Ao Sun

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

Zekai Qu, Yinxu Pan, Ao Sun et al.

Reinforcement learning (RL) post-training has become a trending paradigm for enhancing the capabilities of large language models (LLMs). Most existing RL systems for LLMs operate in a fully synchronous manner, where training must wait for the rollout of an entire batch to complete. This design leads to severe inefficiencies, as extremely long trajectories can stall the entire rollout process and leave many GPUs idle. To address this issue, we propose Concurrency- Controlled Partial Rollout with Importance Sampling (CoPRIS), which mitigates long-tail inefficiencies by maintaining a fixed number of concurrent rollouts, early-terminating once sufficient samples are collected, and reusing unfinished trajectories in subsequent rollouts. To mitigate the impact of off-policy trajectories, we introduce Cross-stage Importance Sampling Correction, which concatenates buffered log probabilities from the previous policy with those recomputed under the current policy for importance sampling correction. Experiments on challenging mathematical reasoning benchmarks show that CoPRIS achieves up to 1.94x faster training while maintaining comparable or superior performance to synchronous RL systems. The code of CoPRIS is available at https://github.com/777pomingzi/CoPRIS.

Hongsong Wang, Ao Sun, Jie Gui et al.

Gesture recognition is an important research area in the field of computer vision. Most gesture recognition efforts focus on close-set scenarios, thereby limiting the capacity to effectively handle unseen or novel gestures. We aim to address class-incremental gesture recognition, which entails the ability to accommodate new and previously unseen gestures over time. Specifically, we introduce a Prototype-Guided Pseudo Feature Replay (PGPFR) framework for data-free class-incremental gesture recognition. This framework comprises four components: Pseudo Feature Generation with Batch Prototypes (PFGBP), Variational Prototype Replay (VPR) for old classes, Truncated Cross-Entropy (TCE) for new classes, and Continual Classifier Re-Training (CCRT). To tackle the issue of catastrophic forgetting, the PFGBP dynamically generates a diversity of pseudo features in an online manner, leveraging class prototypes of old classes along with batch class prototypes of new classes. Furthermore, the VPR enforces consistency between the classifier's weights and the prototypes of old classes, leveraging class prototypes and covariance matrices to enhance robustness and generalization capabilities. The TCE mitigates the impact of domain differences of the classifier caused by pseudo features. Finally, the CCRT training strategy is designed to prevent overfitting to new classes and ensure the stability of features extracted from old classes. Extensive experiments conducted on two widely used gesture recognition datasets, namely SHREC 2017 3D and EgoGesture 3D, demonstrate that our approach outperforms existing state-of-the-art methods by 11.8\% and 12.8\% in terms of mean global accuracy, respectively. The code is available on https://github.com/sunao-101/PGPFR-3/.

Linlin Hu, Ao Sun, Shijie Hao et al.

Currently, most low-light image enhancement methods only consider information from a single view, neglecting the correlation between cross-view information. Therefore, the enhancement results produced by these methods are often unsatisfactory. In this context, there have been efforts to develop methods specifically for low-light stereo image enhancement. These methods take into account the cross-view disparities and enable interaction between the left and right views, leading to improved performance. However, these methods still do not fully exploit the interaction between left and right view information. To address this issue, we propose a model called Toward Sufficient Dual-View Interaction for Low-light Stereo Image Enhancement (SDI-Net). The backbone structure of SDI-Net is two encoder-decoder pairs, which are used to learn the mapping function from low-light images to normal-light images. Among the encoders and the decoders, we design a module named Cross-View Sufficient Interaction Module (CSIM), aiming to fully exploit the correlations between the binocular views via the attention mechanism. The quantitative and visual results on public datasets validate the superiority of our method over other related methods. Ablation studies also demonstrate the effectiveness of the key elements in our model.

Ao Sun, Xiaoyu Wang, Zhe Tan et al.

As Large Language Models (LLMs) serve a global audience, alignment must transition from enforcing universal consensus to respecting cultural pluralism. We demonstrate that dense models, when forced to fit conflicting value distributions, suffer from \textbf{Mean Collapse}, converging to a generic average that fails to represent diverse groups. We attribute this to \textbf{Cultural Sparsity}, where gradient interference prevents dense parameters from spanning distinct cultural modes. To resolve this, we propose \textbf{\textsc{CuMA}} (\textbf{Cu}ltural \textbf{M}ixture of \textbf{A}dapters), a framework that frames alignment as a \textbf{conditional capacity separation} problem. By incorporating demographic-aware routing, \textsc{CuMA} internalizes a \textit{Latent Cultural Topology} to explicitly disentangle conflicting gradients into specialized expert subspaces. Extensive evaluations on WorldValuesBench, Community Alignment, and PRISM demonstrate that \textsc{CuMA} achieves state-of-the-art performance, significantly outperforming both dense baselines and semantic-only MoEs. Crucially, our analysis confirms that \textsc{CuMA} effectively mitigates mean collapse, preserving cultural diversity. Our code is available at https://github.com/Throll/CuMA.

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.

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.

Ao Sun, Hongtao Zhang, Heng Zhou et al.

Linear Attention Large Language Models (LLMs) offer a compelling recurrent formulation that compresses context into a fixed-size state matrix, enabling constant-time inference. However, the internal dynamics of this compressed state remain largely opaque. In this work, we present a comprehensive study on the runtime state dynamics of state-of-the-art Linear Attention models. We uncover a fundamental phenomenon termed State Rank Stratification, characterized by a distinct spectral bifurcation among linear attention heads: while one group maintains an effective rank oscillating near zero, the other exhibits rapid growth that converges to an upper bound. Extensive experiments across diverse inference contexts reveal that these dynamics remain strikingly consistent, indicating that the identity of a head,whether low-rank or high-rank,is an intrinsic structural property acquired during pre-training, rather than a transient state dependent on the input data. Furthermore, our diagnostic probes reveal a surprising functional divergence: low-rank heads are indispensable for model reasoning, whereas high-rank heads exhibit significant redundancy. Leveraging this insight, we propose Joint Rank-Norm Pruning, a zero-shot strategy that achieves a 38.9\% reduction in KV-cache overhead while largely maintaining model accuracy.

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.

Yichen Lu, Wei Dai, Jiaen Liu et al.

LLM-based translation agents have achieved highly human-like translation results and are capable of handling longer and more complex contexts with greater efficiency. However, they are typically limited to text-only inputs. In this paper, we introduce ViDove, a translation agent system designed for multimodal input. Inspired by the workflow of human translators, ViDove leverages visual and contextual background information to enhance the translation process. Additionally, we integrate a multimodal memory system and long-short term memory modules enriched with domain-specific knowledge, enabling the agent to perform more accurately and adaptively in real-world scenarios. As a result, ViDove achieves significantly higher translation quality in both subtitle generation and general translation tasks, with a 28% improvement in BLEU scores and a 15% improvement in SubER compared to previous state-of-the-art baselines. Moreover, we introduce DoveBench, a new benchmark for long-form automatic video subtitling and translation, featuring 17 hours of high-quality, human-annotated data. Our code is available here: https://github.com/pigeonai-org/ViDove

Ao Sun, Yuanyuan Yuan, Pingchuan Ma et al.

Concept Bottleneck Models (CBMs) aim to deliver interpretable and interventionable predictions by bridging features and labels with human-understandable concepts. While recent CBMs show promising potential, they suffer from information leakage, where unintended information beyond the concepts (either when concepts are represented with probabilities or binary states) are leaked to the subsequent label prediction. Consequently, distinct classes are falsely classified via indistinguishable concepts, undermining the interpretation and intervention of CBMs. This paper alleviates the information leakage issue by introducing label supervision in concept predication and constructing a hierarchical concept set. Accordingly, we propose a new paradigm of CBMs, namely SupCBM, which achieves label predication via predicted concepts and a deliberately-designed intervention matrix. SupCBM focuses on concepts that are mostly relevant to the predicted label and only distinguishes classes when different concepts are presented. Our evaluations show that SupCBM outperforms SOTA CBMs over diverse datasets. It also manifests better generality across different backbone models. With proper quantification of information leakage in different CBMs, we demonstrate that SupCBM significantly reduces the information leakage.

Yiran Qin, Ao Sun, Yuze Hong et al.

Navigating unfamiliar environments presents significant challenges for household robots, requiring the ability to recognize and reason about novel decoration and layout. Existing reinforcement learning methods cannot be directly transferred to new environments, as they typically rely on extensive mapping and exploration, leading to time-consuming and inefficient. To address these challenges, we try to transfer the logical knowledge and the generalization ability of pre-trained foundation models to zero-shot navigation. By integrating a large vision-language model with a diffusion network, our approach named \mname ~constructs a visual predictor that continuously predicts the agent's potential observations in the next step which can assist robots generate robust actions. Furthermore, to adapt the temporal property of navigation, we introduce temporal historical information to ensure that the predicted image is aligned with the navigation scene. We then carefully designed an information fusion framework that embeds the predicted future frames as guidance into goal-reaching policy to solve downstream image navigation tasks. This approach enhances navigation control and generalization across both simulated and real-world environments. Through extensive experimentation, we demonstrate the robustness and versatility of our method, showcasing its potential to improve the efficiency and effectiveness of robotic navigation in diverse settings.

Pengyu Xu, Shijia Li, Ao Sun et al.

We propose OutboundEval, a comprehensive benchmark for evaluating large language models (LLMs) in expert-level intelligent outbound calling scenarios. Unlike existing methods that suffer from three key limitations - insufficient dataset diversity and category coverage, unrealistic user simulation, and inaccurate evaluation metrics - OutboundEval addresses these issues through a structured framework. First, we design a benchmark spanning six major business domains and 30 representative sub-scenarios, each with scenario-specific process decomposition, weighted scoring, and domain-adaptive metrics. Second, we develop a large-model-driven User Simulator that generates diverse, persona-rich virtual users with realistic behaviors, emotional variability, and communication styles, providing a controlled yet authentic testing environment. Third, we introduce a dynamic evaluation method that adapts to task variations, integrating automated and human-in-the-loop assessment to measure task execution accuracy, professional knowledge application, adaptability, and user experience quality. Experiments on 12 state-of-the-art LLMs reveal distinct trade-offs between expert-level task completion and interaction fluency, offering practical insights for building reliable, human-like outbound AI systems. OutboundEval establishes a practical, extensible, and domain-oriented standard for benchmarking LLMs in professional applications.

Shiju Wang, Yujie Wang, Ao Sun et al.

Long context training is crucial for LLM's context extension. Existing schemes, such as sequence parallelism, incur substantial communication overhead. Pipeline parallelism (PP) reduces this cost, but its effectiveness hinges on partitioning granularity. Batch-level PP dividing input samples exhibits high memory consumption in long-context scenario, whereas token-level PP splitting sequences into slices alleviates memory overhead but may incur hardware under-utilization. This trade-off motivates adaptively selecting PP granularity to match resource and workload characteristics. Moreover, sequence length distribution of the real-world dataset exhibits skewness, posing a challenge on PP's workload balance and efficient scheduling. Current static PP scheduling methods overlook the variance of sequence length, leading to suboptimal performance. In this paper, we propose Elastic Pipeline Parallelism (EPP) that orchestrates token-level PP and batch-level PP to adapt to resource and workload heterogeneity. We build InfiniPipe, a distributed training system that unleashes the potential of EPP via (1) a resource-aware and workload-balanced sequence processor that splits long sequences and packs short ones; and (2) a co-optimization methodology that jointly optimizes pipeline schedule and gradient checkpointing via a mechanism named stage-aware chunk-level adaptive checkpointing. Comprehensive experiments demonstrate that InfiniPipe achieves a 1.69x speedup over state-of-the-art systems.

Yingying Fan, Jingyuan Liu, Jinchi Lv et al.

We propose a new inference framework, named MOSAIC, for change-point detection in dynamic networks with the simultaneous low-rank and sparse-change structure. We establish the minimax rate of detection boundary, which relies on the sparsity of changes. We then develop an eigen-decomposition-based test with screened signals that approaches the minimax rate in theory, with only a minor logarithmic loss. For practical implementation of MOSAIC, we adjust the theoretical test by a novel residual-based technique, resulting in a pivotal statistic that converges to a standard normal distribution via the martingale central limit theorem under the null hypothesis and achieves full power under the alternative hypothesis. We also analyze the minimax rate of testing boundary for dynamic networks without the low-rank structure, which almost aligns with the results in high-dimensional mean-vector change-point inference. We showcase the effectiveness of MOSAIC and verify our theoretical results with several simulation examples and a real data application.

Yingying Fan, Jinchi Lv, Ao Sun et al.

Forecasting the Consumer Price Index (CPI) is an important yet challenging task in economics, where most existing approaches rely on low-frequency, survey-based data. With the recent advances of large language models (LLMs), there is growing potential to leverage high-frequency online text data for improved CPI prediction, an area still largely unexplored. This paper proposes LLM-CPI, an LLM-based approach for CPI prediction inference incorporating online text time series. We collect a large set of high-frequency online texts from a popularly used Chinese social network site and employ LLMs such as ChatGPT and the trained BERT models to construct continuous inflation labels for posts that are related to inflation. Online text embeddings are extracted via LDA and BERT. We develop a joint time series framework that combines monthly CPI data with LLM-generated daily CPI surrogates. The monthly model employs an ARX structure combining observed CPI data with text embeddings and macroeconomic variables, while the daily model uses a VARX structure built on LLM-generated CPI surrogates and text embeddings. We establish the asymptotic properties of the method and provide two forms of constructed prediction intervals. The finite-sample performance and practical advantages of LLM-CPI are demonstrated through both simulation and real data examples.

Ao Sun

In this work, we introduce CHAIR (Classifier of Hallucination As ImproveR), a supervised framework for detecting hallucinations by analyzing internal logits from each layer of every token. Our method extracts a compact set of features such as maximum, minimum, mean, standard deviation, and slope-from the token logits across all layers, enabling effective hallucination detection without overfitting. Experiments on TruthfulQA and MMLU datasets demonstrate that CHAIR significantly improves detection accuracy, particularly in zero-shot scenarios, showcasing its robustness and generalizability. Beyond hallucination detection, CHAIR highlights the potential of using internal representations for designing advanced decoding strategies. By leveraging patterns in logits, we suggest that more sophisticated models and adaptive decoding methods could further reduce hallucinations and enhance text completion quality. CHAIR not only offers a practical solution for detecting hallucinations but also lays the groundwork for exploring richer representations in LLMs to improve their factuality and coherence.

Ao Sun, Weilin Zhao, Xu Han et al.

Effective attention modules have played a crucial role in the success of Transformer-based large language models (LLMs), but the quadratic time and memory complexities of these attention modules also pose a challenge when processing long sequences. One potential solution for the long sequence problem is to utilize distributed clusters to parallelize the computation of attention modules across multiple devices (e.g., GPUs). However, adopting a distributed approach inevitably introduces extra memory overheads to store local attention results and incurs additional communication costs to aggregate local results into global ones. In this paper, we propose a distributed attention framework named ``BurstAttention'' to optimize memory access and communication operations at both the global cluster and local device levels. In our experiments, we compare BurstAttention with other competitive distributed attention solutions for long sequence processing. The experimental results under different length settings demonstrate that BurstAttention offers significant advantages for processing long sequences compared with these competitive baselines, reducing 40% communication overheads and achieving 1.37 X speedup during training 128K sequence length on 32 X A100.

Ao Sun, Pingchuan Ma, Yuanyuan Yuan et al.

EXplainable AI (XAI) is an essential topic to improve human understanding of deep neural networks (DNNs) given their black-box internals. For computer vision tasks, mainstream pixel-based XAI methods explain DNN decisions by identifying important pixels, and emerging concept-based XAI explore forming explanations with concepts (e.g., a head in an image). However, pixels are generally hard to interpret and sensitive to the imprecision of XAI methods, whereas "concepts" in prior works require human annotation or are limited to pre-defined concept sets. On the other hand, driven by large-scale pre-training, Segment Anything Model (SAM) has been demonstrated as a powerful and promotable framework for performing precise and comprehensive instance segmentation, enabling automatic preparation of concept sets from a given image. This paper for the first time explores using SAM to augment concept-based XAI. We offer an effective and flexible concept-based explanation method, namely Explain Any Concept (EAC), which explains DNN decisions with any concept. While SAM is highly effective and offers an "out-of-the-box" instance segmentation, it is costly when being integrated into defacto XAI pipelines. We thus propose a lightweight per-input equivalent (PIE) scheme, enabling efficient explanation with a surrogate model. Our evaluation over two popular datasets (ImageNet and COCO) illustrate the highly encouraging performance of EAC over commonly-used XAI methods.

Pingchuan Ma, Zongjie Li, Ao Sun et al.

As the popularity of large language models (LLMs) soars across various applications, ensuring their alignment with human values has become a paramount concern. In particular, given that LLMs have great potential to serve as general-purpose AI assistants in daily life, their subtly unethical suggestions become a serious and real concern. Tackling the challenge of automatically testing and repairing unethical suggestions is thus demanding. This paper introduces the first framework for testing and repairing unethical suggestions made by LLMs. We first propose ETHICSSUITE, a test suite that presents complex, contextualized, and realistic moral scenarios to test LLMs. We then propose a novel suggest-critic-reflect (SCR) process, serving as an automated test oracle to detect unethical suggestions. We recast deciding if LLMs yield unethical suggestions (a hard problem; often requiring human expertise and costly to decide) into a PCR task that can be automatically checked for violation. Moreover, we propose a novel on-the-fly (OTF) repairing scheme that repairs unethical suggestions made by LLMs in real-time. The OTF scheme is applicable to LLMs in a black-box API setting with moderate cost. With ETHICSSUITE, our study on seven popular LLMs (e.g., ChatGPT, GPT-4) uncovers in total 109,824 unethical suggestions. We apply our OTF scheme on two LLMs (Llama-13B and ChatGPT), which generates valid repair to a considerable amount of unethical ones, paving the way for more ethically conscious LLMs.