arXiv subjects

Kai Jiang

Publications and source records attributed to Kai Jiang.

At least 55 records · Page 3Linked to original sources

arXiv2026

Multimodal Continual Learning with MLLMs from Multi-scenario Perspectives

Multimodal large language models (MLLMs) deployed on devices must adapt to continuously changing visual scenarios such as variations in background and perspective, to effectively perform complex visual tasks. To investigate catastrophic forgetting under real-world scenario shifts, we construct a multimodal visual understanding dataset (MSVQA), covering four distinct scenarios and perspectives: high-altitude, underwater, low-altitude, and indoor environments. Furthermore, we propose UNIFIER (mUltimodal coNtInual learning with MLLMs From multi-scenarIo pERspectives), a continual learning (CL) framework designed to address visual discrepancies while learning different scenarios. Compared to existing CL methods, UNIFIER enables knowledge accumulation within the same scenario and mutual enhancement across different scenarios via Vision Representation Expansion (VRE) and Vision Consistency Constraint (VCC). Experimental results show that UNIFIER improves the last-step VQA scores by 2.70%~10.62% and the last-step F1 scores by 3.40%~7.69% compared to the state-of-the-art method, QUAD, in 20-step cross-scenario continual learning tasks. MSVQA dataset is available at https://huggingface.co/datasets/Kaij00/MSVQA.

cs.CV↗

arXiv2026

Exploring Strategies for Personalized Radiation Therapy Part IV: An Interaction-Picture Approach to Quantify the Abscopal Effect

We revisit the controversial "abscopal" effect in the context of Personalized Ultra-Fractionated Stereotactic Adaptive Radiotherapy (PULSAR). By allowing long interval between fractions, PULSAR may enhance systemic immune activation and increase the likelihood of abscopal responses compared with conventional daily fractionation. To quantify treatment-induced effects, we introduce an interaction-picture transformation adapted from quantum mechanics, which separates intrinsic tumor growth from radiation and immune-mediated perturbations. In this preliminary study, we tested this method to two preclinical bilateral tumor models (4T1 and MC38). Our model provides a quantitative measure of the interaction strength between primary and secondary tumors at the individual level, capturing dynamics over time rather than relying solely on cohort averages. This approach frames the abscopal effect as a continuous, stochastic phenomenon rather than a binary response. The framework is flexible for future studies, particularly in concurrent radiation and immunotherapy with PULSAR, where different radiation doses and fractionation schedules can be compared, and immune checkpoint inhibitors (ICIs) can be incorporated to further enhance systemic anti-tumor immunity. The framework can also help us make cross-study comparison of abscopal effects and standardizes the reporting of abscopal magnitude beyond simple statistical significance.

q-bio.QM↗

arXiv2026

SageBwd: A Trainable Low-bit Attention

Low-bit attention, such as SageAttention, has emerged as an effective approach for accelerating model inference, but its applicability to training remains poorly understood. In prior work, we introduced SageBwd, a trainable INT8 attention that quantizes six of seven attention matrix multiplications while preserving fine-tuning performance. However, SageBwd exhibited a persistent performance gap to full-precision attention (FPA) during pre-training. In this work, we investigate why this gap occurs and demonstrate that SageBwd matches full-precision attention during pretraining. Through experiments and theoretical analysis, we reach a few important insights and conclusions: (i) QK-norm is necessary for stable training at large tokens per step, (ii) quantization errors primarily arise from the backward-pass score gradient dS, (iii) reducing tokens per step enables SageBwd to match FPA performance in pre-training, and (iv) K-smoothing remains essential for training stability, while Q-smoothing provides limited benefit during pre-training.

cs.LG↗

arXiv2026

Exploring transition pathways in the Landau-Brazovskii model

The Landau-Brazovskii model provides a theoretical framework for describing various phases arising from competing short- and long-range interactions in many physical systems. In this work, we investigate phase transitions among various ordered phases within the three-dimensional Landau-Brazovskii model. We construct the phase diagram of this model, which encompasses eight distinct phases, and systematically compute the transition pathways connecting various metastable and stable states using the Landau-Brazovskii saddle dynamics. Along each transition pathway, the critical nucleus is identified with some detailed analyses of its shape, energy barrier, and Hessian eigenvalues. Furthermore, we explore how the transition state is influenced by model parameters, revealing systematic trends in critical nucleus sizes and energy barrier heights. Our results provide a comprehensive characterization of the nucleation mechanisms within the Landau-Brazovskii model and offer valuable insights into the structural transformations of modulated-phase systems.

cond-mat.mtrl-sci↗

arXiv2026

DiffusionNFT: Online Diffusion Reinforcement with Forward Process

Online reinforcement learning (RL) has been central to post-training language models, but its extension to diffusion models remains challenging due to intractable likelihoods. Recent works discretize the reverse sampling process to enable GRPO-style training, yet they inherit fundamental drawbacks, including solver restrictions, forward-reverse inconsistency, and complicated integration with classifier-free guidance (CFG). We introduce Diffusion Negative-aware FineTuning (DiffusionNFT), a new online RL paradigm that optimizes diffusion models directly on the forward process via flow matching. DiffusionNFT contrasts positive and negative generations to define an implicit policy improvement direction, naturally incorporating reinforcement signals into the supervised learning objective. This formulation enables training with arbitrary black-box solvers, eliminates the need for likelihood estimation, and requires only clean images rather than sampling trajectories for policy optimization. DiffusionNFT is up to $25\times$ more efficient than FlowGRPO in head-to-head comparisons, while being CFG-free. For instance, DiffusionNFT improves the GenEval score from 0.24 to 0.98 within 1k steps, while FlowGRPO achieves 0.95 with over 5k steps and additional CFG employment. By leveraging multiple reward models, DiffusionNFT significantly boosts the performance of SD3.5-Medium in every benchmark tested.

cs.LG↗

arXiv2026

Phason-Driven Diversity of Nucleation Pathways in Icosahedral Quasicrystals

The nucleation of quasicrystals remains a fundamental puzzle, primarily due to the absence of a periodic translational template. Here, we demonstrate that phasons - hidden degrees of freedom unique to quasiperiodic order - drive diverse nucleation pathways in icosahedral quasicrystals (IQCs). Combining a Landau free-energy model with the spring pair method, we compute distinct critical nuclei and their corresponding minimum energy paths. At low temperatures, a direct, symmetry-preserving pathway dominates. In contrast, higher temperatures promote a "symmetry detour" that reduces the nucleation barrier via a lower-symmetry critical nucleus. Remarkably, while the resulting bulk IQCs exhibit distinct real-space symmetries, they remain thermodynamically degenerate with identical diffraction patterns. We resolve this paradox within the high-dimensional projection framework, showing that phason shifts modulate real-space symmetry without altering bulk thermodynamics. Our findings establish phasons as the structural origin of pathway diversity, offering a new physical picture for the emergence of quasiperiodic order.

cond-mat.soft↗

arXiv2026

Compact LED-Based Displacement Sensing for Robot Fingers

In this paper, we introduce a sensor designed for integration in robot fingers, where it can provide information on the displacements induced by external contact. Our sensor uses LEDs to sense the displacement between two plates connected by a transparent elastomer; when a force is applied to the finger, the elastomer displaces and the LED signals change. We show that using LEDs as both light emitters an receivers in this context provides high sensitivity, allowing such an emitter and receiver pairs to detect very small displacements. We characterize the standalone performance of the sensor by testing the ability of a supervised learning model to predict complete force and torque data from its raw signals, and obtain a mean error between 0.05 and 0.07 N across the three directions of force applied to the finger. Our method allows for finger-size packaging with no amplification electronics, low cost manufacturing, easy integration into a complete hand, and high overload shear forces and bending torques, suggesting future applicability to complete manipulation tasks.

cs.RO↗

arXiv2026

SageAttention3: Microscaling FP4 Attention for Inference and An Exploration of 8-Bit Training

The efficiency of attention is important due to its quadratic time complexity. We enhance the efficiency of attention through two key contributions: First, we leverage the new FP4 Tensor Cores in Blackwell GPUs to accelerate attention computation. Our implementation achieves 1038 TOPS on RTX5090, which is a 5x speedup over the fastest FlashAttention on RTX5090. Experiments show that our FP4 attention can accelerate inference of various models in a plug-and-play way. Second, we pioneer low-bit attention to training tasks. Existing low-bit attention works like FlashAttention3 and SageAttention focus only on inference. However, the efficiency of training large models is also important. To explore whether low-bit attention can be effectively applied to training tasks, we design an accurate and efficient 8-bit attention for both forward and backward propagation. Experiments indicate that 8-bit attention achieves lossless performance in fine-tuning tasks but exhibits slower convergence in pretraining tasks. The code is available at https://github.com/thu-ml/SageAttention.

cs.LG↗

arXiv2026

Spectral Distribution of one-dimensional Photonic Quasicrystals: The Role of Irrational Numbers

In this paper, we construct a one-dimensional photonic quasicrystal by combining two incommensurate spatial harmonics, where the ratio of their periods is the irrational number β. We evaluate the photonic quasicrystal accurately by a generalized spectral method that embeds the quasiperiodic structure into a higher-dimensional periodic system. We study the spectral distribution of one-dimensional photonic quasicrystals and find some interesting phenomena. As the computational resolution N increases, there are more eigenvalues within finite frequency bandwidths, and the maximum localization always occurs at spectral gap edges for states near index N + 1. By varying βwithin the range of (0,1), we present a butterfly-shaped spectral structure with abundant band gaps. We find that the spectral structure factor Q (defined as I_{mg}/N, where I_{mg} is the maximum gap index) exhibits different linear patterns as βchanges: Q = 1 - βwhen β< βc, while Q = βwhen β> βc, where βc \approx 0.424 is the transition point. This linear relationship holds robustly in the strong quasiperiodic regime (βaway from 0 or 1) and is independent of the specific type of irrational number used. The relationship disappears (weak quasiperiodic regime) near β= 0 or β= 1. It demonstrates that the intrinsic spectral properties of one-dimensional photonic quasicrystals are fundamentally governed by the magnitude of the irrational parameter β.

physics.optics↗

arXiv2025

TurboDiffusion: Accelerating Video Diffusion Models by 100-200 Times

We introduce TurboDiffusion, a video generation acceleration framework that can speed up end-to-end diffusion generation by 100-200x while maintaining video quality. TurboDiffusion mainly relies on several components for acceleration: (1) Attention acceleration: TurboDiffusion uses low-bit SageAttention and trainable Sparse-Linear Attention (SLA) to speed up attention computation. (2) Step distillation: TurboDiffusion adopts rCM for efficient step distillation. (3) W8A8 quantization: TurboDiffusion quantizes model parameters and activations to 8 bits to accelerate linear layers and compress the model. In addition, TurboDiffusion incorporates several other engineering optimizations. We conduct experiments on the Wan2.2-I2V-14B-720P, Wan2.1-T2V-1.3B-480P, Wan2.1-T2V-14B-720P, and Wan2.1-T2V-14B-480P models. Experimental results show that TurboDiffusion achieves 100-200x speedup for video generation even on a single RTX 5090 GPU, while maintaining comparable video quality. The GitHub repository, which includes model checkpoints and easy-to-use code, is available at https://github.com/thu-ml/TurboDiffusion.

cs.CV↗

arXiv2025

Representation of quasi-periodic functions and Hausdorff-Young inequalities for Besicovitch almost periodic functions

For a class of $\mathbb{R}^d$-ations and $\mathbb{Z}^d$-actions on the $n$-dimensional torus $\mathbb{T}^n$, we characterize their unique ergodicity and establish a theorem of Weyl type. This result allows us to establish an isomorphism between the Banach algebra of quasi-periodic functions with spectrum in a given $\mathbb{Z}$-module and the Banach algebra of periodic functions on a torus. This, in return, allows us to give a very simple proof of Hausdorff-Young inequalities for Besicovitch almost periodic functions. The regularity of the parent function of a quasi-periodic function is also studied.

math.CA↗

arXiv2025

SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse-Linear Attention

In Diffusion Transformer (DiT) models, particularly for video generation, attention latency is a major bottleneck due to the long sequence length and the quadratic complexity. We find that attention weights can be separated into two parts: a small fraction of large weights with high rank and the remaining weights with very low rank. This naturally suggests applying sparse acceleration to the first part and low-rank acceleration to the second. Based on this finding, we propose SLA (Sparse-Linear Attention), a trainable attention method that fuses sparse and linear attention to accelerate diffusion models. SLA classifies attention weights into critical, marginal, and negligible categories, applying O(N^2) attention to critical weights, O(N) attention to marginal weights, and skipping negligible ones. SLA combines these computations into a single GPU kernel and supports both forward and backward passes. With only a few fine-tuning steps using SLA, DiT models achieve a 20x reduction in attention computation, resulting in significant acceleration without loss of generation quality. Experiments show that SLA reduces attention computation by 95% without degrading end-to-end generation quality, outperforming baseline methods. In addition, we implement an efficient GPU kernel for SLA, which yields a 13.7x speedup in attention computation and a 2.2x end-to-end speedup in video generation on Wan2.1-1.3B. The code is available at https://github.com/thu-ml/SLA.

cs.LG↗

arXiv2025

An image-guided high-precision research platform for ultra-high dose rate spinal cord toxicity studies

Objective: While FLASH radiotherapy is recognized for short-term normal tissue sparing, its durability in late-responding organs remains uncertain, limiting clinical adoption. With its clinical importance and steep dose-response, the spinal cord is an ideal model for evaluating FLASH effect on late toxicity. This work introduces a robust image-guided research platform for high-precision irradiation at both CONV and UHDR to enable FLASH late toxicity studies using a rat spinal cord model. Approach: A modified LINAC was employed to irradiate the C1-T2 rat spinal cord with 18 MeV UHDR and CONV beams. A custom rat immobilization device, a portable X-ray imaging system, and an ion-chamber-based UHDR output monitoring system were integrated to ensure accurate C1-T2 localization and precise dose delivery. A Monte Carlo (MC) dose engine was developed to provide accurate dosimetry and support interpretation of in vivo results. Scintillator measurements were performed within the spinal cord to verify MC results and the precision of our platform. Results: We achieved submillimeter C1-T2 setup accuracy and maintained submillimeter intrafraction motion. Ion chamber readings showed linear correlation with UHDR output. MC indicated uniform irradiation along the central ~13 mm cord. Our CONV beam exhibited distribution close to that of the UHDR beam, with difference <3%, isolating dose rate as the only variable. Scintillator-measured dose agreed with MC within 4%, confirming both MC accuracy and the platform's high-precision delivery. Significance: We developed the first comprehensive, image-guided preclinical platform for accurate UHDR and CONV irradiation to investigate FLASH-mitigated spinal cord toxicity in rats. This work thus establishes a robust foundation for systematic evaluation of the FLASH effect in late-responding organs and for determining clinical applicability of FLASH.

physics.med-ph↗

arXiv2025

Nullspace-preserving high-index saddle dynamics method for degenerate multiple solution problems

We propose the nullspace-preserving high-index saddle dynamics (NPHiSD) method for degenerating multiple solution systems in constrained and unconstrained settings. The NPHiSD efficiently locates high-index saddle points and provides parent states for downward searches of lower-index saddles, thereby constructing the solution landscape systematically. The NPHiSD method searches along multiple efficient ascent directions by excluding the nullspace, which is the key for upward searches in degenerate problems. To reduce the cost of frequent nullspace updates, the search is divided into segments, within which the ascent directions remain orthogonal to the nullspace of the initial state of each segment. A sufficient and necessary condition for characterizing the segment that admits efficient ascent directions is proved. Extensive numerical experiments for typical problems such as Lifshitz-Petrich, Gross-Pitaevskii, and Lennard-Jones models are performed to show the universality and effectiveness of the NPHiSD method.

math.NA↗

arXiv2025

Dielectric Deposition Enhanced Crystallization in Atomic-Layer-Deposited Indium Oxide Transistors Achieving High Gated-Hall Mobility Exceeding 100 cm2/Vs at Room Temperature

In this work, we report high-performance atomic-layer-deposited indium oxide (In2O3) transistors with high gated-Hall mobility (μH) exceeding 100 cm2/Vs at room temperature (RT). It is found that the deposition of top hafnium oxide (HfO2) above the In2O3 channel significantly enhances its crystallization, leading to an average grain size of 97.2 nm in a 4.2-nm In2O3 channel. The ALD of In2O3 exhibits an epitaxy-like growth behavior, with its (222) planes aligning parallel to the (111) planes of both the top and bottom HfO2 dielectrics. As a result, bottom-gate In2O3 transistors with a high μH of 100.9 cm2/Vs and a decent subthreshold swing (SS) of 94 mV/dec are achieved by gated-Hall measurement at RT. Furthermore, the devices maintain excellent performance at low temperatures, achieving a μH of 162.2 cm2/Vs at 100 K. Our study reveals the critical role of dielectric deposition induced crystallization in enhancing carrier transport and offers a scalable pathway toward high-mobility devices.

cond-mat.mtrl-sci↗

arXiv2025

Amplitude Expansion Phase Field Crystal (APFC) Modeling based Efficient Dislocation Simulations using Fourier Pseudospectral Method

Crystalline defects critically influence material properties, necessitating accurate simulation methods. Existing approaches, from atomic-scale configurations to continuum elasticity, face inherent limitations in modeling dislocation-induced lattice deformation. The amplitude expansion of the phase field crystal (APFC) model bridges this gap with a mesoscopic description. This paper introduces a computationally efficient Fourier pseudospectral method for solving the APFC equations. The method exploits system periodicity and solution analyticity--the latter's rigorous proof remaining an open question, as discussed herein--to enable precise implementation of periodic boundary conditions. Numerical experiments on 2D triangular and 3D body-centered cubic lattices demonstrate that the method accurately reproduces the strain fields of edge dislocations, matching continuum theory predictions. These results confirm the APFC model's potential for capturing complex defect structures at the mesoscale, paving the way for simulating more intricate defect dynamics.

cond-mat.mtrl-sci↗

arXiv2025

Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning

Class Incremental Learning (CIL) aims to continuously learn new categories while retaining the knowledge of old ones. Pre-trained models (PTMs) show promising capabilities in CIL. However, existing approaches that apply lightweight fine-tuning to backbones still induce parameter drift, thereby compromising the generalization capability of pre-trained models. Parameter drift can be conceptualized as a form of noise that obscures critical patterns learned for previous tasks. However, recent researches have shown that noise is not always harmful. For example, the large number of visual patterns learned from pre-training can be easily abused by a single task, and introducing appropriate noise can suppress some low-correlation features, thus leaving a margin for future tasks. To this end, we propose learning beneficial noise for CIL guided by information theory and propose Mixture of Noise (Min), aiming to mitigate the degradation of backbone generalization from adapting new tasks. Specifically, task-specific noise is learned from high-dimension features of new tasks. Then, a set of weights is adjusted dynamically for optimal mixture of different task noise. Finally, Min embeds the beneficial noise into the intermediate features to mask the response of inefficient patterns. Extensive experiments on six benchmark datasets demonstrate that Min achieves state-of-the-art performance in most incremental settings, with particularly outstanding results in 50-steps incremental settings. This shows the significant potential for beneficial noise in continual learning. Code is available at https://github.com/ASCIIJK/MiN-NeurIPS2025.

cs.CV↗

arXiv2025

Visual Generation Without Guidance

Classifier-Free Guidance (CFG) has been a default technique in various visual generative models, yet it requires inference from both conditional and unconditional models during sampling. We propose to build visual models that are free from guided sampling. The resulting algorithm, Guidance-Free Training (GFT), matches the performance of CFG while reducing sampling to a single model, halving the computational cost. Unlike previous distillation-based approaches that rely on pretrained CFG networks, GFT enables training directly from scratch. GFT is simple to implement. It retains the same maximum likelihood objective as CFG and differs mainly in the parameterization of conditional models. Implementing GFT requires only minimal modifications to existing codebases, as most design choices and hyperparameters are directly inherited from CFG. Our extensive experiments across five distinct visual models demonstrate the effectiveness and versatility of GFT. Across domains of diffusion, autoregressive, and masked-prediction modeling, GFT consistently achieves comparable or even lower FID scores, with similar diversity-fidelity trade-offs compared with CFG baselines, all while being guidance-free. Code will be available at https://github.com/thu-ml/GFT.

cs.CV↗