Huan Zhou

ByteDance Seed, Jiaze Chen, Tiantian Fan et al. · bytedance

We introduce Seed1.5-Thinking, capable of reasoning through thinking before responding, resulting in improved performance on a wide range of benchmarks. Seed1.5-Thinking achieves 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA, demonstrating excellent reasoning abilities in STEM and coding. Beyond reasoning tasks, the method demonstrates notable generalization across diverse domains. For instance, it surpasses DeepSeek R1 by 8% in win rate on non-reasoning tasks, indicating its broader applicability. Compared to other state-of-the-art reasoning models, Seed1.5-Thinking is a Mixture-of-Experts (MoE) model with a relatively small size, featuring 20B activated and 200B total parameters. As part of our effort to assess generalized reasoning, we develop two internal benchmarks, BeyondAIME and Codeforces, both of which will be publicly released to support future research. Model trial link: https://www.volcengine.com/experience/ark.

Gangwei Xu, Huan Zhou, Xin Yang

In this paper, we propose CGI-Stereo, a novel neural network architecture that can concurrently achieve real-time performance, competitive accuracy, and strong generalization ability. The core of our CGI-Stereo is a Context and Geometry Fusion (CGF) block which adaptively fuses context and geometry information for more effective cost aggregation and meanwhile provides feedback to feature learning to guide more effective contextual feature extraction. The proposed CGF can be easily embedded into many existing stereo matching networks, such as PSMNet, GwcNet and ACVNet. The resulting networks show a significant improvement in accuracy. Specially, the model which incorporates our CGF with ACVNet ranks $1^{st}$ on the KITTI 2012 and 2015 leaderboards among all the published methods. We further propose an informative and concise cost volume, named Attention Feature Volume (AFV), which exploits a correlation volume as attention weights to filter a feature volume. Based on CGF and AFV, the proposed CGI-Stereo outperforms all other published real-time methods on KITTI benchmarks and shows better generalization ability than other real-time methods. Code is available at https://github.com/gangweiX/CGI-Stereo.

Hong Liu, Wen-Dong Xu, Zi-Hao Shang et al.

Molecular subtypes of breast cancer are important references to personalized clinical treatment. For cost and labor savings, only one of the patient's paraffin blocks is usually selected for subsequent immunohistochemistry (IHC) to obtain molecular subtypes. Inevitable sampling error is risky due to tumor heterogeneity and could result in a delay in treatment. Molecular subtype prediction from conventional H&E pathological whole slide images (WSI) using AI method is useful and critical to assist pathologists pre-screen proper paraffin block for IHC. It's a challenging task since only WSI level labels of molecular subtypes can be obtained from IHC. Gigapixel WSIs are divided into a huge number of patches to be computationally feasible for deep learning. While with coarse slide-level labels, patch-based methods may suffer from abundant noise patches, such as folds, overstained regions, or non-tumor tissues. A weakly supervised learning framework based on discriminative patch selecting and multi-instance learning was proposed for breast cancer molecular subtype prediction from H&E WSIs. Firstly, co-teaching strategy was adopted to learn molecular subtype representations and filter out noise patches. Then, a balanced sampling strategy was used to handle the imbalance in subtypes in the dataset. In addition, a noise patch filtering algorithm that used local outlier factor based on cluster centers was proposed to further select discriminative patches. Finally, a loss function integrating patch with slide constraint information was used to finetune MIL framework on obtained discriminative patches and further improve the performance of molecular subtyping. The experimental results confirmed the effectiveness of the proposed method and our models outperformed even senior pathologists, with potential to assist pathologists to pre-screen paraffin blocks for IHC in clinic.

Kai Liu, Ziqing Du, Xucheng Wan et al.

Target speech extraction (TSE) systems are designed to extract target speech from a multi-talker mixture. The popular training objective for most prior TSE networks is to enhance reconstruction performance of extracted speech waveform. However, it has been reported that a TSE system delivers high reconstruction performance may still suffer low-quality experience problems in practice. One such experience problem is wrong speaker extraction (called speaker confusion, SC), which leads to strong negative experience and hampers effective conversations. To mitigate the imperative SC issue, we reformulate the training objective and propose two novel loss schemes that explore the metric of reconstruction improvement performance defined at small chunk-level and leverage the metric associated distribution information. Both loss schemes aim to encourage a TSE network to pay attention to those SC chunks based on the said distribution information. On this basis, we present X-SepFormer, an end-to-end TSE model with proposed loss schemes and a backbone of SepFormer. Experimental results on the benchmark WSJ0-2mix dataset validate the effectiveness of our proposals, showing consistent improvements on SC errors (by 14.8% relative). Moreover, with SI-SDRi of 19.4 dB and PESQ of 3.81, our best system significantly outperforms the current SOTA systems and offers the top TSE results reported till date on the WSJ0-2mix.

Jiashuo Liu, Jiayun Wu, Chunjie Wu et al.

The rapid proliferation of Large Language Models (LLMs) and diverse specialized benchmarks necessitates a shift from fragmented, task-specific metrics to a holistic, competitive ranking system that effectively aggregates performance across multiple ability dimensions. Primarily using static scoring, current evaluation methods are fundamentally limited. They struggle to determine the proper mix ratio across diverse benchmarks, and critically, they fail to capture a model's dynamic competitive fitness or its vulnerability when confronted with sequential, high-stakes tasks. To address this, we introduce the novel Competitive Swiss-System Dynamics (CSD) framework. CSD simulates a multi-round, sequential contest where models are dynamically paired across a curated sequence of benchmarks based on their accumulated win-loss record. And Monte Carlo Simulation ($N=100,000$ iterations) is used to approximate the statistically robust Expected Win Score ($E[S_m]$), which eliminates the noise of random pairing and early-round luck. Furthermore, we implement a Failure Sensitivity Analysis by parameterizing the per-round elimination quantity ($T_k$), which allows us to profile models based on their risk appetite--distinguishing between robust generalists and aggressive specialists. We demonstrate that CSD provides a more nuanced and context-aware ranking than traditional aggregate scoring and static pairwise models, representing a vital step towards risk-informed, next-generation LLM evaluation.

Seed, Baisheng Li, Banggu Wu et al.

We introduce Virtual Width Networks (VWN), a framework that delivers the benefits of wider representations without incurring the quadratic cost of increasing the hidden size. VWN decouples representational width from backbone width, expanding the embedding space while keeping backbone compute nearly constant. In our large-scale experiment, an 8-times expansion accelerates optimization by over 2 times for next-token and 3 times for next-2-token prediction. The advantage amplifies over training as both the loss gap grows and the convergence-speedup ratio increases, showing that VWN is not only token-efficient but also increasingly effective with scale. Moreover, we identify an approximately log-linear scaling relation between virtual width and loss reduction, offering an initial empirical basis and motivation for exploring virtual-width scaling as a new dimension of large-model efficiency.

Titouan Duston, Shuo Xin, Yang Sun et al.

We introduce AInsteinBench, a large-scale benchmark for evaluating whether large language model (LLM) agents can operate as scientific computing development agents within real research software ecosystems. Unlike existing scientific reasoning benchmarks which focus on conceptual knowledge, or software engineering benchmarks that emphasize generic feature implementation and issue resolving, AInsteinBench evaluates models in end-to-end scientific development settings grounded in production-grade scientific repositories. The benchmark consists of tasks derived from maintainer-authored pull requests across six widely used scientific codebases, spanning quantum chemistry, quantum computing, molecular dynamics, numerical relativity, fluid dynamics, and cheminformatics. All benchmark tasks are carefully curated through multi-stage filtering and expert review to ensure scientific challenge, adequate test coverage, and well-calibrated difficulty. By leveraging evaluation in executable environments, scientifically meaningful failure modes, and test-driven verification, AInsteinBench measures a model's ability to move beyond surface-level code generation toward the core competencies required for computational scientific research.

Huan Zhou, Feng Xue, Yucong Li et al.

Achieving real-time and accuracy on embedded platforms has always been the pursuit of road segmentation methods. To this end, they have proposed many lightweight networks. However, they ignore the fact that roads are "stuff" (background or environmental elements) rather than "things" (specific identifiable objects), which inspires us to explore the feasibility of representing roads with low-level instead of high-level features. Surprisingly, we find that the primary stage of mainstream network models is sufficient to represent most pixels of the road for segmentation. Motivated by this, we propose a Low-level Feature Dominated Road Segmentation network (LFD-RoadSeg). Specifically, LFD-RoadSeg employs a bilateral structure. The spatial detail branch is firstly designed to extract low-level feature representation for the road by the first stage of ResNet-18. To suppress texture-less regions mistaken as the road in the low-level feature, the context semantic branch is then designed to extract the context feature in a fast manner. To this end, in the second branch, we asymmetrically downsample the input image and design an aggregation module to achieve comparable receptive fields to the third stage of ResNet-18 but with less time consumption. Finally, to segment the road from the low-level feature, a selective fusion module is proposed to calculate pixel-wise attention between the low-level representation and context feature, and suppress the non-road low-level response by this attention. On KITTI-Road, LFD-RoadSeg achieves a maximum F1-measure (MaxF) of 95.21% and an average precision of 93.71%, while reaching 238 FPS on a single TITAN Xp and 54 FPS on a Jetson TX2, all with a compact model size of just 936k parameters. The source code is available at https://github.com/zhouhuan-hust/LFD-RoadSeg.

Han Nie, Bin Luo, Jun Liu et al.

The ideal goal of image matching is to achieve stable and efficient performance in unseen domains. However, many existing learning-based optical-SAR image matching methods, despite their effectiveness in specific scenarios, exhibit limited generalization and struggle to adapt to practical applications. Repeatedly training or fine-tuning matching models to address domain differences is not only not elegant enough but also introduces additional computational overhead and data production costs. In recent years, general foundation models have shown great potential for enhancing generalization. However, the disparity in visual domains between natural and remote sensing images poses challenges for their direct application. Therefore, effectively leveraging foundation models to improve the generalization of optical-SAR image matching remains challenge. To address the above challenges, we propose PromptMID, a novel approach that constructs modality-invariant descriptors using text prompts based on land use classification as priors information for optical and SAR image matching. PromptMID extracts multi-scale modality-invariant features by leveraging pre-trained diffusion models and visual foundation models (VFMs), while specially designed feature aggregation modules effectively fuse features across different granularities. Extensive experiments on optical-SAR image datasets from four diverse regions demonstrate that PromptMID outperforms state-of-the-art matching methods, achieving superior results in both seen and unseen domains and exhibiting strong cross-domain generalization capabilities. The source code will be made publicly available https://github.com/HanNieWHU/PromptMID.

Ziqing Du, Kai Liu, Xucheng Wan et al.

Overlapped speech detection (OSD) is critical for speech applications in scenario of multi-party conversion. Despite numerous research efforts and progresses, comparing with speech activity detection (VAD), OSD remains an open challenge and its overall performance is far from satisfactory. The majority of prior research typically formulates the OSD problem as a standard classification problem, to identify speech with binary (OSD) or three-class label (joint VAD and OSD) at frame level. In contrast to the mainstream, this study investigates the joint VAD and OSD task from a new perspective. In particular, we propose to extend traditional classification network with multi-exit architecture. Such an architecture empowers our system with unique capability to identify class using either low-level features from early exits or high-level features from last exit. In addition, two training schemes, knowledge distillation and dense connection, are adopted to further boost our system performance. Experimental results on benchmark datasets (AMI and DIHARD-III) validated the effectiveness and generality of our proposed system. Our ablations further reveal the complementary contribution of proposed schemes. With $F_1$ score of 0.792 on AMI and 0.625 on DIHARD-III, our proposed system outperforms several top performing models on these datasets, but also surpasses the current state-of-the-art by large margins across both datasets. Besides the performance benefit, our proposed system offers another appealing potential for quality-complexity trade-offs, which is highly preferred for efficient OSD deployment.

Kai Liu, Xucheng Wan, Ziqing Du et al.

As a practical alternative of speech separation, target speaker extraction (TSE) aims to extract the speech from the desired speaker using additional speaker cue extracted from the speaker. Its main challenge lies in how to properly extract and leverage the speaker cue to benefit the extracted speech quality. The cue extraction method adopted in majority existing TSE studies is to directly utilize discriminative speaker embedding, which is extracted from the pre-trained models for speaker verification. Although the high speaker discriminability is a most desirable property for speaker verification task, we argue that it may be too sophisticated for TSE. In this study, we propose that a simplified speaker cue with clear class separability might be preferred for TSE. To verify our proposal, we introduce several forms of speaker cues, including naive speaker embedding (such as, x-vector and xi-vector) and new speaker embeddings produced from sparse LDA-transform. Corresponding TSE models are built by integrating these speaker cues with SepFormer (one SOTA speech separation model). Performances of these TSE models are examined on the benchmark WSJ0-2mix dataset. Experimental results validate the effectiveness and generalizability of our proposal, showing up to 9.9% relative improvement in SI-SDRi. Moreover, with SI-SDRi of 19.4 dB and PESQ of 3.78, our best TSE system significantly outperforms the current SOTA systems and offers the top TSE results reported till date on the WSJ0-2mix.

Guanbao Liang, Yuanchen Bei, Sheng Zhou et al.

Automatic prompt optimization is a promising approach for adapting large language models (LLMs) to downstream tasks, yet existing methods typically search for a specific prompt specialized to a fixed task. This paradigm limits generalization across heterogeneous queries and prevents models from accumulating reusable prompting knowledge over time. In this paper, we propose MemAPO, a memory-driven framework that reconceptualizes prompt optimization as generalizable and self-evolving experience accumulation. MemAPO maintains a dual-memory mechanism that distills successful reasoning trajectories into reusable strategy templates while organizing incorrect generations into structured error patterns that capture recurrent failure modes. Given a new prompt, the framework retrieves both relevant strategies and failure patterns to compose prompts that promote effective reasoning while discouraging known mistakes. Through iterative self-reflection and memory editing, MemAPO continuously updates its memory, enabling prompt optimization to improve over time rather than restarting from scratch for each task. Experiments on diverse benchmarks show that MemAPO consistently outperforms representative prompt optimization baselines while substantially reducing optimization cost.

Huan Zhou, Shi Gong, Yu Zhou et al.

This paper aims to address the problem of anomaly discovery in semantic segmentation. Our key observation is that semantic classification plays a critical role in existing approaches, while the incorrectly classified pixels are easily regarded as anomalies. Such a phenomenon frequently appears and is rarely discussed, which significantly reduces the performance of anomaly discovery. To this end, we propose a novel Distillation Comparison Network (DiCNet). It comprises of a teacher branch which is a semantic segmentation network that removed the semantic classification head, and a student branch that is distilled from the teacher branch through a distribution distillation. We show that the distillation guarantees the semantic features of the two branches hold consistency in the known classes, while reflect inconsistency in the unknown class. Therefore, we leverage the semantic feature discrepancy between the two branches to discover the anomalies. DiCNet abandons the semantic classification head in the inference process, and hence significantly alleviates the issue caused by incorrect semantic classification. Extensive experimental results on StreetHazards dataset and BDD-Anomaly dataset are conducted to verify the superior performance of DiCNet. In particular, DiCNet obtains a 6.3% improvement in AUPR and a 5.2% improvement in FPR95 on StreetHazards dataset, achieves a 4.2% improvement in AUPR and a 6.8% improvement in FPR95 on BDD-Anomaly dataset. Codes are available at https://github.com/zhouhuan-hust/DiCNet.

Xucheng Wan, Naijun Zheng, Kai Liu et al.

Contextualized ASR models have been demonstrated to effectively improve the recognition accuracy of uncommon phrases when a predefined phrase list is available. However, these models often struggle with bilingual settings, which are prevalent in code-switching speech recognition. In this study, we make the initial attempt to address this challenge by introducing a Cross-lingual Contextual Biasing(XCB) module. Specifically, we augment a pre-trained ASR model for the dominant language by integrating an auxiliary language biasing module and a supplementary language-specific loss, aimed at enhancing the recognition of phrases in the secondary language. Experimental results conducted on our in-house code-switching dataset have validated the efficacy of our approach, demonstrating significant improvements in the recognition of biasing phrases in the secondary language, even without any additional inference overhead. Additionally, our proposed system exhibits both efficiency and generalization when is applied by the unseen ASRU-2019 test set.

Xucheng Wan, Kai Liu, Ziqing Du et al.

To address the monaural speech enhancement problem, numerous research studies have been conducted to enhance speech via operations either in time-domain on the inner-domain learned from the speech mixture or in time--frequency domain on the fixed full-band short time Fourier transform (STFT) spectrograms. Very recently, a few studies on sub-band based speech enhancement have been proposed. By enhancing speech via operations on sub-band spectrograms, those studies demonstrated competitive performances on the benchmark dataset of DNS2020. Despite attractive, this new research direction has not been fully explored and there is still room for improvement. As such, in this study, we delve into the latest research direction and propose a sub-band based speech enhancement system with perceptually-motivated optimization and dual transformations, called PT-FSE. Specially, our proposed PT-FSE model improves its backbone, a full-band and sub-band fusion model, by three efforts. First, we design a frequency transformation module that aims to strengthen the global frequency correlation. Then a temporal transformation is introduced to capture long range temporal contexts. Lastly, a novel loss, with leverage of properties of human auditory perception, is proposed to facilitate the model to focus on low frequency enhancement. To validate the effectiveness of our proposed model, extensive experiments are conducted on the DNS2020 dataset. Experimental results show that our PT-FSE system achieves substantial improvements over its backbone, but also outperforms the current state-of-the-art while being 27\% smaller than the SOTA. With average NB-PESQ of 3.57 on the benchmark dataset, our system offers the best speech enhancement results reported till date.

He Wang, Pengcheng Guo, Xucheng Wan et al.

Automatic lip-reading (ALR) aims to automatically transcribe spoken content from a speaker's silent lip motion captured in video. Current mainstream lip-reading approaches only use a single visual encoder to model input videos of a single scale. In this paper, we propose to enhance lip-reading by incorporating multi-scale video data and multi-encoder. Specifically, we first propose a novel multi-scale lip motion extraction algorithm based on the size of the speaker's face and an Enhanced ResNet3D visual front-end (VFE) to extract lip features at different scales. For the multi-encoder, in addition to the mainstream Transformer and Conformer, we also incorporate the recently proposed Branchformer and E-Branchformer as visual encoders. In the experiments, we explore the influence of different video data scales and encoders on ALR system performance and fuse the texts transcribed by all ALR systems using recognizer output voting error reduction (ROVER). Finally, our proposed approach placed second in the ICME 2024 ChatCLR Challenge Task 2, with a 21.52% reduction in character error rate (CER) compared to the official baseline on the evaluation set.

Qintao Du, Ran Li, Weiyi Lv et al.

Conservation Voltage Reduction (CVR) relies on the effective coordination of slow-acting devices, such as OLTCs and CBs, and fast-acting devices, such as SVGs and PV inverters, typically implemented through a hierarchical multi-stage Volt-Var Control (VVC) spanning day-ahead scheduling, intra-day dispatch, and real-time control. However, existing sequential methods fail to account for the cas-cading impact of forecast errors on multi-stage decision-making. This oversight results in suboptimal day-ahead schedules for OLTCs and CBs that hinder the ef-fective coordination with fast-acting SVGs and inverters, inevitably driving a trade-off between real-time voltage security and CVR efficiency. To improve the Pareto front of this trade-off, this paper proposes a novel bi-level multi-timescale forecasting (Bi-MTF) framework for multi-stage VVC optimization. By integrating the downstream multi-stage VVC optimization into the upstream forecasting mod-els training, the decision-focused forecasting models are able to learn the trade-offs across temporal horizons. To solve the computationally challenging bi-level for-mulation, a modified sensitivity-driven integer L-shaped method is developed. It utilizes a hybrid gradient feedback mechanism that integrates numerical sensitivity analysis for discrete variables with analytical dual information for continuous fore-cast parameters to ensure tractability. Numerical results on a modified IEEE 33-bus system demonstrate that the proposed approach yields superior energy savings and operational safety compared to conventional MSE-based sequential paradigms. Specifically, as the capacity of fast-acting devices increases, the energy savings of the proposed method rise from 2.74% to 3.41%, which is far superior to the 1.50% to 1.76% achieved by conventional MSE-based sequential paradigms.

Jingzhe Ding, Shengda Long, Changxin Pu et al.

Recent advances in coding agents suggest rapid progress toward autonomous software development, yet existing benchmarks fail to rigorously evaluate the long-horizon capabilities required to build complete software systems. Most prior evaluations focus on localized code generation, scaffolded completion, or short-term repair tasks, leaving open the question of whether agents can sustain coherent reasoning, planning, and execution over the extended horizons demanded by real-world repository construction. To address this gap, we present NL2Repo Bench, a benchmark explicitly designed to evaluate the long-horizon repository generation ability of coding agents. Given only a single natural-language requirements document and an empty workspace, agents must autonomously design the architecture, manage dependencies, implement multi-module logic, and produce a fully installable Python library. Our experiments across state-of-the-art open- and closed-source models reveal that long-horizon repository generation remains largely unsolved: even the strongest agents achieve below 40% average test pass rates and rarely complete an entire repository correctly. Detailed analysis uncovers fundamental long-horizon failure modes, including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of interaction steps. NL2Repo Bench establishes a rigorous, verifiable testbed for measuring sustained agentic competence and highlights long-horizon reasoning as a central bottleneck for the next generation of autonomous coding agents.