Yiming Dong

Huan Li, Yiming Dong, Zhouchen Lin

This paper studies the AdamW-style Shampoo optimizer, an effective implementation of classical Shampoo that notably won the external tuning track of the AlgoPerf neural network training algorithm competition. Our analysis unifies one-sided and two-sided preconditioning and establishes the convergence rate $\frac{1}{K}\sum_{k=1}^K E\left[\|\nabla f(X_k)\|_*\right]\leq O(\frac{\sqrt{m+n}C}{K^{1/4}})$ measured by nuclear norm, where $K$ represents the iteration number, $(m,n)$ denotes the size of matrix parameters, and $C$ matches the constant in the optimal convergence rate of SGD. Theoretically, we have $\|\nabla f(X)\|_F\leq \|\nabla f(X)\|_*\leq \sqrt{m+n}\|\nabla f(X)\|_F$, supporting that our convergence rate can be considered to be analogous to the optimal $\frac{1}{K}\sum_{k=1}^KE\left[\|\nabla f(X_k)\|_F\right]\leq O(\frac{C}{K^{1/4}})$ convergence rate of SGD in the ideal case of $\|\nabla f(X)\|_*= Θ(\sqrt{m+n})\|\nabla f(X)\|_F$.

Yiming Dong, Yacheng Kang, Junjie Zhao et al.

With the detection of gravitational waves (GWs), multi-messenger astronomy has opened a new window for advancing our understanding of astrophysics, dense matter, gravitation, and cosmology. The GW sources detected to date are from mergers of compact object binaries, which possess the potential to generate detectable electromagnetic (EM) counterparts. Searching for associations between GW signals and their EM counterparts is an essential step toward enabling subsequent multi-messenger studies. In the era of next-generation GW and EM detectors, the rapid increase in the number of events brings not only unprecedented scientific opportunities, but also substantial challenges to the existing data analysis paradigm. To help address these challenges, we develop GW-Eyes, an agentic framework powered by large language models (LLMs). For the first time, GW-Eyes integrates domain-specific tools and autonomously performs counterpart association tasks between GW and candidate EM events. It supports natural language interaction to assist human experts with auxiliary tasks such as catalog management, skymap visualization, and rapid verification. Our framework leverages the complex decision-making capabilities of LLMs and their traceable reasoning processes, offering a new perspective to the multi-messenger astronomy.

Junjie Wang, Pan Zhou, Yiming Dong et al.

Large language models (LLMs) have demonstrated impressive generalization and emergent capabilities, yet their pre-training remains computationally expensive and sensitive to optimization dynamics. While Adam-based optimizers offer fast convergence by adapting learning rates coordinate-wise, recent studies reveal that their updates often suffer from poor spectral conditioning and low-rank structures, hindering efficiency. Muon addresses this issue via global spectral normalization but lacks the per-coordinate adaptivity of Adam. In this work, we propose Column-Normalized Adam (Conda), a novel optimizer that bridges the strengths of both approaches. Conda projects updates into an orthogonal subspace and applies column-wise second moment normalization based on the projected gradients, thereby achieving both improved spectral conditioning and maintaining coordinate-wise adaptivity. This design alleviates the spectral pathologies of Adam while preserving its fast convergence behavior. Extensive experiments on the LLaMA and GPT-2 series show that Conda consistently outperforms AdamW, Muon, and other baselines in pre-training. Remarkably, on the LLaMA series, Conda achieves 2-2.5 the convergence speed of AdamW, measured in both training steps and training time. Further ablations demonstrate its robustness under diverse training setups. These results collectively highlight Conda as an effective and broadly applicable optimizer for large-scale LLM training. The code is released on https://github.com/jie040109/Conda

Haoyu Li, Xuhong Li, Yiming Dong et al.

Dataset diversity plays a pivotal role for the successful training of many machine learning models, particularly in the supervised fine-tuning (SFT) stage of large language model (LLM) development. Despite increasing recognition of its importance, systematic analyses of dataset diversity still remain underexplored. To address this gap, this work presents a systematic taxonomy of existing diversity-control strategies, which primarily focus on the instruction component, operating at either macroscopic (entire instruction semantics) or mesoscopic levels (instruction units), and furthermore introduces a novel analysis of microscopic diversity within the response component, specifically analyzing the statistical distribution of tokens in SFT training samples. In the experimental evaluation, we construct fixed-size datasets (e.g., 10,000 samples each) from a corpus of 117,000 open-source SFT samples, incorporating six distinct diversity-control strategies spanning macro-, meso-, and microscopic levels applied to both instructions and responses. We then fine-tune LLMs on these datasets to assess the six diversity-control strategies. Results reveal that while macroscopic and mesoscopic strategies lead to higher performance with increasing diversity, the microscopic strategy in responses exhibits both a stronger correlation between model performance and the degree of diversity and superior performance with maximum diversity across all strategies. These findings offer actionable insights for constructing high-performance SFT datasets.

Yiming Dong, Huan Li, Zhouchen Lin

The LION (evoLved sIgn mOmeNtum) optimizer for deep neural network training was found by Google via program search, with the simple sign update yet showing impressive performance in training large scale networks. Although previous studies have investigated its convergence properties, a comprehensive analysis, especially the convergence rate, is still desirable. Recognizing that LION can be regarded as solving a specific constrained problem, this paper focuses on demonstrating its convergence to the Karush-Kuhn-Tucker (KKT) point at the rate of $\cal O(\sqrt{d}K^{-1/4})$ measured by gradient $\ell_1$ norm, where $d$ is the problem dimension and $K$ is the number of iteration steps. Step further, we remove the constraint and establish that LION converges to the critical point of the general unconstrained problem at the same rate. This rate not only delivers the currently optimal dependence on the problem dimension $d$ but also tightly matches the theoretical lower bound for nonconvex stochastic optimization algorithms, which is typically measured using the gradient $\ell_2$ norm, with respect to the number of iterations $K$. Through extensive experiments, we not only demonstrate that LION achieves lower loss and higher performance compared to standard SGD, but also empirically confirm that the gradient $\ell_1/\ell_2$ norm ratio aligns with $Θ(\sqrt{d})$, thus proving that our convergence rate matches the theoretical lower bound with respect to $d$ in the empirical sense.

Yiming Dong, Kun Fu, Haoyu Li et al.

Prolonged reinforcement learning with verifiable rewards (RLVR) has been shown to drive continuous improvements in the reasoning capabilities of large language models, but the training is often prone to instabilities, especially in Mixture-of-Experts (MoE) architectures. Training instability severely undermines model capability improvement, yet its underlying causes and mechanisms remain poorly understood. In this work, we introduce a principled framework for understanding RLVR instability through the lens of objective-level hacking. Unlike reward hacking, which arises from exploitable verifiers, objective-level hacking emerges from token-level credit misalignment and is manifested as system-level spurious signals in the optimization objective. Grounded in our framework, together with extensive experiments on a 30B MoE model, we trace the origin and formalize the mechanism behind a key pathological training dynamic in MoE models: the abnormal growth of the training-inference discrepancy, a phenomenon widely associated with instability but previously lacking a mechanistic explanation. These findings provide a concrete and causal account of the training dynamics underlying instabilities in MoE models, offering guidance for the design of stable RLVR algorithms.

Zhoutong Wu, Yuan Zhang, Yiming Dong et al.

Transformer models have driven breakthroughs across various language tasks by their strong capability to learn rich contextual representations. Scaling them to improve representation, however, often demands substantial memory and compute costs, such as the Key-Value (KV) cache used during auto-regressive decoding. Skip connections offer a promising way to improve representation without bloating resource usage, yet most prior works either improve expressivity while leaving KV costs unchanged, or reduce memory at the cost of weaker representation. In this work, we propose SkipV1Former, a Transformer variant that uses skip connections from the first layer's Value heads to strengthen model representation and reduce KV cache. Specifically, from the second block onward, each layer reuses half of its Value heads from the very first layer, while computing the other half as usual-cutting Value projections and V cache by nearly 50 \%. Theoretically, we show that routing uncompressed first-layer Values into deeper layers restores information lost to compression and accelerates the model's implicit mesa-optimization-a key pattern of Transformer in auto-regressive tasks. Empirically, across different model scales, SkipV1Former delivers consistent reductions of approximately 25 \% in KV cache while improving perplexity relative to standard Multi-Head Attention (MHA) Transformers and some advanced variants. Moreover, we propose a recipe for uptraining existing MHA Transformer checkpoints to SkipV1Former with only 10-15\% additional compute. Finally, SkipV1Former can seamlessly combine advanced methods like Group-Query Attention and Multi-Latent Attention to achieve further KV cache savings and performance improvement. When combined with YOCO, it cuts KV cache size by nearly 50 \% while still improving performance.

Jiawei Shan, Yiming Dong, Jiwei Zhao

Real-world applications often face scarce labeled data due to the high cost and time requirements of gold-standard experiments, whereas unlabeled data are typically abundant. With the growing adoption of machine learning techniques, it has become increasingly feasible to generate multiple predicted labels using a variety of models and algorithms, including deep learning, large language models, and generative AI. In this paper, we propose a novel approach that safely and adaptively aggregates multiple black-box predictions with unknown quality while preserving valid statistical inference. Our method provides two key guarantees: (i) it never performs worse than using the labeled data alone, regardless of the quality of the predictions; and (ii) if any one of the predictions (without knowing which one) perfectly fits the ground truth, the algorithm adaptively exploits this to achieve either a faster convergence rate or the semiparametric efficiency bound. We demonstrate the effectiveness of the proposed algorithm through experiments on both synthetic and benchmark datasets.

Yibo Jin, Yixu Xu, Yue Chen et al.

Serving disaggregated large language models has been widely adopted in industrial practice for enhanced performance. However, too many tokens generated in decoding phase, i.e., occupying the resources for a long time, essentially hamper the cloud from achieving a higher throughput. Meanwhile, due to limited on-device resources, the time to first token (TTFT), i.e., the latency of prefill phase, increases dramatically with the growth on prompt length. In order to concur with such a bottleneck on resources, i.e., long occupation in cloud and limited on-device computing capacity, we propose to separate large language model between cloud and devices. That is, the cloud helps a portion of the content for each device, only in its prefill phase. Specifically, after receiving the first token from the cloud, decoupling with its own prefill, the device responds to the user immediately for a lower TTFT. Then, the following tokens from cloud are presented via a speed controller for smoothed TPOT (the time per output token), until the device catches up with the progress. On-device prefill is then amortized using received tokens while the resource usage in cloud is controlled. Moreover, during cloud prefill, the prompt can be refined, using those intermediate data already generated, to further speed up on-device inference. We implement such a scheme P/D-Device, and confirm its superiority over other alternatives. We further propose an algorithm to decide the best settings. Real-trace experiments show that TTFT decreases at least 60%, maximum TPOT is about tens of milliseconds, and cloud throughput increases by up to 15x.

Anda Tang, Yiming Dong, Yutao Zeng et al. · bytedance

The expanding computational costs and limited resources underscore the critical need for budgeted-iteration training, which aims to achieve optimal learning within predetermined iteration budgets. While learning rate schedules fundamentally govern the performance of different networks and tasks, particularly in budgeted-iteration scenarios, their design remains largely heuristic, lacking theoretical foundations. In addition, the optimal learning rate schedule requires extensive trial-and-error selection, making the training process inefficient. In this work, we propose the Unified Budget-Aware (UBA) schedule, a theoretically grounded learning rate schedule that consistently outperforms commonly-used schedules among diverse architectures and tasks under different constrained training budgets. First, we bridge the gap by constructing a novel training budget-aware optimization framework, which explicitly accounts for the robustness to landscape curvature variations. From this framework, we derive the UBA schedule, controlled by a single hyper-parameter \varphi that provides a trade-off between flexibility and simplicity, eliminating the need for per-network numerical optimization. Moreover, we establish a theoretical connection between \varphi and the condition number, adding interpretation and justification to our approach. Besides, we prove the convergence for different values of \varphi. We offer practical guidelines for its selection via theoretical analysis and empirical results. Extensive experimental results show that UBA consistently surpasses the commonly-used schedules across diverse vision and language tasks, spanning network architectures (e.g., ResNet, OLMo) and scales, under different training-iteration budgets.

Huan Li, Yiming Dong, Zhouchen Lin

As the default optimizer for training large language models, AdamW has achieved remarkable success in deep learning. However, its convergence behavior is not theoretically well-understood. This paper establishes the convergence rate $\frac{1}{K}\sum_{k=1}^KE\left[||\nabla f(x^k)||_1\right]\leq O(\frac{\sqrt{d}C}{K^{1/4}})$ for AdamW measured by $\ell_1$ norm, where $K$ represents the iteration number, $d$ denotes the model dimension, and $C$ matches the constant in the optimal convergence rate of SGD. Theoretically, we have $||\nabla f(x)||_2\ll ||\nabla f(x)||_1\leq \sqrt{d}||\nabla f(x)||_2$ for any high-dimensional vector $x$ and $E\left[||\nabla f(x)||_1\right]\geq\sqrt{\frac{2d}π}E\left[||\nabla f(x)||_2\right]$ when each element of $\nabla f(x)$ is generated from Gaussian distribution $\mathcal N(0,1)$. Empirically, our experimental results on real-world deep learning tasks reveal $||\nabla f(x)||_1=\varTheta(\sqrt{d})||\nabla f(x)||_2$. Both support that our convergence rate can be considered to be analogous to the optimal $\frac{1}{K}\sum_{k=1}^KE\left[||\nabla f(x^k)||_2\right]\leq O(\frac{C}{K^{1/4}})$ convergence rate of SGD in the ideal case. We also extend our result to NAdamW, an AdamW variant that employs a double-momentum mechanism, and demonstrate that it maintains the same convergence rate.

Huan Li, Yiming Dong, Zhouchen Lin

Although adaptive gradient methods have been extensively used in deep learning, their convergence rates proved in the literature are all slower than that of SGD, particularly with respect to their dependence on the dimension. This paper considers the classical RMSProp and its momentum extension and establishes the convergence rate of $\frac{1}{T}\sum_{k=1}^T E\left[\|\nabla f(x^k)\|_1\right]\leq O(\frac{\sqrt{d}C}{T^{1/4}})$ measured by $\ell_1$ norm without the bounded gradient assumption, where $d$ is the dimension of the optimization variable, $T$ is the iteration number, and $C$ is a constant identical to that appeared in the optimal convergence rate of SGD. Our convergence rate matches the lower bound with respect to all the coefficients except the dimension $d$. Since $\|x\|_2\ll\|x\|_1\leq\sqrt{d}\|x\|_2$ for problems with extremely large $d$, our convergence rate can be considered to be analogous to the $\frac{1}{T}\sum_{k=1}^T E\left[\|\nabla f(x^k)\|_2\right]\leq O(\frac{C}{T^{1/4}})$ rate of SGD in the ideal case of $\|\nabla f(x)\|_1=\varTheta(\sqrt{d}\|\nabla f(x)\|_2)$.