Villalobos 等人 (2024) 预测，公开可用的文本数据将在未来十年内枯竭。因此，在无法获取真实标注（ground-truth labels）的情况下提升模型性能变得愈发重要。我们提出了一种无需标注的后处理框架，利用一个虽然较弱但校准效果更好的参考模型，来改进一个性能强大但校准不良的模型。我们的框架能够在不依赖标签的情况下，确保最坏情况下的性能得到严格提升。我们的方法基于对“何时能实现严格提升”这一条件的刻画，即当强模型与参考模型未实现相互校准时。我们对这一条件进行了形式化处理，并将其与经济学中的“无套利”（no-arbitrage）结果建立了联系。我们证明了该问题可以采用 Bregman 投影作为解决方案，前人研究（Mohri 等人，2025）曾将其应用于语言模型的自我提升。我们将 Bregman 投影实现为一种针对强模型输出的高效后处理算法。在不同规模的代表性大语言模型（LLMs）上进行的实验表明，我们的无标注方法显著降低了适当损失（proper losses）和校准误差，实现了与有监督基线相当的性能。

Pre-training LLMs is extremely resource-intensive, making optimizer efficiency critical. Matrix-based optimizers such as Muon and SOAP improve over AdamW by leveraging curvature, but their updates can be overly isotropic—conservative in flat directions and aggressive in sharp ones. We develop a unified Riemannian ODE view showing that preconditioning defines the geometry while momentum acts as Riemannian damping. Building on this insight, we propose LITE, an acceleration strategy that increases effective damping and learning rates along flat trajectories to speed progress in anisotropic landscapes. Experiments across Dense/MoE models (130M–1.3B), datasets (C4, Pile), and schedules validate consistent speedups for Muon and SOAP, supported by theory showing faster convergence along flat directions.

基于人类反馈的强化学习（Reinforcement Learning from Human Feedback, RLHF）及其变体方法已成为将大型语言模型与人类意图对齐的主导性技术范式。尽管经验层面效果显著，但这些方法在高维场景下的理论泛化特性仍有待深入探索。为此，我们通过算法稳定性理论框架，在线性奖励模型假设下构建了面向大语言模型RLHF的泛化理论体系。与现有基于奖励模型最大似然估计一致性分析的研究不同，本研究的分析立足于端到端学习框架，这更符合实际应用场景。具体而言，我们证明在关键的特征覆盖条件下，策略模型的经验最优解具有阶泛化误差界。进一步地，该结论可推广至基于梯度的学习算法（即梯度上升算法与随机梯度上升算法）所获得的参数解。因此，我们主张本文结果为RLHF处理后大语言模型经验性泛化能力提供了新的理论佐证。

Bandit models are a key framework for designing algorithms in interactive decision-making. While stochastic linear bandits are well-studied, real-world complexities have led to important extensions like generalized linear bandits (GLB) with nonlinear link functions, and heavy-tailed linear bandits (HvLB) to handle heavy-tailed noise. While optimal regret bounds have been established, existing algorithms are computationally impractical, requiring full data storage and repeated passes over all historical data. In this talk, I will introduce a "one-pass" method based on the Online Mirror Descent framework, a textbook-standard approach for regret optimization whereas we here use it as a statistical estimator. This approach achieves O(1) per-round computational cost while preserving optimal regret for GLB and HvLB. Then I will discuss extensions to online RL theory: (i) RL with multinomial logit function approximation, and (ii) RLHF with on-policy active data collection.

In this talk, we focus on decentralized online convex optimization (OCO) in changing environments, and aim to minimize adaptive regret and dynamic regret. It is well-known that in the standard OCO, plenty of algorithms with (nearly) optimal bounds on these two metrics have been proposed. However, none of them has been extended into the decentralized OCO, possibly due to the difficulty in handling their commonly used two-level structure. To fill the gap, we first provide novel reductions from minimizing these two metrics of the decentralized OCO to minimizing them of OCO with delayed feedback. Furthermore, we revisit an existing black-box reduction from the delayed OCO to the standard OCO, and prove that it can also convert non-delayed algorithms for adaptive regret and dynamic regret into the delayed setting. Finally, we demonstrate the power of our reductions by establishing nearly optimal bounds on adaptive regret and dynamic regret of the decentralized OCO.

模仿学习从专家示例中学习策略模型，是大语言模型与具身智能模型的关键底层技术。然而，在长决策时域任务下，模仿学习往往面临误差累积与分布偏移等困难。值得关注的是，对抗模仿学习这类方法展现了优异的实践性能：仅依赖极少量甚至单条专家轨迹，也能在机器人运动控制等长序列决策任务中取得接近专家的性能。这一现象引出了两个长期未解的核心问题：对抗模仿学习为何能够在小样本条件下有效，以及其性能为何不随决策时域增长而显著退化。本报告围绕上述问题，介绍我们对一种基于总变差距离的对抗模仿学习方法（TV-AIL）的理论分析结果。针对一类从机器人运动控制任务抽象得到的MDP，我们证明了TV-AIL 具有$\mathcal{O} (\min \{1, \sqrt{|S|/N}\})$的模仿误差界，该误差界与决策时域无关，且在小样本与大样本条件下均具有理论意义。该结果为对抗模仿学习的优异实践性能提供了理论解释，并揭示了其缓解分布偏移问题的内在机制。分析过程中，我们利用TV-AIL 特有的多阶段策略优化结构，提出了一种新的阶段耦合分析方法；同时，该分析方法也帮助刻画了TV-AIL 在一般MDP 中的最坏情况表现，从而明确其适用范围与潜在局限。

Recent developments of stochastic optimization often suggest biased gradient estimators to improve either the robustness, communication efficiency or computational speed. Representative biased stochastic gradient methods (BSGMs) include Zeroth-order stochastic gradient descent (SGD), Clipped-SGD and SGD with delayed gradients. In this talk, we present the first framework to study the stability and generalization of BSGMs for convex and smooth problems. We apply our general result to develop the first stability bound for Zeroth-order SGD with reasonable step size sequences, and the first stability bound for Clipped-SGD. While our stability analysis is developed for general BSGMs, the resulting stability bounds for both Zeroth-order SGD and Clipped-SGD match those of SGD under appropriate smoothing/clipping parameters.

We study multi-bit watermarking for data generated by stochastic processes, where a hidden message is embedded during sampling and must be decodable by an authorized detector that possesses side information unavailable to unauthorized observers. In high-stakes deployments, a practical watermark must simultaneously control false alarms, preserve generation quality without distorting the output distribution, and support reliable multi-bit decoding. Satisfying all three goals at once inevitably creates fundamental trade-offs. We formulate watermark embedding as a distributional information-embedding problem and watermark detection as a multiple-hypothesis testing problem under distortion and rate constraints, leading to four fundamental metrics: false-alarm probability, detection error probability, distortion, and information rate. Within this information-theoretic framework, we derive matched converse and achievability bounds that characterize the optimal trade-offs and provide scheme-ag.

Condensation (also known as quantization, clustering, or alignment) is a widely observed phenomenon where neurons in the same layer tend to align with one another during the nonlinear training of deep neural networks (DNNs). It is a key characteristic of the feature learning process of neural networks. In recent years, to advance the mathematical understanding of condensation, we uncover structures regarding the dynamical regime, loss landscape and generalization for deep neural networks, based on which a novel theoretical framework emerges. This presentation will cover these findings in detail. First, I will present results regarding the dynamical regime identification of condensation at the infinite width limit, where small initialization is crucial. Then, I will discuss the mechanism of condensation at the initial training stage and the global loss landscape structure underlying condensation in later training stages, highlighting the prevalence of condensed critical points and globa

While transformers are typically assumed to be "blank slates" at random initialization, we demonstrate that untrained models exhibit systematic structural biases, including extreme token preferences. We provide a mechanistic explanation by identifying a contraction of token representations along initialization-dependent directions, which is driven by the interaction of asymmetric MLP activations and self-attention value aggregation. We show these biases persist throughout training, forming a stable model identity that enables SeedPrint, a method to fingerprint LLMs by their random initialization. Finally, we establish that attention mechanism’s intra-sequence contraction is causally linked to the attention-sink phenomenon, providing a principled explanation for sink emergence and a pathway for systematic control.

理解深度学习在实际问题中的性能需要考虑模型特征、数据特征以及连接这两部分的优化算法的特征。该报告将从多种角度来分析数据特征，并设计实验来挖掘模型和优化的特征，以理解深度学习的泛化能力和语言模型的推理能力，并对实际的模型训练提供一些参考。我们发现小初始化会使模型更偏好推理的方式来解释数据，而非记忆的方式，这与模型在小初始化有凝聚的现象紧密相关。另外，数据中的一些重要统计量是形成嵌入结构的驱动力，并影响模型的推理能力。

While Masked Diffusion Language Models (MDLMs) relying on token masking and unmasking have shown promise in language modeling, the training efficiency, generation quality and flexibility remain constrained by the masking paradigm. We propose Deletion-Insertion Diffusion language models (DID) that rigorously formulate token deletion and insertion as discrete diffusion processes, replacing the masking and unmasking processes in current MDLMs, which improves training and inference efficiency by eliminating computational overhead in MDLMs. Moreover, we propse Variational Autoencoding Discrete Diffusion (VADD), a novel framework that enhances discrete diffusion with latent variable modeling to implicitly capture correlations among dimensions, to improve the parallel decording.

Low-precision training is essential for scaling transformer pretraining, but it can introduce delayed instability that is often invisible in the early phase of optimization. In this talk, I present an invariance-based view of this phenomenon. The key observation is that many transformer components contain shift-invariant directions: for example, adding a constant offset to all logits does not change the softmax loss, and LayerNorm is likewise insensitive to constant shifts. In exact arithmetic, gradients cancel along these directions. Under low-precision computation, however, this cancellation is imperfect, and small arithmetic errors can accumulate into harmful parameter drift. I will show how this perspective leads to a simple theoretical picture in which the unstable component behaves like a variance-accumulating stochastic process, and how that analysis suggests a lightweight energy-based signal for early detection. Experiments on GPT-style pretraining under BF16 and FP8 illustrate that this signal can distinguish stable and unstable runs before the training loss clearly separates. I will conclude by discussing the practical implications of this viewpoint, including how symmetry-aware corrections may offer low-cost routes to more stable low-precision pretraining.

Stein's method is a classical "descriptive" framework in probability theory for proving quantitative central limit theorems (CLT). This talk introduces a somewhat surprising algorithmic application of this framework in adversarial online linear optimization (OLO). Focusing on one-dimensional fixed-time OLO on a bounded domain, we present an algorithm based on Stein's method which is capable of achieving various "additively sharp" performance guarantees, surpassing the conventional big-O optimality. In particular, instantiations of this algorithm improve upon the total loss upper bounds of classical baselines including OGD and MWU. Conceptually, our algorithm can be viewed as a "continuous" refinement of a seminal dynamic programming algorithm of T. Cover (1966), improving its computational complexity. Technically, our construction is inspired by the remarkably clean proof of a Wasserstein CLT due to A. Röllin (2018). This is a joint work with Aaditya Ramdas at CMU.

在现代机器学习应用中，数据往往以流的形式动态到达，要求模型能够基于当前数据样本实时计算梯度并进行持续更新。本报告旨在探讨几类典型的流式学习场景，并系统研究在这些场景下优化训练效率的加速方法。具体内容涵盖以下几个方面：首先，针对非结构化学习问题中的一般性非凸目标函数，我们探讨了方差缩减技术，论证其如何保证算法在寻求一阶与二阶稳定点时达到最快的收敛速度；其次，在结构化问题中，对于广义线性回归所对应的凸优化问题，我们分析了动量方法在宽泛条件下的加速效果；最后，针对张量分解等非凸优化难题，我们研究了模型过参数化表达与适当的梯度归一化方法，论证其是缩小统计与计算之间鸿沟的关键有效策略。

面向真实开放环境，图增强大模型正由静态、封闭、端到端训练范式，转向面向噪声数据、序列任务与交互反馈的自演化智能新范式，对鲁棒表征、持续适配与闭环决策提出更高要求。报告围绕数据自演化、任务自演化和反馈自演化三条主线展开，系统总结了动态鲁棒表征、可持续学习调优和智能推理决策等研究布局，重点介绍了面向稀疏含噪信号的跨模态对齐、面向开放多样场景的持续学习与个性化适配，以及利用稀疏反馈实现关联发现、推理规划与行动生成闭环的阶段性进展。

在大语言模型时代，如何让AI系统真正理解和推理复杂的关系型数据，是一个亟待解决的关键问题。尽管大语言模型在人工智能领域带来了深刻变革，但其在处理结构化信息和理解复杂知识依赖关系方面仍存在根本性的局限。本研究围绕图基础模型展开，致力于应对这一核心挑战。本次报告将首先介绍我们在图基础模型方面的开创性工作。通过提出新颖的大规模模型构建技术，我们的模型能够有效捕获和推理复杂的图结构信息，实现了以往难以达到的鲁棒且可泛化的图学习能力。我们所提出的框架在处理大规模图结构的同时，能够有效保留关键的拓扑信息，从而突破了传统方法的瓶颈。随后，报告将展示我们的图学习框架如何通过检索增强生成和自主智能体等技术，将图结构化知识融入大语言模型，从而显著提升其推理能力。图基础模型与大语言模型的深度协同，使得对结构化数据的复杂推理成为可能，相关应用涵盖推荐系统、生物信息学、城市计算以及AI智能体等多个领域。本研究为基础模型的发展开辟了新方向，是迈向更强大、更具泛化能力的人工智能系统的重要一步。

Molecular identification and discovery are vital in bioanalysis, environmental governance, and customs inspection. While spectroscopic tools (e.g., MS, IR, NMR, X-ray) enable molecular probing, current analysis relies heavily on database matching and expert interpretation—limiting novelty detection, efficiency, accuracy, and scalability. To overcome this, we developed SpectraAI: a holistic framework integrating (1) high-quality spectral data curation, (2) spectroscopy-specific large foundation models, (3) AI Agents for automated analysis, (4) biochemically informed algorithms, and (5) real-world deployment in biomedicine and environmental monitoring. In this talk, I will present SpectraAI’s architecture, recent advances, and key future directions—including de novo molecule discovery and multimodal spectral reasoning.

Transformer models rely on attention mechanism to capture long-range dependencies but suffer from quadratic complexity, limiting their scalability to long sequences. Kernel-based linear attention reduces this complexity but typically relies on fixed or weakly learnable kernels, restricting expressiveness and performance. In this work, we propose Flexformer, a flexible linear Transformer that learns attention kernels in a fully data-driven manner. Flexformer builds on random Fourier feature-based linear attention and treats spectral frequencies as trainable parameters, enabling the model to learn a broad family of attention kernels. We develop both stationary and nonstationary variants, with the latter offering strictly greater expressiveness. Extensive experiments on language modeling and sequence classification demonstrate that Flexformer consistently outperforms baselines. Moreover, Flexformer can be effectively distilled from pretrained Transformers to recover softmax attention and exhibits strong kernel transferability across domains, achieving both high efficiency and competitive performance on long-sequence tasks.

Semantic segmentation labels each pixel in an image with its corresponding class, and is typically evaluated using the Intersection over Union (IoU) and Dice metrics to quantify the overlap between predicted and ground-truth segmentation masks. In the literature, most existing methods estimate pixel-wise class probabilities, then apply argmax or thresholding to obtain the final prediction. These methods have been shown to generally lead to inconsistent or suboptimal results, as they do not directly maximize segmentation metrics. To address this issue, a novel consistent segmentation framework, RankSEG, has been proposed, which includes RankDice and RankIoU specifically designed to optimize the Dice and IoU metrics, respectively. Although RankSEG almost guarantees improved performance, it suffers from two major drawbacks. First, it is its computational expense-RankDice has a complexity of O(d log d) with a substantial constant factor (where d represents the number of pixels), while RankIoU exhibits even higher complexity O(d^2), thus limiting its practical application. For instance, in LiTS, prediction with RankSEG takes 16.33 seconds compared to just 0.01 seconds with the argmax rule. Second, RankSEG is only applicable to overlapping segmentation settings, where multiple classes can occupy the same pixel, which contrasts with standard benchmarks that typically assume non-overlapping segmentation. In this paper, we overcome these two drawbacks via a reciprocal moment approximation (RMA) of RankSEG with the following contributions: (i) we improve RankSEG using RMA, namely RankSEG-RMA, reduces the complexity of both algorithms to O(d) while maintaining comparable performance; (ii) inspired by RMA, we develop a pixel-wise score function that allows efficient implementation for non-overlapping segmentation settings.

Optimal transport (OT) has emerged as a fundamental tool in modern machine learning, yet its computational cost remains a significant bottleneck for large-scale applications. While harnessing the massive parallelism of modern GPU hardware is critical for efficiency, the de facto standard Sinkhorn algorithm, despite its ease of parallelization, often suffers from slow convergence in challenging problems. More recently, the sparse-plus-low-rank quasi-Newton method offers a balance between convergence rate and per-iteration complexity; however, its efficiency on GPUs is severely hindered by the serial nature of sparse matrix symbolic analysis and irregular memory access patterns. To bridge this gap, we present cuRegOT, a high-performance GPU solver tailored for entropic-regularized OT. We introduce a suite of algorithmic and architectural optimizations, including an amortized symbolic analysis strategy to mitigate CPU bottlenecks, an asynchronous Sinkhorn iterates generation mechanism, and a fused kernel for bandwidth-efficient gradient evaluation. These strategies are backed by rigorous theoretical guarantees ensuring algorithmic convergence. Extensive numerical experiments demonstrate that cuRegOT achieves significant speedups over state-of-the-art GPU-based solvers across a variety of benchmark tasks.

TBD

LLM training now spans multiple stages (pretraining → mid-training → SFT → post-training RL), and dataset choice increasingly determines benchmark outcomes and safety. We focus on two data-selection problems. First, in mid-training, we study how models acquire knowledge from mixed sources and show phase-transition behavior: the optimal response flips between sources rather than interpolating, and the critical mixture ratio depends on model size, implying scale-dependent mixture recipes. We then present a practical framework that learns per-domain losses as a function of (model size, mixture) and optimizes mixtures for benchmark scores. Second, for post-training RLVR, we argue for training on hard problems while addressing sparse rewards by augmenting out-of-distribution tasks with hints/partial solutions to speed learning.

在当前大模型对齐研究中，一个备受关注的现象是弱到强泛化（Weak-to-Strong Generalization, W2SG），即通过弱教师模型生成伪标签，指导强学生模型进行训练，从而实现学生在目标任务中反超教师的现象。尽管这一现象已被实证观察到，但其理论机理仍未被充分揭示。本报告围绕W2SG的理论分析展开，重点在于用Bregman散度下的广义偏差-方差分解刻画学生与教师之间的风险差异，首次在不依赖假设空间凸性这一强假设的前提下，推导出基于“预测不匹配”的W2SG不等式。我们进一步理论证明对于容量足够大的学生模型，W2SG现象更有可能出现。与此同时，在W2SG损失函数选择方面，我们理论上比较了标准交叉熵与反向交叉熵在W2SG场景下的表现，指出后者在面对教师预测不确定性时更加稳健。此外，我们通过实证分析验证了上述理论发现，包括学生模型容量对W2SG的影响，以及来自多个教师模型的平均监督对提升学生性能的作用。最后我们也会探讨预训练过程对弱到强泛化的促进作用。