Fengzhuo Zhang

Shuche Wang, Fengzhuo Zhang, Jiaxiang Li et al.

Muon improves training efficiency over Adam in large language-model training by about two times, but the local geometric source of this advantage remains unclear. Our work takes a first step toward demystifying Muon's superiority over Adam from a curvature perspective. First, we apply a second-order Taylor approximation to the training landscape and show that Muon achieves a larger one-step loss decrease than Adam at matched validation loss. The two optimizers have comparable first-order gains, but Muon consistently incurs a smaller second-order curvature penalty. Second, we decompose this curvature penalty into the squared update norm and Normalized Directional Sharpness (NDS). We find that Muon and Adam have comparable update norms, so Muon's smaller curvature penalty is driven by lower NDS, not update scale. Third, we study how training data and model structure shape Muon's NDS advantage. Using Zipf-Probabilistic Context-Free Grammar (PCFG) data with controlled imbalance, we show that data imbalance amplifies Muon's NDS advantage over Adam. A within-/cross-layer decomposition further shows that, in the middle and late stages of training, Muon's lower NDS is mainly sustained by smaller within-layer curvature. Beyond empirical evidence, we analyze stylized quadratic problems with heterogeneous curvature and gradient alignment toward high-curvature modes. We prove that Muon attains a smaller average NDS than GD by balancing update energy across curvature groups; when curvature heterogeneity is sufficiently strong, this also yields lower local quadratic loss after the same number of steps.

Jianliang He, Leda Wang, Fengzhuo Zhang et al.

Understanding how structured internal structure emerges during neural network training is central to the study of deep learning. We investigate this phenomenon through the group composition task, where a two-layer neural network is trained to predict $g_1 \star g_2$ for elements of a finite group $G$. By lifting the projected gradient flow to the Fourier domain, we demonstrate that the training dynamics are governed by a Riemannian gradient ascent on a representation-theoretic energy functional. We prove that, under random initialization, this flow drives each neuron to converge almost surely toward a single irreducible representation, while the cross-layer Fourier coefficients achieve a rotational rank-one alignment. This framework provides a representation-theoretic account of feature learning and characterizes a novel low-rank compression phenomenon for matrix-valued group representations. Moreover, for Abelian groups, we provide a complete population-level description: random initialization promotes uniform diversification across nontrivial representations and induces Haar-uniform phases, jointly approximating the indicator via a majority-vote mechanism. We further prove that both phase alignment and representation competition emerge with exponential convergence rates.

Fengzhuo Zhang, Boyi Liu, Kaixin Wang et al.

The cooperative Multi-A gent R einforcement Learning (MARL) with permutation invariant agents framework has achieved tremendous empirical successes in real-world applications. Unfortunately, the theoretical understanding of this MARL problem is lacking due to the curse of many agents and the limited exploration of the relational reasoning in existing works. In this paper, we verify that the transformer implements complex relational reasoning, and we propose and analyze model-free and model-based offline MARL algorithms with the transformer approximators. We prove that the suboptimality gaps of the model-free and model-based algorithms are independent of and logarithmic in the number of agents respectively, which mitigates the curse of many agents. These results are consequences of a novel generalization error bound of the transformer and a novel analysis of the Maximum Likelihood Estimate (MLE) of the system dynamics with the transformer. Our model-based algorithm is the first provably efficient MARL algorithm that explicitly exploits the permutation invariance of the agents. Our improved generalization bound may be of independent interest and is applicable to other regression problems related to the transformer beyond MARL.

Fengzhuo Zhang, Vincent Y. F. Tan, Zhaoran Wang et al.

We design and analyze reinforcement learning algorithms for Graphon Mean-Field Games (GMFGs). In contrast to previous works that require the precise values of the graphons, we aim to learn the Nash Equilibrium (NE) of the regularized GMFGs when the graphons are unknown. Our contributions are threefold. First, we propose the Proximal Policy Optimization for GMFG (GMFG-PPO) algorithm and show that it converges at a rate of $O(T^{-1/3})$ after $T$ iterations with an estimation oracle, improving on a previous work by Xie et al. (ICML, 2021). Second, using kernel embedding of distributions, we design efficient algorithms to estimate the transition kernels, reward functions, and graphons from sampled agents. Convergence rates are then derived when the positions of the agents are either known or unknown. Results for the combination of the optimization algorithm GMFG-PPO and the estimation algorithm are then provided. These algorithms are the first specifically designed for learning graphons from sampled agents. Finally, the efficacy of the proposed algorithms are corroborated through simulations. These simulations demonstrate that learning the unknown graphons reduces the exploitability effectively.

Yuan Cheng, Fengzhuo Zhang, Yunlong Hou et al.

Large Language Models (LLMs) often exhibit slash attention patterns, where attention scores concentrate along the $Δ$-th sub-diagonal for some offset $Δ$. These patterns play a key role in passing information across tokens. But why do they emerge? In this paper, we demystify the emergence of these Slash-Dominant Heads (SDHs) from both empirical and theoretical perspectives. First, by analyzing open-source LLMs, we find that SDHs are intrinsic to models and generalize to out-of-distribution prompts. To explain the intrinsic emergence, we analyze the queries, keys, and Rotary Position Embedding (RoPE), which jointly determine attention scores. Our empirical analysis reveals two characteristic conditions of SDHs: (1) Queries and keys are almost rank-one, and (2) RoPE is dominated by medium- and high-frequency components. Under these conditions, queries and keys are nearly identical across tokens, and interactions between medium- and high-frequency components of RoPE give rise to SDHs. Beyond empirical evidence, we theoretically show that these conditions are sufficient to ensure the emergence of SDHs by formalizing them as our modeling assumptions. Particularly, we analyze the training dynamics of a shallow Transformer equipped with RoPE under these conditions, and prove that models trained via gradient descent exhibit SDHs. The SDHs generalize to out-of-distribution prompts.

Xiangming Gu, Tianyu Pang, Chao Du et al.

Language Models (LMs) assign significant attention to the first token, even if it is not semantically important, which is known as attention sink. This phenomenon has been widely adopted in applications such as streaming/long context generation, KV cache optimization, inference acceleration, model quantization, and others. Despite its widespread use, a deep understanding of attention sink in LMs is still lacking. In this work, we first demonstrate that attention sinks exist universally in LMs with various inputs, even in small models. Furthermore, attention sink is observed to emerge during the LM pre-training, motivating us to investigate how optimization, data distribution, loss function, and model architecture in LM pre-training influence its emergence. We highlight that attention sink emerges after effective optimization on sufficient training data. The sink position is highly correlated with the loss function and data distribution. Most importantly, we find that attention sink acts more like key biases, storing extra attention scores, which could be non-informative and not contribute to the value computation. We also observe that this phenomenon (at least partially) stems from tokens' inner dependence on attention scores as a result of softmax normalization. After relaxing such dependence by replacing softmax attention with other attention operations, such as sigmoid attention without normalization, attention sinks do not emerge in LMs up to 1B parameters. The code is available at https://github.com/sail-sg/Attention-Sink.

Xinyang Hu, Fengzhuo Zhang, Siyu Chen et al.

Chain-of-Thought (CoT) prompting and its variants have gained popularity as effective methods for solving multi-step reasoning problems using pretrained large language models (LLMs). In this work, we analyze CoT prompting from a statistical estimation perspective, providing a comprehensive characterization of its sample complexity. To this end, we introduce a multi-step latent variable model that encapsulates the reasoning process, where the latent variable encodes the task information. Under this framework, we demonstrate that when the pretraining dataset is sufficiently large, the estimator formed by CoT prompting is equivalent to a Bayesian estimator. This estimator effectively solves the multi-step reasoning problem by aggregating a posterior distribution inferred from the demonstration examples in the prompt. Moreover, we prove that the statistical error of the CoT estimator can be decomposed into two main components: (i) a prompting error, which arises from inferring the true task using CoT prompts, and (ii) the statistical error of the pretrained LLM. We establish that, under appropriate assumptions, the prompting error decays exponentially to zero as the number of demonstrations increases. Additionally, we explicitly characterize the approximation and generalization errors of the pretrained LLM. Notably, we construct a transformer model that approximates the target distribution of the multi-step reasoning problem with an error that decreases exponentially in the number of transformer blocks. Our analysis extends to other variants of CoT, including Self-Consistent CoT, Tree-of-Thought, and Selection-Inference, offering a broad perspective on the efficacy of these methods. We also provide numerical experiments to validate the theoretical findings.

Penghui Yang, Cunxiao Du, Fengzhuo Zhang et al.

As Large Language Models (LLMs) can now process extremely long contexts, efficient inference over these extended inputs has become increasingly important, especially for emerging applications like LLM agents that highly depend on this capability. Speculative decoding (SD) offers a promising lossless acceleration technique compared to lossy alternatives such as quantization and model cascades. However, most state-of-the-art SD methods are trained on short texts (typically fewer than 4k tokens), making them unsuitable for long-context scenarios. Specifically, adapting these methods to long contexts presents three key challenges: (1) the excessive memory demands posed by draft models due to large Key-Value (KV) cache; (2) performance degradation resulting from the mismatch between short-context training and long-context inference; and (3) inefficiencies in tree attention mechanisms when managing long token sequences. This work introduces LongSpec, a framework that addresses these challenges through three core innovations: a memory-efficient draft model with a constant-sized KV cache; novel position indices that mitigate the training-inference mismatch; and an attention aggregation strategy that combines fast prefix computation with standard tree attention to enable efficient decoding. Experimental results confirm the effectiveness of LongSpec, achieving up to a 3.26x speedup over strong Flash Attention baselines across five long-context understanding datasets, as well as a 2.25x reduction in wall-clock time on the AIME24 long reasoning task with the QwQ model, demonstrating significant latency improvements for long-context applications. The code is available at https://github.com/sail-sg/LongSpec.

Xingyao Li, Fengzhuo Zhang, Cunxiao Du et al.

Despite significant progress in autoregressive image generation, inference remains slow due to the sequential nature of AR models and the ambiguity of image tokens, even when using speculative decoding. Recent works attempt to address this with relaxed speculative decoding but lack theoretical grounding. In this paper, we establish the theoretical basis of relaxed SD and propose COOL-SD, an annealed relaxation of speculative decoding built on two key insights. The first analyzes the total variation (TV) distance between the target model and relaxed speculative decoding and yields an optimal resampling distribution that minimizes an upper bound of the distance. The second uses perturbation analysis to reveal an annealing behaviour in relaxed speculative decoding, motivating our annealed design. Together, these insights enable COOL-SD to generate images faster with comparable quality, or achieve better quality at similar latency. Experiments validate the effectiveness of COOL-SD, showing consistent improvements over prior methods in speed-quality trade-offs.

Xuan Zhang, Fengzhuo Zhang, Cunxiao Du et al.

Scaling language models to handle longer contexts introduces substantial memory challenges due to the growing cost of key-value (KV) caches. Motivated by the efficiency gains of hybrid models and the broad availability of pretrained large transformer backbones, we explore transitioning transformer models into hybrid architectures for a more efficient generation. In this work, we propose LightTransfer, a lightweight method that transforms models such as LLaMA into hybrid variants. Our approach identifies lazy layers -- those focusing on recent or initial tokens -- and replaces their full attention with streaming attention. This transformation can be performed without any training for long-context understanding tasks or with minimal fine-tuning for o1-like long reasoning generation tasks that require stronger reasoning capabilities. Experiments across diverse benchmarks and models (e.g., LLaMA, Mistral, QwQ-STILL) demonstrate that, even when half of the layers are identified as lazy, LightTransfer achieves up to 2.17$\times$ throughput improvement with minimal performance loss ($<1.5\%$ on LongBench) and achieves 53.3\% on math benchmark AIME24 of advanced o1-like long reasoning model QwQ-STILL.

Yunlong Hou, Fengzhuo Zhang, Cunxiao Du et al.

Speculative decoding has emerged as a popular method to accelerate the inference of Large Language Models (LLMs) while retaining their superior text generation performance. Previous methods either adopt a fixed speculative decoding configuration regardless of the prefix tokens, or train draft models in an offline or online manner to align them with the context. This paper proposes a training-free online learning framework to adaptively choose the configuration of the hyperparameters for speculative decoding as text is being generated. We first formulate this hyperparameter selection problem as a Multi-Armed Bandit problem and provide a general speculative decoding framework BanditSpec. Furthermore, two bandit-based hyperparameter selection algorithms, UCBSpec and EXP3Spec, are designed and analyzed in terms of a novel quantity, the stopping time regret. We upper bound this regret under both stochastic and adversarial reward settings. By deriving an information-theoretic impossibility result, it is shown that the regret performance of UCBSpec is optimal up to universal constants. Finally, extensive empirical experiments with LLaMA3 and Qwen2 demonstrate that our algorithms are effective compared to existing methods, and the throughput is close to the oracle best hyperparameter in simulated real-life LLM serving scenarios with diverse input prompts.

Jing Wang, Fengzhuo Zhang, Xiaoli Li et al.

A variety of Auto-Regressive Video Diffusion Models (ARVDM) have achieved remarkable successes in generating realistic long-form videos. However, theoretical analyses of these models remain scant. In this work, we develop theoretical underpinnings for these models and use our insights to improve the performance of existing models. We first develop Meta-ARVDM, a unified framework of ARVDMs that subsumes most existing methods. Using Meta-ARVDM, we analyze the KL-divergence between the videos generated by Meta-ARVDM and the true videos. Our analysis uncovers two important phenomena inherent to ARVDM -- error accumulation and memory bottleneck. By deriving an information-theoretic impossibility result, we show that the memory bottleneck phenomenon cannot be avoided. To mitigate the memory bottleneck, we design various network structures to explicitly use more past frames. We also achieve a significantly improved trade-off between the mitigation of the memory bottleneck and the inference efficiency by compressing the frames. Experimental results on DMLab and Minecraft validate the efficacy of our methods. Our experiments also demonstrate a Pareto-frontier between the error accumulation and memory bottleneck across different methods.

Shuche Wang, Fengzhuo Zhang, Jiaxiang Li et al.

The Muon optimizer is consistently faster than Adam in training Large Language Models (LLMs), yet the mechanism underlying its success remains unclear. This paper demystifies this mechanism through the lens of associative memory. By ablating the transformer components optimized by Muon, we reveal that the associative memory parameters of LLMs, namely the Value and Output (VO) attention weights and Feed-Forward Networks (FFNs), are the primary contributors to Muon's superiority. Motivated by this associative memory view, we then explain Muon's superiority on real-world corpora, which are intrinsically heavy-tailed: a few classes (tail classes) appear far less frequently than others. The superiority is explained through two key properties: (i) its update rule consistently yields a more isotropic singular spectrum than Adam; and as a result, (ii) on heavy-tailed data, it optimizes tail classes more effectively than Adam. Beyond empirical evidence, we theoretically confirm these findings by analyzing a one-layer associative memory model under class-imbalanced data. We prove that Muon consistently achieves balanced learning across classes regardless of feature embeddings, whereas Adam can induce large disparities in learning errors depending on embedding properties. In summary, our empirical observations and theoretical analyses reveal Muon's core advantage: its update rule aligns with the outer-product structure of linear associative memories, enabling more balanced and effective learning of tail classes in heavy-tailed distributions than Adam.

Xuan Zhang, Cunxiao Du, Sicheng Yu et al.

Due to the auto-regressive nature of current video large language models (Video-LLMs), the inference latency increases as the input sequence length grows, posing challenges for the efficient processing of video sequences that are usually very long. We observe that during decoding, the attention scores of most tokens in Video-LLMs tend to be sparse and concentrated, with only certain tokens requiring comprehensive full attention. Based on this insight, we introduce Sparse-to-Dense (StD), a novel decoding strategy that integrates two distinct modules: one leveraging sparse top-K attention and the other employing dense full attention. These modules collaborate to accelerate Video-LLMs without loss. The fast (sparse) model speculatively decodes multiple tokens, while the slow (dense) model verifies them in parallel. StD is a tuning-free, plug-and-play solution that achieves up to a 1.94$\times$ walltime speedup in video processing. It maintains model performance while enabling a seamless transition from a standard Video-LLM to a sparse Video-LLM with minimal code modifications.

Xingyao Li, Fengzhuo Zhang, Jiachun Pan et al.

Despite the considerable progress achieved in the long video generation problem, there is still significant room to improve the consistency of the generated videos, particularly in terms of their smoothness and transitions between scenes. We address these issues to enhance the consistency and coherence of videos generated with either single or multiple prompts. We propose the Time-frequency based temporal Attention Reweighting Algorithm (TiARA), which judiciously edits the attention score matrix based on the Discrete Short-Time Fourier Transform. This method is supported by a frequency-based analysis, ensuring that the edited attention score matrix achieves improved consistency across frames. It represents the first-of-its-kind for frequency-based methods in video diffusion models. For videos generated by multiple prompts, we further uncover key factors such as the alignment of the prompts affecting prompt interpolation quality. Inspired by our analyses, we propose PromptBlend, an advanced prompt interpolation pipeline that systematically aligns the prompts. Extensive experimental results validate the efficacy of our proposed method, demonstrating consistent and substantial improvements over multiple baselines.

Yufeng Zhang, Fengzhuo Zhang, Zhuoran Yang et al.

In this paper, we conduct a comprehensive study of In-Context Learning (ICL) by addressing several open questions: (a) What type of ICL estimator is learned by large language models? (b) What is a proper performance metric for ICL and what is the error rate? (c) How does the transformer architecture enable ICL? To answer these questions, we adopt a Bayesian view and formulate ICL as a problem of predicting the response corresponding to the current covariate, given a number of examples drawn from a latent variable model. To answer (a), we show that, without updating the neural network parameters, ICL implicitly implements the Bayesian model averaging algorithm, which is proven to be approximately parameterized by the attention mechanism. For (b), we analyze the ICL performance from an online learning perspective and establish a $\mathcal{O}(1/T)$ regret bound for perfectly pretrained ICL, where $T$ is the number of examples in the prompt. To answer (c), we show that, in addition to encoding Bayesian model averaging via attention, the transformer architecture also enables a fine-grained statistical analysis of pretraining under realistic assumptions. In particular, we prove that the error of pretrained model is bounded by a sum of an approximation error and a generalization error, where the former decays to zero exponentially as the depth grows, and the latter decays to zero sublinearly with the number of tokens in the pretraining dataset. Our results provide a unified understanding of the transformer and its ICL ability with bounds on ICL regret, approximation, and generalization, which deepens our knowledge of these essential aspects of modern language models.

Fengzhuo Zhang, Anshoo Tandon, Vincent Y. F. Tan

The Chow-Liu algorithm (IEEE Trans.~Inform.~Theory, 1968) has been a mainstay for the learning of tree-structured graphical models from i.i.d.\ sampled data vectors. Its theoretical properties have been well-studied and are well-understood. In this paper, we focus on the class of trees that are arguably even more fundamental, namely {\em homogeneous} trees in which each pair of nodes that forms an edge has the same correlation $ρ$. We ask whether we are able to further reduce the error probability of learning the structure of the homogeneous tree model when {\em active learning} or {\em active sampling of nodes or variables} is allowed. Our figure of merit is the {\em error exponent}, which quantifies the exponential rate of decay of the error probability with an increasing number of data samples. At first sight, an improvement in the error exponent seems impossible, as all the edges are statistically identical. We design and analyze an algorithm Active Learning Algorithm for Trees with Homogeneous Edge (Active-LATHE), which surprisingly boosts the error exponent by at least 40\% when $ρ$ is at least $0.8$. For all other values of $ρ$, we also observe commensurate, but more modest, improvements in the error exponent. Our analysis hinges on judiciously exploiting the minute but detectable statistical variation of the samples to allocate more data to parts of the graph in which we are less confident of being correct.

Fengzhuo Zhang, Vincent Y. F. Tan

We consider learning the structures of Gaussian latent tree models with vector observations when a subset of them are arbitrarily corrupted. First, we present the sample complexities of Recursive Grouping (RG) and Chow-Liu Recursive Grouping (CLRG) without the assumption that the effective depth is bounded in the number of observed nodes, significantly generalizing the results in Choi et al. (2011). We show that Chow-Liu initialization in CLRG greatly reduces the sample complexity of RG from being exponential in the diameter of the tree to only logarithmic in the diameter for the hidden Markov model (HMM). Second, we robustify RG, CLRG, Neighbor Joining (NJ) and Spectral NJ (SNJ) by using the truncated inner product. These robustified algorithms can tolerate a number of corruptions up to the square root of the number of clean samples. Finally, we derive the first known instance-dependent impossibility result for structure learning of latent trees. The optimalities of the robust version of CLRG and NJ are verified by comparing their sample complexities and the impossibility result.