Searcharxiv⌕ Search

arXiv subjects

Rui Tian

Publications and source records attributed to Rui Tian.

At least 19 recordsLinked to original sources
arXiv2026

DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

Speculative decoding accelerates Large Language Model (LLM) inference by decoupling draft generation from target verification. While recent parallel drafters efficiently propose long token sequences in a single forward pass, they suffer from rapid acceptance decay due to a lack of inter-token dependencies. Furthermore, indiscriminately verifying these extended blocks wastes critical batch capacity on tokens with high rejection risks, severely degrading throughput in high-concurrency serving systems. We introduce DSpark, a speculative decoding framework that unifies high-throughput parallel generation with adaptive, load-aware verification. To maintain draft quality, DSpark utilizes a semi-autoregressive architecture, coupling a parallel backbone with a lightweight sequential module, to introduce intra-block dependency modeling and mitigate suffix decay. To optimize system efficiency, DSpark employs confidence-scheduled verification, dynamically tailoring the verification length for each request based on estimated prefix survival probabilities and engine-specific throughput profiles. On offline benchmarks across diverse domains, DSpark substantially improves the accepted length over state-of-the-art autoregressive and parallel drafters. When deployed within the DeepSeek-V4 serving system under live user traffic, DSpark successfully mitigates verification waste. Compared to the established production baseline (MTP-1), DSpark accelerates per-user generation speeds by 60 to 85 percent at matched throughput levels. More importantly, by preventing severe throughput degradation under strict interactivity constraints, it enables performance tiers that were previously unattainable, shifting the Pareto frontier of our serving system.

cs.AI↗
arXiv2026

Compile to Compress: Boosting Formal Theorem Provers by Compiler Outputs

Large language models (LLMs) have demonstrated significant potential in formal theorem proving, yet state-of-the-art performance often necessitates prohibitive test-time compute via massive roll-outs or extended context windows. In this work, we address this scalability bottleneck by exploiting an informative structure in formal verification: the observation that compilers map a vast space of diverse proof attempts to a compact set of structured failure modes. We introduce a learning-to-refine framework that leverages this compression to perform efficient learning and proof exploration. We perform tree search that corrects errors locally conditioned on explicit verifier feedback, thereby circumventing the costs associated with accumulating a long history of proof attempts. Extensive evaluations show that our method consistently amplifies the reasoning capabilities of base provers across varying scales. Notably, our approach achieves state-of-the-art performance on PutnamBench among publicly reported $\sim$8B and $\sim$32B parameter models under comparable test-time budgets, offering a scalable paradigm for next-generation verifier-guided reasoning.

cs.LG↗
arXiv2026

Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

While Mixture-of-Experts (MoE) scales capacity via conditional computation, Transformers lack a native primitive for knowledge lookup, forcing them to inefficiently simulate retrieval through computation. To address this, we introduce conditional memory as a complementary sparsity axis, instantiated via Engram, a module that modernizes classic $N$-gram embedding for O(1) lookup. By formulating the Sparsity Allocation problem, we uncover a U-shaped scaling law that optimizes the trade-off between neural computation (MoE) and static memory (Engram). Guided by this law, we scale Engram to 27B parameters, achieving superior performance over a strictly iso-parameter and iso-FLOPs MoE baseline. Most notably, while the memory module is expected to aid knowledge retrieval (e.g., MMLU +3.4; CMMLU +4.0), we observe even larger gains in general reasoning (e.g., BBH +5.0; ARC-Challenge +3.7) and code/math domains~(HumanEval +3.0; MATH +2.4). Mechanistic analyses reveal that Engram relieves the backbone's early layers from static reconstruction, effectively deepening the network for complex reasoning. Furthermore, by delegating local dependencies to lookups, it frees up attention capacity for global context, substantially boosting long-context retrieval (e.g., Multi-Query NIAH: 84.2 to 97.0). Finally, Engram establishes infrastructure-aware efficiency: its deterministic addressing enables runtime prefetching from host memory, incurring negligible overhead. We envision conditional memory as an indispensable modeling primitive for next-generation sparse models.

cs.CL↗
arXiv2026

Group-kernel auto-calibration and group-patch k-space reconstruction: Fast MRI with time-variant B0 kernels partitioned into time-invariant subsets

Purpose: Pushing MRI speed further demands more spatially-encoded information captured per unit time, e.g., by superimposing additional field modulations during oversampled readout. However, this can introduce calibration errors and increase reconstruction time. Thus, we propose a continuous field calibration approach and an efficient k-space reconstruction technique. Theory and Methods: Our auto-calibration generalizes GRAPPA kernels to explicitly extract continuous B0 modulation kernels, solving interpolation relationships between two ACS regions differing only in the extra field modulation. The k-space locations sharing the same instantaneous image-space modulation are grouped, so that subsets of time-invariant kernels can be separately estimated, as a generalized solution for Wave-CAIPI/FRONSAC-type scans. This view further inspires a k-space subregion-wise reconstruction technique, as an efficient alternative to conventional hybrid-space reconstruction. At 9.4T, FLASH accelerated by a local B0 coil array and Wave-CAIPI were tested with retrospective undersampling. Results: Artifact-free images were reconstructed, under diverse rapid B0 modulation schemes, reaching maximum acceleration factors of 8-fold in 2D and 14.6-fold in 3D. Some nonlinear gradients modulation schemes reach similar sampling efficiency as linear gradients modulation. The proposed reconstruction shows potentials in reducing reconstruction time. Conclusion: Rapid B0 modulations and widely-adopted parallel imaging can share a common mathematical framework, and consequently, achieve similarly-robust reconstructions. Specifically, for scans using dynamic B0 and static RF kernels, not only signal encoding, but also auto-calibration and reconstruction can be performed in k-space. This paves the way to robustly remove eddy currents, and explore more complex B0 modulation strategies towards ultimate MRI speed.

physics.med-ph↗
arXiv2026

Group-Patch Joint Compression: Compressing Dynamic B0 and Static RF Spatial Modulations Across k-Space Subregion Groups for Highly Accelerated MRI

Purpose: To accelerate MRI further, rapid B0 field modulations can be applied during oversampled readout to capture additional physical information, as in Wave-CAIPI/FRONSAC/local B0 coils modulation techniques. These methods, however, turn the Fourier readout into a non-Fourier-encoded dimension that cannot be reconstructed by FFT, posing significant reconstruction challenges especially in compressed-sensing or neural-network frameworks. Theory and methods: Because the rapid B0 modulations still vary slowly relative to the oversampled ADC dwell time, we exploit this encoding redundancy by compressing k-space patch-by-patch across subregions, each of which is jointly encoded by a distinct subset of B0 and RF (receive) spatial encoding functions. For each subset, a compression matrix is computed once and reused to compress all patches encoded by the same B0-RF spatial modulations. This can be implemented by feeding subsets of B0 and RF spatial encoding maps into an adapted conventional RF array compression algorithm, mimicking an expanded set of virtual receiver channels. This approach was evaluated on human brain scans at 9.4 T/3 T. Results: The proposed group-patch joint compression achieves substantially higher compression factors than conventional RF-only compression, while minimally compromising encoding efficiency. Typically, joint compression factors of 11x-20x led to negligible encoding loss, dramatically reducing reconstruction time and peak memory usage. For example, compressed-sensing reconstruction took 1.4-5.1 s/2D slice, 177 s-10.1 min/3D volume on a high-memory CPU node. Conclusion: Given joint encoding of dynamic B0 and static RF fields, compressing multidimensional k-space patches in separate groups outperforms compressing RF receivers alone. This substantially mitigates a fundamental computational bottleneck when combining rapid B0 and RF-receiver modulations.

physics.med-ph↗
arXiv2025

UniGen-1.5: Enhancing Image Generation and Editing through Reward Unification in Reinforcement Learning

We present UniGen-1.5, a unified multimodal large language model (MLLM) for advanced image understanding, generation and editing. Building upon UniGen, we comprehensively enhance the model architecture and training pipeline to strengthen the image understanding and generation capabilities while unlocking strong image editing ability. Especially, we propose a unified Reinforcement Learning (RL) strategy that improves both image generation and image editing jointly via shared reward models. To further enhance image editing performance, we propose a light Edit Instruction Alignment stage that significantly improves the editing instruction comprehension that is essential for the success of the RL training. Experimental results show that UniGen-1.5 demonstrates competitive understanding and generation performance. Specifically, UniGen-1.5 achieves 0.89 and 4.31 overall scores on GenEval and ImgEdit that surpass the state-of-the-art models such as BAGEL and reaching performance comparable to proprietary models such as GPT-Image-1.

cs.CV↗
arXiv2025

LLM as Attention-Informed NTM and Topic Modeling as long-input Generation: Interpretability and long-Context Capability

Topic modeling aims to produce interpretable topic representations and topic--document correspondences from corpora, but classical neural topic models (NTMs) remain constrained by limited representation assumptions and semantic abstraction ability. We study LLM-based topic modeling from both white-box and black-box perspectives. For white-box LLMs, we propose an attention-informed framework that recovers interpretable structures analogous to those in NTMs, including document-topic and topic-word distributions. This validates the view that LLM can serve as an attention-informed NTM. For black-box LLMs, we reformulate topic modeling as a structured long-input task and introduce a post-generation signal compensation method based on diversified topic cues and hybrid retrieval. Experiments show that recovered attention structures support effective topic assignment and keyword extraction, while black-box long-context LLMs achieve competitive or stronger performance than other baselines. These findings suggest a connection between LLMs and NTMs and highlight the promise of long-context LLMs for topic modeling.

cs.CL↗
arXiv2025

HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer

Recent advancements in image generative foundation models have prioritized quality improvements but often at the cost of increased computational complexity and inference latency. To address this critical trade-off, we introduce HiDream-I1, a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds. HiDream-I1 is constructed with a new sparse Diffusion Transformer (DiT) structure. Specifically, it starts with a dual-stream decoupled design of sparse DiT with dynamic Mixture-of-Experts (MoE) architecture, in which two separate encoders are first involved to independently process image and text tokens. Then, a single-stream sparse DiT structure with dynamic MoE architecture is adopted to trigger multi-model interaction for image generation in a cost-efficient manner. To support flexiable accessibility with varied model capabilities, we provide HiDream-I1 in three variants: HiDream-I1-Full, HiDream-I1-Dev, and HiDream-I1-Fast. Furthermore, we go beyond the typical text-to-image generation and remould HiDream-I1 with additional image conditions to perform precise, instruction-based editing on given images, yielding a new instruction-based image editing model namely HiDream-E1. Ultimately, by integrating text-to-image generation and instruction-based image editing, HiDream-I1 evolves to form a comprehensive image agent (HiDream-A1) capable of fully interactive image creation and refinement. To accelerate multi-modal AIGC research, we have open-sourced all the codes and model weights of HiDream-I1-Full, HiDream-I1-Dev, HiDream-I1-Fast, HiDream-E1 through our project websites: https://github.com/HiDream-ai/HiDream-I1 and https://github.com/HiDream-ai/HiDream-E1. All features can be directly experienced via https://vivago.ai/studio.

cs.CV↗
arXiv2025

UniGen: Enhanced Training & Test-Time Strategies for Unified Multimodal Understanding and Generation

We introduce UniGen, a unified multimodal large language model (MLLM) capable of image understanding and generation. We study the full training pipeline of UniGen from a data-centric perspective, including multi-stage pre-training, supervised fine-tuning, and direct preference optimization. More importantly, we propose a new Chain-of-Thought Verification (CoT-V) strategy for test-time scaling, which significantly boosts UniGen's image generation quality using a simple Best-of-N test-time strategy. Specifically, CoT-V enables UniGen to act as both image generator and verifier at test time, assessing the semantic alignment between a text prompt and its generated image in a step-by-step CoT manner. Trained entirely on open-source datasets across all stages, UniGen achieves state-of-the-art performance on a range of image understanding and generation benchmarks, with a final score of 0.78 on GenEval and 85.19 on DPG-Bench. Through extensive ablation studies, our work provides actionable insights and addresses key challenges in the full life cycle of building unified MLLMs, contributing meaningful directions to the future research.

cs.CV↗
arXiv2025

Observation of non-Hermitian bulk-boundary correspondence in non-chiral non-unitary quantum dynamics of single photons

The breakdown of conventional bulk-boundary correspondence, a cornerstone of topological physics, is one of counter-intuitive phenomena in non-Hermitian systems, that is deeply rooted in symmetry. In particular, preserved chiral symmetry is one of the key ingredients, which plays a pivotal role in determining non-Hermitian topology. Nevertheless, chiral symmetry breaking in non-Hermitian systems disrupts topological protection, modifies topological invariants, and substantially reshapes spectral and edge-state behavior. The corresponding fundamentally important bulk-boundary correspondence thus needs to be drastically reconstructed. However, it has so far eluded experimental efforts. Here, we theoretically predict and experimentally demonstrate the bulk-boundary correspondence of a one-dimensional (1D) non-Hermitian system with chiral symmetry breaking in discrete-time non-chiral non-unitary quantum walks of single photons. Through constructing a domain-wall configuration, we experimentally observe the photon localization at the interface of domain-wall structure, clearly indicating the presence of the topological edge mode. The appearance of that matches excellently with the prediction of our introduced non-chiral non-Bloch topological invariants pair. Our work thus unequivocally builds the non-Hermitian bulk-boundary correspondence as a general principle for studying topological physics in non-Hermitian systems with chiral symmetry breaking.

cond-mat.mes-hall↗
arXiv2025

Non-Hermiticity enhanced topological immunity of one-dimensional $p$-wave superconducting chain

Studying the immunity of topological superconductors against non-local disorder is one of the key issues in both fundamental researches and potential applications. Here, we demonstrate that the non-Hermiticity can enhance the robustness of topological edge states against non-local disorder. To illustrate that, we consider a one-dimensional (1D) generalized Kitaev model with the asymmetric hopping in the presence of disorder. It is shown that the region supporting Majorana zero modes (MZMs) against non-local disorder will be enlarged by the non-Hermiticity. Through both the numerical and analytical analyses, we show that non-Hermiticity can stabilize the topological superconducting (SC) phase against higher disorder strength. Our studies would offer new insights into the interplay between non-Hermiticity and topology.

cond-mat.supr-con↗
arXiv2024

Multi-Modality Driven LoRA for Adverse Condition Depth Estimation

The autonomous driving community is increasingly focused on addressing corner case problems, particularly those related to ensuring driving safety under adverse conditions (e.g., nighttime, fog, rain). To this end, the task of Adverse Condition Depth Estimation (ACDE) has gained significant attention. Previous approaches in ACDE have primarily relied on generative models, which necessitate additional target images to convert the sunny condition into adverse weather, or learnable parameters for feature augmentation to adapt domain gaps, resulting in increased model complexity and tuning efforts. Furthermore, unlike CLIP-based methods where textual and visual features have been pre-aligned, depth estimation models lack sufficient alignment between multimodal features, hindering coherent understanding under adverse conditions. To address these limitations, we propose Multi-Modality Driven LoRA (MMD-LoRA), which leverages low-rank adaptation matrices for efficient fine-tuning from source-domain to target-domain. It consists of two core components: Prompt Driven Domain Alignment (PDDA) and Visual-Text Consistent Contrastive Learning(VTCCL). During PDDA, the image encoder with MMD-LoRA generates target-domain visual representations, supervised by alignment loss that the source-target difference between language and image should be equal. Meanwhile, VTCCL bridges the gap between textual features from CLIP and visual features from diffusion model, pushing apart different weather representations (vision and text) and bringing together similar ones. Through extensive experiments, the proposed method achieves state-of-the-art performance on the nuScenes and Oxford RobotCar datasets, underscoring robustness and efficiency in adapting to varied adverse environments.

cs.CV↗
arXiv2024

Many-body interferometry of one-dimensional integrable systems

We propose using many-body Ramsey interferometry to measure non-equilibrium correlation functions of one-dimensional (1D) integrable systems. The 1D transverse-field Ising model, which is conjectured to equilibrate into non-thermal Gibbs ensemble (GGE) steady states, is studied. It is shown that retarded Green's functions, as opposed to ordinary spin-spin correlators considered previously, can convincingly distinguish between the GGE and thermal post-quench steady states, justifying the assumption of convergence towards the GGE as the system equilibrates. We also propose the experimental protocol for measuring the response functions with Ramsey interferometry, which can be used to distinguish between different post-quench phases of the model. Our proposal can be realized with current ultracold atom techniques, and opens up the possibility to study dynamics in non-thermal ensembles.

cond-mat.quant-gas↗
arXiv2024

Experimental investigation of the uncertainty relation in pre- and postselected systems

Uncertainty principle is one of the fundamental principles of quantum mechanics. Exploring such uncertainty relations in pre- and postselected (PPS) systems, where weak measurements on post-selected states have been used as a powerful tool for exploring the foundation of quantum mechanics, has so far eluded experimental effort. In this work, we experimentally investigate the Robertson-Heisenberg-type uncertainty relation for two incompatible observables in a PPS system. Through conducting a von Neumann-type measurement, uncertainty relations between two non-commuting Pauli observables are experimentally unveiled. Such uncertainty relations in the PPS system impose limitations on the joint sharp preparation of pre- and postselected states for two incompatible observables. Our experiments would be useful for the development of amplification techniques for precision measurements.

physics.optics↗
arXiv2024

REDUCIO! Generating 1K Video within 16 Seconds using Extremely Compressed Motion Latents

Commercial video generation models have exhibited realistic, high-fidelity results but are still restricted to limited access. One crucial obstacle for large-scale applications is the expensive training and inference cost. In this paper, we argue that videos contain significantly more redundant information than images, allowing them to be encoded with very few motion latents. Towards this goal, we design an image-conditioned VAE that projects videos into extremely compressed latent space and decode them based on content images. This magic Reducio charm enables 64x reduction of latents compared to a common 2D VAE, without sacrificing the quality. Building upon Reducio-VAE, we can train diffusion models for high-resolution video generation efficiently. Specifically, we adopt a two-stage generation paradigm, first generating a condition image via text-to-image generation, followed by text-image-to-video generation with the proposed Reducio-DiT. Extensive experiments show that our model achieves strong performance in evaluation. More importantly, our method significantly boosts the training and inference efficiency of video LDMs. Reducio-DiT is trained in just 3.2K A100 GPU hours in total and can generate a 16-frame 1024$\times$1024 video clip within 15.5 seconds on a single A100 GPU. Code released at https://github.com/microsoft/Reducio-VAE .

cs.CV↗
arXiv2024

ezyMRI: How to build an MRI machine from scratch -- Experience from a four-day hackathon

Nuclear magnetic resonance instruments are becoming available to the do-it-yourself community. The challenges encountered in the endeavor to build a magnetic resonance imaging instrument from scratch were confronted in a four-day hackathon at Singapore University of Technology and Design in spring 2024. One day was devoted to educational lectures and three days to system construction and testing. Seventy young researchers from all parts of the world formed six teams focusing on magnet, gradient coil, RF coil, console, system integration, and design, which together produced a working MRI instrument in three days. The different steps, encountered challenges, and their solutions are reported.

physics.med-ph↗
arXiv2024

Nonlinearity-induced dynamical self-organized twisted-bilayer lattices in Bose-Einstein condensates

Creating crystal bilayers twisted with respect to each other would lead to large periodic supercell structures, which can support a wide range of novel electron correlated phenomena, where the full understanding is still under debate. Here, we propose a new scheme to realize a nonlinearity-induced dynamical self-organized twisted-bilayer lattice in an atomic Bose-Einstein condensate (BEC). The key idea here is to utilize the nonlinear effect from the intrinsic atomic interactions to couple different layers and induce a dynamical self-organized supercell structure, dramatically distinct from the conventional wisdom to achieve the static twisted-bilayer lattices. To illustrate that, we study the dynamics of a two-component BEC and show that the nonlinear interaction effect naturally emerged in the Gross-Pitaevskii equation of interacting bosonic ultracold atoms can dynamically induce both periodic (commensurable) and aperiodic (incommensurable) moir\'{e} structures. One of the interesting moir\'{e} phenomena, i.e., the flat-band physics, is shown through investigating the dynamics of the wave packet of BEC. Our proposal can be implemented using available state-of-the-art experimental techniques and reveal a profound connection between the nonlinearity and twistronics in cold atom quantum simulators.

cond-mat.quant-gas↗