arXiv subjects

Qian Yao

Publications and source records attributed to Qian Yao.

12 recordsLinked to original sources

arXiv2026

NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction

We introduce NCP-ArchPreview, a latent-space language model that pushes autoregressive pretraining beyond standard next-token prediction (NTP). Alongside NTP, the model learns through Next Concept Prediction (NCP) to predict discrete concepts that span multiple tokens, introducing an explicit and more challenging concept-level objective while preserving standard token-level autoregressive generation. NCP-ArchPreview builds a latent space by constructing a product-quantized concept vocabulary directly from its hidden states, and subsequently learns to predict future concepts via a dedicated Concept Module. These predicted concepts are then fed back to the token level to guide subsequent generation, with NTP and NCP trained jointly end-to-end. We scale this architecture to 8.9B parameters and train it on 5.73T tokens from the Dolma-3 dataset, marking the largest demonstration of a latent-space language model to date. Remarkably, by consuming only 51.3% of the total training tokens, NCP-ArchPreview achieves the final pretraining loss of OLMo-3-7B. Following full pretraining, it outperforms OLMo-3-7B by 2.45 points on the downstream macro-average, including a notable 5.99-point gain on GSM8K. Controlled experiments isolate a clear progression of performance gains stemming from both the latent architecture and the NCP objective. Furthermore, utilizing only 85% of the standard computation, NCP-ArchPreview approaches the training loss of a strictly parameter-aligned 8.9B baseline. The learned latent space remains highly valuable after the pretraining stage: updating just the 17M-parameter VQ module yields a novel, lightweight interface for domain adaptation, while a simple injection of concept representations into a DFlash2 drafter improves the mean accepted length by 4.17% with negligible overhead.

cs.CL↗

arXiv2026

Intern-S2-Preview: Scientific Agentic Foundation Model

Scientific discovery increasingly requires AI systems that can reason over scientific evidence of heterogeneous modalities, interact with scientific tools and environments, and sustain progress across long task horizons. We present Intern-S2-Preview, a series of scientific agentic foundation models designed to support multimodal scientific understanding, reasoning, generation, and long-horizon tasks. The training pipeline begins with scientific multimodal pre-training over rendered scientific documents, interleaved image-text data, and diverse scientific corpora. Starting from the pretrained checkpoint, we apply a unified post-training pipeline consisting of supervised fine-tuning, scalable multi-task reinforcement learning (RL), black- and white-box agentic RL, and on-policy distillation. This pipeline is supported by practical techniques that improve rollout and training stability and efficiency, including partial rollout with off-policy correction, adaptive length regularization, online speculative decoding, robust multi-task optimization, and trace-aware experience assembly for agentic tasks. At the architecture level, Intern-S2-Preview-397B extends time series modelling from efficient long-sequence understanding to numerical forecasting, while Memory Decoder is studied as a separate memory-augmented path for rapid scientific specialization without modifying the frozen 397B backbone. Evaluations across scientific, multimodal, agentic, and general-purpose benchmarks show that Intern-S2-Preview-397B achieves competitive or leading results in multiple settings. The time series modules improve scientific signal understanding and forecasting on SciTS, while the separate Intern-MemDec-4B extension improves the Biology-Instructions average score from 56.92 to 60.32 without modifying the frozen 397B backbone.

cs.LG↗

arXiv2026

LHSDet: High-Resolution AI-Generated Image Detection via Visual Question Answering

Driven by advances in diffusion models and autoregressive models, the fidelity and resolution of AI-generated images now rival those of real images. However, existing AI-generated image detection methods often downsample the images, inevitably overlooking critical low-level texture details in high-resolution AI-generated images, therefore limiting their detection performance. In addition, the ceaseless emergence of unknown generative models makes large-scale pre-training datasets inaccessible. To address these challenges, we propose a novel high-resolution AI-generated image detector, termed LHSDet. Specifically, we formulate the AI-generated image detection task as a Visual Question Answering problem, leveraging a fine-tuned vision-language framework to fully exploit the complementary information between visual and textual modalities. Recognizing that the default visual encoder of existing vision-language models is not tailored for AI-generated image detection, we redesign a visual encoder to better capture both the low-level and high-level artifacts inherent in AI-generated images. Furthermore, we incorporate a semantic-level textual branch to enable multi-modal feature fusion and detection. Consequently, LHSDet employs a triple-branch architecture to extract complementary multi-modal features: a low-level visual branch that aggregates non-overlapping patches for local texture cues, a high-level visual branch based on SigLIP2 for global perception feature extraction, and a semantic-level textual branch that generates captions using BLIP-2. Extensive experimental results demonstrate that LHSDet achieves high detection accuracy and robust performance across diverse generative models, including both diffusion and autoregressive models.

cs.CV↗

arXiv2026

LMDeploy Accelerates Mixed-Precision LLM Inference with TurboMind

Mixed-precision inference techniques reduce the memory and computational demands of Large Language Models (LLMs) by applying hybrid precision formats to model weights, activations, and KV caches. However, existing systems struggle to (i) automatically generalize across diverse hardware architectures and precision formats, often requiring fragmented, hand-tuned kernels, and (ii) fully exploit available memory and compute resources, often causing performance bottlenecks. To address these problems, we propose TurboMind, a generalizable and efficient mixed-precision LLM inference engine of LMDeploy. TurboMind is built around two hardware-aware mixed-precision pipelines: A General Matrix Multiply (GEMM) pipeline that optimizes matrix operations through offline weight packing and online acceleration, and an attention pipeline that enables efficient attention computation with different Query, Key, and Value precision combinations. These pipelines are enabled by four key techniques: (i) Hardware-aware weight packing and (ii) adaptive head alignment for generalizability, and (iii) instruction-level parallelism and (iv) a KV memory loading pipeline for efficiency. We conduct comprehensive evaluations of LMDeploy powered by TurboMind across sixteen popular LLMs and four representative GPU architectures. Results demonstrate that LMDeploy achieves up to 61% lower serving latency (30% on average) and up to 156% higher throughput (58% on average) in mixed-precision workloads compared to existing mixed-precision frameworks, establishing consistent performance improvements across all tested configurations and hardware types. This work is open-sourced and publicly available at https://github.com/InternLM/lmdeploy.

cs.DC↗

arXiv2026

Intern-S1-Pro: Scientific Multimodal Foundation Model at Trillion Scale

We introduce Intern-S1-Pro, the first one-trillion-parameter scientific multimodal foundation model. Scaling to this unprecedented size, the model delivers a comprehensive enhancement across both general and scientific domains. Beyond stronger reasoning and image-text understanding capabilities, its intelligence is augmented with advanced agent capabilities. Simultaneously, its scientific expertise has been vastly expanded to master over 100 specialized tasks across critical science fields, including chemistry, materials, life sciences, and earth sciences. Achieving this massive scale is made possible by the robust infrastructure support of XTuner and LMDeploy, which facilitates highly efficient Reinforcement Learning (RL) training at the 1-trillion parameter level while ensuring strict precision consistency between training and inference. By seamlessly integrating these advancements, Intern-S1-Pro further fortifies the fusion of general and specialized intelligence, working as a Specializable Generalist, demonstrating its position in the top tier of open-source models for general capabilities, while outperforming proprietary models in the depth of specialized scientific tasks.

cs.LG↗

arXiv2025

SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generation

We propose SDAR, a Synergistic Diffusion-Autoregression paradigm that unifies the training efficiency of autoregressive models with the parallel inference capability of diffusion. Instead of costly end-to-end diffusion training, SDAR performs a lightweight paradigm conversion that transforms a well-trained autoregressive (AR) model into a blockwise diffusion model through brief, data-efficient adaptation. During inference, SDAR generates sequences autoregressively across blocks for global coherence while decoding all tokens within each block in parallel via a discrete diffusion process. Extensive experiments show that AR models remain substantially more compute-efficient than masked diffusion models, providing a strong foundation for adaptation. Building on this insight, SDAR achieves efficient AR-to-diffusion conversion with minimal cost, preserving AR-level performance while enabling parallel generation. Scaling studies across dense and Mixture-of-Experts architectures confirm that SDAR scales without compromise: larger models exhibit stronger robustness to block size and decoding thresholds, yielding greater speedups without accuracy loss. Beyond efficiency, SDAR demonstrates enhanced reasoning and domain adaptability. Our 30B MoE model surpasses its AR counterpart on challenging scientific reasoning benchmarks such as GPQA and ChemBench, and gains further improvements under test-time scaling methods like majority voting and pass@k. Together, these results establish SDAR as a practical paradigm that combines the strengths of autoregression and diffusion for scalable, high-throughput reasoning.

cs.LG↗

arXiv2024

Blockchain for Finance: A Survey

As an innovative technology for enhancing authenticity, security, and risk management, blockchain is being widely adopted in trade and finance systems. The unique capabilities of blockchain, such as immutability and transparency, enable new business models of distributed data storage, point-to-point transactions, and decentralized autonomous organizations. In this paper, we focus on blockchain-based securities trading, in which blockchain technology plays a vital role in financial services as it ultimately lifts trust and frees the need for third-party verification by using consensus-based verification. We investigate the 12 most popular blockchain platforms and elaborate on 6 platforms that are related to finance, seeking to provide a panorama of securities trading practices. Meanwhile, this survey provides a comprehensive summary of blockchain-based securities trading applications. We gather numerous practical applications of blockchain-based securities trading and categorize them into four distinct categories. For each category, we introduce a typical example and explain how blockchain contributes to solving the key problems faced by FinTech companies and researchers. Finally, we provide interesting observations ranging from mainstream blockchain-based financial institutions to security issues of decentralized finance applications, aiming to picture the current blockchain ecosystem in finance.

cs.CR↗

arXiv2023

The Li + CaF $\to$ Ca + LiF chemical reaction under cold conditions

The calcium monofluoride (CaF) molecule has emerged as a promising candidate for precision measurements, quantum simulation, and ultracold chemistry experiments. Inelastic and reactive collisions of laser cooled CaF molecules in optical tweezers have recently been reported and collisions of cold Li atoms with CaF are of current experimental interest. In this paper, we report ab initio electronic structure and full-dimensional quantum dynamical calculations of the Li + CaF $\to$ LiF + Ca chemical reaction. The electronic structure calculations are performed using the internally contracted multi-reference configuration-interaction method with Davidson correction (MRCI+Q). An analytic fit of the interaction energies is obtained using a many-body expansion method. A coupled-channel quantum reactive scattering approach implemented in hyperspherical coordinates is adopted for the scattering calculations under cold conditions. Results show that the Li + CaF reaction populates several low-lying vibrational levels and many rotational levels of the product LiF molecule and that the reaction is inefficient in the 1-100 mK regime allowing sympathetic cooling of CaF by collisions with cold Li atoms.

physics.chem-ph↗

arXiv2022

Causal Estimation of Position Bias in Recommender Systems Using Marketplace Instruments

Information retrieval systems, such as online marketplaces, news feeds, and search engines, are ubiquitous in today's digital society. They facilitate information discovery by ranking retrieved items on predicted relevance, i.e. likelihood of interaction (click, share) between users and items. Typically modeled using past interactions, such rankings have a major drawback: interaction depends on the attention items receive. A highly-relevant item placed outside a user's attention could receive little interaction. This discrepancy between observed interaction and true relevance is termed the position bias. Position bias degrades relevance estimation and when it compounds over time, it can silo users into false relevant items, causing marketplace inefficiencies. Position bias may be identified with randomized experiments, but such an approach can be prohibitive in cost and feasibility. Past research has also suggested propensity score methods, which do not adequately address unobserved confounding; and regression discontinuity designs, which have poor external validity. In this work, we address these concerns by leveraging the abundance of A/B tests in ranking evaluations as instrumental variables. Historical A/B tests allow us to access exogenous variation in rankings without manually introducing them, harming user experience and platform revenue. We demonstrate our methodology in two distinct applications at LinkedIn - feed ads and the People-You-May-Know (PYMK) recommender. The marketplaces comprise users and campaigns on the ads side, and invite senders and recipients on PYMK. By leveraging prior experimentation, we obtain quasi-experimental variation in item rankings that is orthogonal to user relevance. Our method provides robust position effect estimates that handle unobserved confounding well, greater generalizability, and easily extends to other information retrieval systems.

econ.EM↗

arXiv2020

Stereodynamic control of overlapping resonances in cold molecular collisions

Stereodynamic control of resonant molecular collisions has emerged as a new frontier in cold molecule research. Recent experimental studies have focused on weakly interacting molecular systems such as HD collisions with H$_2$, D$_2$ and He. We report here the possibility of such control in strongly interacting systems taking rotational relaxation in cold collisions of HCl and H$_2$. Using explicit quantum scattering calculations in full six dimensions it is shown that robust control of the collision dynamics is possible even when multiple (overlapping) shape-resonances coexist in a narrow energy range, indicating that cold stereochemistry offers great promise for many molecules beyond simple systems. We demonstrate a striking case where two prominent peaks in overlapping resonances are switched-off simultaneously by suitable alignment of the HCl molecule.

physics.chem-ph↗

arXiv2018

Ni Foam Assisted Synthesis of High Quality Hexagonal Boron Nitride with Large Domain Size and Controllable Thickness

The scalable synthesis of two-dimensional (2D) hexagonal boron nitride (h-BN) is of great interest for its numerous applications in novel electronic devices. Highly-crystalline h-BN films, with single-crystal sizes up to hundreds of microns, are demonstrated via a novel Ni foam assisted technique reported here for the first time. The nucleation density of h-BN domains can be significantly reduced due to the high boron solubility, as well as the large specific surface area of the Ni foam. The crystalline structure of the h-BN domains is found to be well aligned with, and therefore strongly dependent upon, the underlying Pt lattice orientation. Growth-time dependent experiments confirm the presence of a surface mediated self-limiting growth mechanism for monolayer h-BN on the Pt substrate. However, utilizing remote catalysis from the Ni foam, bilayer h-BN films can be synthesized breaking the self-limiting effect. This work provides further understanding of the mechanisms involved in the growth of h-BN and proposes a facile synthesis technique that may be applied to further applications in which control over the crystal alignment, and the numbers of layers is crucial.

cond-mat.mtrl-sci↗

arXiv2012

First-principles study on the Li storage performance of silicon clusters and graphene composite structure

This paper focuses on the performance of the storage of Li and the stability of the hybrid structure of different lattice planes of the silicon clusters and graphene by the first-principle theory. In this paper, we calculate the binding energy, adsorption energy and PDOS of the hybrid structure of the different height and size of the silicon clusters and graphene. We figure out that there can form strong Si-C bonds between the silicon cluster and graphene. Especially, the hybrid structure of the silicon clusters with plane (111) and graphene performs best with the highest formation energy and the outstanding stability. According to the calculation of Li absorption energy, we conclude that the location of the silicon cluster near the graphene has higher possibility and higher absorption energy of the Li storage, with the reason that the charge transfer between the lithium and the carbon and the silicon. Because of the graphene used, the deformation of the interface of the silicon clusters can be obviously reduced during the absorption of Li, which brings about a good future for the hybrid structure using for the battery anode materials.

cond-mat.mes-hall↗