-Base generalizes to tasks absent from its training data, such as voice conversion, style transfer, and speech editing. MiMo-Audio-7B-Base also demonstrates powerful speech continuation capabilities, capable of generating highly realistic talk shows, recitations, livestreaming and debates. At the post-training stage, we curate a diverse instruction-tuning corpus and introduce thinking mechanisms into both audio understanding and generation. MiMo-Audio-7B-Instruct achieves open-source SOTA on audio understanding benchmarks (MMSU, MMAU, MMAR, MMAU-Pro), spoken dialogue benchmarks (Big Bench Audio, MultiChallenge Audio) and instruct-TTS evaluations, approaching or surpassing closed-source models. Model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-Audio.

Xu Wen, Wanling Gao, Anzheng Li et al.

Classical machine learning (CML) occupies nearly half of machine learning pipelines in production applications. Unfortunately, it fails to utilize the state-of-the-practice devices fully and performs poorly. Without a unified framework, the hybrid deployments of deep learning (DL) and CML also suffer from severe performance and portability issues. This paper presents the design of a unified compiler, called CMLCompiler, for CML inference. We propose two unified abstractions: operator representations and extended computational graphs. The CMLCompiler framework performs the conversion and graph optimization based on two unified abstractions, then outputs an optimized computational graph to DL compilers or frameworks. We implement CMLCompiler on TVM. The evaluation shows CMLCompiler's portability and superior performance. It achieves up to 4.38$\times$ speedup on CPU, 3.31$\times$ speedup on GPU, and 5.09$\times$ speedup on IoT devices, compared to the state-of-the-art solutions -- scikit-learn, intel sklearn, and hummingbird. Our performance of CML and DL mixed pipelines achieves up to 3.04x speedup compared with cross-framework implementations. The project documents and source code are available at https://www.computercouncil.org/cmlcompiler.

Lichen Jia, Bowen Tang, Chenggang Wu et al.

The binary code similarity detection (BCSD) method measures the similarity of two binary executable codes. Recently, the learning-based BCSD methods have achieved great success, outperforming traditional BCSD in detection accuracy and efficiency. However, the existing studies are rather sparse on the adversarial vulnerability of the learning-based BCSD methods, which cause hazards in security-related applications. To evaluate the adversarial robustness, this paper designs an efficient and black-box adversarial code generation algorithm, namely, FuncFooler. FuncFooler constrains the adversarial codes 1) to keep unchanged the program's control flow graph (CFG), and 2) to preserve the same semantic meaning. Specifically, FuncFooler consecutively 1) determines vulnerable candidates in the malicious code, 2) chooses and inserts the adversarial instructions from the benign code, and 3) corrects the semantic side effect of the adversarial code to meet the constraints. Empirically, our FuncFooler can successfully attack the three learning-based BCSD models, including SAFE, Asm2Vec, and jTrans, which calls into question whether the learning-based BCSD is desirable.

LLM-Core Xiaomi, Bingquan Xia, Bowen Shen et al. · pku

We present MiMo-7B, a large language model born for reasoning tasks, with optimization across both pre-training and post-training stages. During pre-training, we enhance the data preprocessing pipeline and employ a three-stage data mixing strategy to strengthen the base model's reasoning potential. MiMo-7B-Base is pre-trained on 25 trillion tokens, with additional Multi-Token Prediction objective for enhanced performance and accelerated inference speed. During post-training, we curate a dataset of 130K verifiable mathematics and programming problems for reinforcement learning, integrating a test-difficulty-driven code-reward scheme to alleviate sparse-reward issues and employing strategic data resampling to stabilize training. Extensive evaluations show that MiMo-7B-Base possesses exceptional reasoning potential, outperforming even much larger 32B models. The final RL-tuned model, MiMo-7B-RL, achieves superior performance on mathematics, code and general reasoning tasks, surpassing the performance of OpenAI o1-mini. The model checkpoints are available at https://github.com/xiaomimimo/MiMo.

Shimao Chen, Zirui Liu, Zhiying Wu et al.

As the foundation of large language models (LLMs), self-attention module faces the challenge of quadratic time and memory complexity with respect to sequence length. FlashAttention accelerates attention computation and reduces its memory usage by leveraging the GPU memory hierarchy. A promising research direction is to integrate FlashAttention with quantization methods. This paper introduces INT-FlashAttention, the first INT8 quantization architecture compatible with the forward workflow of FlashAttention, which significantly improves the inference speed of FlashAttention on Ampere GPUs. We implement our INT-FlashAttention prototype with fully INT8 activations and general matrix-multiplication (GEMM) kernels, making it the first attention operator with fully INT8 input. As a general token-level post-training quantization framework, INT-FlashAttention is also compatible with other data formats like INT4, etc. Experimental results show INT-FlashAttention achieves 72% faster inference speed and 82% smaller quantization error compared to standard FlashAttention with FP16 and FP8 data format.

Yizhao Gao, Jianyu Wei, Qihao Zhang et al.

This work introduces Hybrid Sparse Attention (HySparse), a new architecture that interleaves each full attention layer with several sparse attention layers. While conceptually simple, HySparse strategically derives each sparse layer's token selection and KV caches directly from the preceding full attention layer. This architecture resolves two fundamental limitations of prior sparse attention methods. First, conventional approaches typically rely on additional proxies to predict token importance, introducing extra complexity and potentially suboptimal performance. In contrast, HySparse uses the full attention layer as a precise oracle to identify important tokens. Second, existing sparse attention designs often reduce computation without saving KV cache. HySparse enables sparse attention layers to reuse the full attention KV cache, thereby reducing both computation and memory. We evaluate HySparse on both 7B dense and 80B MoE models. Across all settings, HySparse consistently outperforms both full attention and hybrid SWA baselines. Notably, in the 80B MoE model with 49 total layers, only 5 layers employ full attention, yet HySparse achieves substantial performance gains while reducing KV cache storage by nearly 10x.

Xiaomi LLM-Core Team, Zihao Yue, Zhenru Lin et al. · pku

We open-source MiMo-VL-7B-SFT and MiMo-VL-7B-RL, two powerful vision-language models delivering state-of-the-art performance in both general visual understanding and multimodal reasoning. MiMo-VL-7B-RL outperforms Qwen2.5-VL-7B on 35 out of 40 evaluated tasks, and scores 59.4 on OlympiadBench, surpassing models with up to 78B parameters. For GUI grounding applications, it sets a new standard with 56.1 on OSWorld-G, even outperforming specialized models such as UI-TARS. Our training combines four-stage pre-training (2.4 trillion tokens) with Mixed On-policy Reinforcement Learning (MORL) integrating diverse reward signals. We identify the importance of incorporating high-quality reasoning data with long Chain-of-Thought into pre-training stages, and the benefits of mixed RL despite challenges in simultaneous multi-domain optimization. We also contribute a comprehensive evaluation suite covering 50+ tasks to promote reproducibility and advance the field. The model checkpoints and full evaluation suite are available at https://github.com/XiaomiMiMo/MiMo-VL.

Yuanhao Liu, Dehui Du, Zihan Jiang et al.

Causal inference permits us to discover covert relationships of various variables in time series. However, in most existing works, the variables mentioned above are the dimensions. The causality between dimensions could be cursory, which hinders the comprehension of the internal relationship and the benefit of the causal graph to the neural networks (NNs). In this paper, we find that causality exists not only outside but also inside the time series because it reflects a succession of events in the real world. It inspires us to seek the relationship between internal subsequences. However, the challenges are the hardship of discovering causality from subsequences and utilizing the causal natural structures to improve NNs. To address these challenges, we propose a novel framework called Mining Causal Natural Structure (MCNS), which is automatic and domain-agnostic and helps to find the causal natural structures inside time series via the internal causality scheme. We evaluate the MCNS framework and impregnation NN with MCNS on time series classification tasks. Experimental results illustrate that our impregnation, by refining attention, shape selection classification, and pruning datasets, drives NN, even the data itself preferable accuracy and interpretability. Besides, MCNS provides an in-depth, solid summary of the time series and datasets.

Zihan Jiang, Wanling Gao, Lei Wang et al.

In recent years, with the trend of applying deep learning (DL) in high performance scientific computing, the unique characteristics of emerging DL workloads in HPC raise great challenges in designing, implementing HPC AI systems. The community needs a new yard stick for evaluating the future HPC systems. In this paper, we propose HPC AI500 --- a benchmark suite for evaluating HPC systems that running scientific DL workloads. Covering the most representative scientific fields, each workload from HPC AI500 is based on real-world scientific DL applications. Currently, we choose 14 scientific DL benchmarks from perspectives of application scenarios, data sets, and software stack. We propose a set of metrics for comprehensively evaluating the HPC AI systems, considering both accuracy, performance as well as power and cost. We provide a scalable reference implementation of HPC AI500. HPC AI500 is a part of the open-source AIBench project, the specification and source code are publicly available from \url{http://www.benchcouncil.org/AIBench/index.html}.

Zihan Jiang, Fanghao Liu, Huawei Wang et al.

Computational complexity and storage requirements are crucial factors influencing the performance and efficiency of convolutional neural networks (CNNs) in resource-constrained environments. This paper presents a high-performance embedded target detection system based on FPGA and YOLOv3-Tiny, specifically designed for embedded artificial intelligence applications. By integrating lightweight CNN optimization techniques with hardware accelerator design, significant improvements are made in both computational efficiency and resource utilization. Key optimizations, including low-bit quantization, batch normalization fusion, and table lookup mapping, reduce model parameters and computational complexity. Additionally, an FPGA hardware accelerator with a pipelined architecture is developed to enhance the efficiency of convolution operations while minimizing off-chip data transmission through modular design and on-chip cache optimization. On the ZYNQ-XC7Z035 platform, the system achieves an inference latency of 0.211 seconds, outperforming comparable designs by 75.58% in speed. The system achieves an power efficiency of 10.11 GOPS/W, surpassing comparable designs by at least 29.45%. Furthermore, hardware resource utilization is reduced by up to 51.94% compared to similar systems. This study offers innovative design methodologies and practical application examples for the efficient deployment of deep learning models on embedded platforms.

Lin Sun, Guangxiang Zhao, Xiaoqi Jian et al.

The challenge of reducing the size of Large Language Models (LLMs) while maintaining their performance has gained significant attention. However, existing methods, such as model distillation and transfer learning, often fail to achieve high accuracy. To address this limitation, we introduce the Branch-Merge distillation approach, which enhances model compression through two phases: (1) the Branch Phase, where knowledge from a large teacher model is \textit{selectively distilled} into specialized student models via domain-specific supervised fine-tuning (SFT); And (2) the Merge Phase, where these student models are merged to enable cross-domain knowledge transfer and improve generalization. We validate our distillation approach using DeepSeek-R1 as the teacher and DeepSeek-R1-Distill-Qwen-32B as the student. The resulting merged model, TinyR1-32B-Preview, outperforms its counterpart DeepSeek-R1-Distill-Qwen-32B across multiple benchmarks, including Mathematics (+5.5 points), Coding (+4.4 points) and Science (+2.9 points), while achieving near-equal performance to DeepSeek-R1 on AIME 2024. The Branch-Merge distillation approach provides a scalable solution for creating smaller, high-performing LLMs with reduced computational cost and time.

Feiyu Wang, Guoan Wang, Yihao Zhang et al.

Quantization-Aware Training (QAT) integrates quantization into the training loop, enabling LLMs to learn robust low-bit representations, and is widely recognized as one of the most promising research directions. All current QAT research focuses on minimizing quantization error on full-precision models, where the full-precision accuracy acts as an upper bound (accuracy ceiling). No existing method has even attempted to surpass this ceiling. To break this ceiling, we propose a new paradigm: raising the ceiling (full-precision model), and then still quantizing it efficiently into 2 bits. We propose Fairy$\pm i$, the first 2-bit quantization framework for complex-valued LLMs. Specifically, our method leverages the representational advantages of the complex domain to boost full-precision accuracy. We map weights to the fourth roots of unity $\{\pm1, \pm i\}$, forming a perfectly symmetric and information-theoretically optimal 2-bit representation. Importantly, each quantized weight has either a zero real or imaginary part, enabling multiplication-free inference using only additions and element swaps. Experimental results show that Fairy$\pm i$ outperforms the ceiling of existing 2-bit quantization approaches in terms of both PPL and downstream tasks, while maintaining strict storage and compute efficiency. This work opens a new direction for building highly accurate and practical LLMs under extremely low-bit constraints.

Yunyou Huang, Nana Wang, Suqin Tang et al.

This paper quantitatively reveals the state-of-the-art and state-of-the-practice AI systems only achieve acceptable performance on the stringent conditions that all categories of subjects are known, which we call closed clinical settings, but fail to work in real-world clinical settings. Compared to the diagnosis task in the closed setting, real-world clinical settings pose severe challenges, and we must treat them differently. We build a clinical AI benchmark named Clinical AIBench to set up real-world clinical settings to facilitate researches. We propose an open, dynamic machine learning framework and develop an AI system named OpenClinicalAI to diagnose diseases in real-world clinical settings. The first versions of Clinical AIBench and OpenClinicalAI target Alzheimer's disease. In the real-world clinical setting, OpenClinicalAI significantly outperforms the state-of-the-art AI system. In addition, OpenClinicalAI develops personalized diagnosis strategies to avoid unnecessary testing and seamlessly collaborates with clinicians. It is promising to be embedded in the current medical systems to improve medical services.

Jiansong Li, Xiao Dong, Guangli Li et al.

The training of deep neural networks (DNNs) is usually memory-hungry due to the limited device memory capacity of DNN accelerators. Characterizing the memory behaviors of DNN training is critical to optimize the device memory pressures. In this work, we pinpoint the memory behaviors of each device memory block of GPU during training by instrumenting the memory allocators of the runtime system. Our results show that the memory access patterns of device memory blocks are stable and follow an iterative fashion. These observations are useful for the future optimization of memory-efficient training from the perspective of raw memory access patterns.

Wanling Gao, Fei Tang, Jianfeng Zhan et al.

Modern real-world application scenarios like Internet services consist of a diversity of AI and non-AI modules with huge code sizes and long and complicated execution paths, which raises serious benchmarking or evaluating challenges. Using AI components or micro benchmarks alone can lead to error-prone conclusions. This paper presents a methodology to attack the above challenge. We formalize a real-world application scenario as a Directed Acyclic Graph-based model and propose the rules to distill it into a permutation of essential AI and non-AI tasks, which we call a scenario benchmark. Together with seventeen industry partners, we extract nine typical scenario benchmarks. We design and implement an extensible, configurable, and flexible benchmark framework. We implement two Internet service AI scenario benchmarks based on the framework as proxies to two real-world application scenarios. We consider scenario, component, and micro benchmarks as three indispensable parts for evaluating. Our evaluation shows the advantage of our methodology against using component or micro AI benchmarks alone. The specifications, source code, testbed, and results are publicly available from \url{https://www.benchcouncil.org/aibench/scenario/}.

Fei Tang, Wanling Gao, Jianfeng Zhan et al.

Earlier-stage evaluations of a new AI architecture/system need affordable benchmarks. Only using a few AI component benchmarks like MLPerfalone in the other stages may lead to misleading conclusions. Moreover, the learning dynamics are not well understood, and the benchmarks' shelf-life is short. This paper proposes a balanced benchmarking methodology. We use real-world benchmarks to cover the factors space that impacts the learning dynamics to the most considerable extent. After performing an exhaustive survey on Internet service AI domains, we identify and implement nineteen representative AI tasks with state-of-the-art models. For repeatable performance ranking (RPR subset) and workload characterization (WC subset), we keep two subsets to a minimum for affordability. We contribute by far the most comprehensive AI training benchmark suite. The evaluations show: (1) AIBench Training (v1.1) outperforms MLPerfTraining (v0.7) in terms of diversity and representativeness of model complexity, computational cost, convergent rate, computation, and memory access patterns, and hotspot functions; (2) Against the AIBench full benchmarks, its RPR subset shortens the benchmarking cost by 64%, while maintaining the primary workload characteristics; (3) The performance ranking shows the single-purpose AI accelerator like TPU with the optimized TensorFlowframework performs better than that of GPUs while losing the latter's general support for various AI models. The specification, source code, and performance numbers are available from the AIBench homepage https://www.benchcouncil.org/aibench-training/index.html.

Wanling Gao, Fei Tang, Jianfeng Zhan et al.

Domain-specific software and hardware co-design is encouraging as it is much easier to achieve efficiency for fewer tasks. Agile domain-specific benchmarking speeds up the process as it provides not only relevant design inputs but also relevant metrics, and tools. Unfortunately, modern workloads like Big data, AI, and Internet services dwarf the traditional one in terms of code size, deployment scale, and execution path, and hence raise serious benchmarking challenges. This paper proposes an agile domain-specific benchmarking methodology. Together with seventeen industry partners, we identify ten important end-to-end application scenarios, among which sixteen representative AI tasks are distilled as the AI component benchmarks. We propose the permutations of essential AI and non-AI component benchmarks as end-to-end benchmarks. An end-to-end benchmark is a distillation of the essential attributes of an industry-scale application. We design and implement a highly extensible, configurable, and flexible benchmark framework, on the basis of which, we propose the guideline for building end-to-end benchmarks, and present the first end-to-end Internet service AI benchmark. The preliminary evaluation shows the value of our benchmark suite---AIBench against MLPerf and TailBench for hardware and software designers, micro-architectural researchers, and code developers. The specifications, source code, testbed, and results are publicly available from the web site \url{http://www.benchcouncil.org/AIBench/index.html}.

Zihan Jiang, Jiansong Li, Jiangfeng Zhan

In recent years, domain-specific hardware has brought significant performance improvements in deep learning (DL). Both industry and academia only focus on throughput when evaluating these AI accelerators, which usually are custom ASICs deployed in datacenter to speed up the inference phase of DL workloads. Pursuing higher hardware throughput such as OPS (Operation Per Second) using various optimizations seems to be their main design target. However, they ignore the importance of accuracy in the DL nature. Motivated by this, this paper argue that a single throughput metric can not comprehensively reflect the real-world performance of AI accelerators. To reveal this pitfall, we evaluates several frequently-used optimizations on a typical AI accelerator and quantifies their impact on accuracy and throughout under representative DL inference workloads. Based on our experimental results, we find that some optimizations cause significant loss on accuracy in some workloads, although it can improves the throughout. Furthermore, our results show the importance of end-to-end evaluation in DL.

Fanda Fan, Yunyou Huang, Lei Wang et al.

It is necessary for clinicians to comprehensively analyze patient information from different sources. Medical image fusion is a promising approach to providing overall information from medical images of different modalities. However, existing medical image fusion approaches ignore the semantics of images, making the fused image difficult to understand. In this work, we propose a new evaluation index to measure the semantic loss of fused image, and put forward a Fusion W-Net (FW-Net) for multimodal medical image fusion. The experimental results are promising: the fused image generated by our approach greatly reduces the semantic information loss, and has better visual effects in contrast to five state-of-art approaches. Our approach and tool have great potential to be applied in the clinical setting.