arXiv subjects

Zixuan Gong

Publications and source records attributed to Zixuan Gong.

14 recordsLinked to original sources

arXiv2026

Disentangling Feature Structure: A Mathematically Provable Two-Stage Training Dynamics in Transformers

Transformers may exhibit two-stage training dynamics during the real-world training process. For instance, when training GPT-2 on the Counterfact dataset, the answers progress from syntactically incorrect to syntactically correct to semantically correct. However, existing theoretical analyses hardly account for this feature-level two-stage phenomenon, which could be conceptually attributed to disentangled two-type features like syntax and semantics. In this paper, we theoretically demonstrate how the two-stage training dynamics potentially occur in transformers. Specifically, we analyze the feature learning dynamics induced by the aforementioned disentangled two-type feature structure, grounding our analysis in a simplified yet illustrative setting that comprises normalized ReLU self-attention and structured data. Such disentanglement of feature structure is general in practice, e.g., natural languages contain syntax and semantics, and proteins contain primary and secondary structures. To our best knowledge, this is the first rigorous result regarding a feature-level two-stage optimization process in transformers within this theoretical framework. A corollary further indicates that such a two-stage process is closely related to the spectral properties of attention weights.

cs.CL↗

arXiv2026

Sparsity is Combinatorial Depth: Quantifying MoE Expressivity via Tropical Geometry

While Mixture-of-Experts (MoE) architectures define the state-of-the-art, their theoretical success is often attributed to heuristic efficiency rather than geometric expressivity. In this work, we present the first analysis of MoE through the lens of tropical geometry, establishing that the Top-$k$ routing mechanism is algebraically isomorphic to the $k$-th elementary symmetric tropical polynomial. This isomorphism partitions the input space into the Normal Fan of a Hypersimplex, revealing that \textbf{sparsity is combinatorial depth} which scales geometric capacity by the binomial coefficient $\binom{N}{k}$. Moving beyond ambient bounds, we introduce the concept of \textit{Effective Capacity} under the Manifold Hypothesis. We prove that while dense networks suffer from capacity collapse on low-dimensional data, MoE architectures exhibit \textit{Combinatorial Resilience}, maintaining high expressivity via the transversality of routing cones. Translating these theoretical bounds into architectural principles, we derive asymptotic capacity limits for optimal expert granularity and prove that shared experts are geometrically necessary to prevent routing collapse.

cs.LG↗

arXiv2026

Beyond the Black Box: A Survey on the Theory and Mechanism of Large Language Models

The rapid emergence of Large Language Models (LLMs) has precipitated a profound paradigm shift in Artificial Intelligence, delivering monumental engineering successes that increasingly impact modern society. However, a critical paradox persists within the current field: despite the empirical efficacy, our theoretical understanding of LLMs remains disproportionately nascent, forcing these systems to be treated largely as ``black boxes''. To address this theoretical fragmentation, this survey proposes a unified lifecycle-based taxonomy that organizes the research landscape into six distinct stages: Data Preparation, Model Preparation, Training, Alignment, Inference, and Evaluation. Within this framework, we provide a systematic review of the foundational theories and internal mechanisms driving LLM performance. Specifically, we analyze core theoretical issues such as the mathematical justification for data mixtures, the representational limits of various architectures, and the optimization dynamics of alignment algorithms. Moving beyond current best practices, we identify critical frontier challenges, including the theoretical limits of synthetic data self-improvement, the mathematical bounds of safety guarantees, and the mechanistic origins of emergent intelligence. By connecting empirical observations with rigorous scientific inquiry, this work provides a structured roadmap for transitioning LLM development from engineering heuristics toward a principled scientific discipline.

cs.CL↗

arXiv2026

Effective Frontiers: A Unification of Neural Scaling Laws

Neural scaling laws govern the prediction power-law improvement of test loss with respect to model capacity ($N$), datasize ($D$), and compute ($C$). However, existing theoretical explanations often rely on specific architectures or complex kernel methods, lacking intuitive universality. In this paper, we propose a unified framework that abstracts general learning tasks as the progressive coverage of patterns from a long-tail (Zipfian) distribution. We introduce the Effective Frontier ($k_\star$), a threshold in the pattern rank space that separates learned knowledge from the unlearned tail. We prove that reducible loss is asymptotically determined by the probability mass of the tail a resource-dependent frontier truncation. Based on our framework, we derive the precise scaling laws for $N$, $D$, and $C$, attributing them to capacity, coverage, and optimization bottlenecks, respectively. Furthermore, we unify these mechanisms via a Max-Bottleneck principle, demonstrating that the Kaplan and Chinchilla scaling laws are not contradictory, but equilibrium solutions to the same constrained optimization problem under different active bottlenecks.

cs.LG↗

arXiv2026

What Makes Looped Transformers Perform Better Than Non-Recursive Ones

While looped transformers (termed as Looped-Attn) often outperform standard transformers (termed as Single-Attn) on complex reasoning tasks, the mechanism for this advantage remains underexplored. In this paper, we explain this phenomenon through the lens of loss landscape geometry, inspired by empirical observations of their distinct dynamics at both sample and Hessian levels. To formalize this, we extend the River-Valley landscape model by distinguishing between U-shaped valleys (flat) and V-shaped valleys (steep). Based on empirical observations, we conjecture that the recursive architecture of Looped-Attn induces a landscape-level inductive bias towards River-V-Valley. This inductive bias suggest a better loss convergence along the river due to valley hopping, and further encourage learning about complex patterns compared to the River-U-Valley induced by Single-Attn. Building on this insight, we propose SHIFT (Staged HIerarchical Framework for Progressive Training), a principled training strategy that accelerates the training process of Looped-Attn while achieving comparable performances.

cs.LG↗

arXiv2025

Efficient Diffusion Models: A Survey

Diffusion models have emerged as powerful generative models capable of producing high-quality contents such as images, videos, and audio, demonstrating their potential to revolutionize digital content creation. However, these capabilities come at the cost of their significant computational resources and lengthy generation time, underscoring the critical need to develop efficient techniques for practical deployment. In this survey, we provide a systematic and comprehensive review of research on efficient diffusion models. We organize the literature in a taxonomy consisting of three main categories, covering distinct yet interconnected efficient diffusion model topics from algorithm-level, system-level, and framework perspective, respectively. We have also created a GitHub repository where we organize the papers featured in this survey at https://github.com/AIoT-MLSys-Lab/Efficient-Diffusion-Model-Survey. We hope our survey can serve as a valuable resource to help researchers and practitioners gain a systematic understanding of efficient diffusion model research and inspire them to contribute to this important and exciting field.

cs.LG↗

arXiv2025

MindSimulator: Exploring Brain Concept Localization via Synthetic FMRI

Concept-selective regions within the human cerebral cortex exhibit significant activation in response to specific visual stimuli associated with particular concepts. Precisely localizing these regions stands as a crucial long-term goal in neuroscience to grasp essential brain functions and mechanisms. Conventional experiment-driven approaches hinge on manually constructed visual stimulus collections and corresponding brain activity recordings, constraining the support and coverage of concept localization. Additionally, these stimuli often consist of concept objects in unnatural contexts and are potentially biased by subjective preferences, thus prompting concerns about the validity and generalizability of the identified regions. To address these limitations, we propose a data-driven exploration approach. By synthesizing extensive brain activity recordings, we statistically localize various concept-selective regions. Our proposed MindSimulator leverages advanced generative technologies to learn the probability distribution of brain activity conditioned on concept-oriented visual stimuli. This enables the creation of simulated brain recordings that reflect real neural response patterns. Using the synthetic recordings, we successfully localize several well-studied concept-selective regions and validate them against empirical findings, achieving promising prediction accuracy. The feasibility opens avenues for exploring novel concept-selective regions and provides prior hypotheses for future neuroscience research.

q-bio.NC↗

arXiv2025

Towards Auto-Regressive Next-Token Prediction: In-Context Learning Emerges from Generalization

Large language models (LLMs) have demonstrated remarkable in-context learning (ICL) abilities. However, existing theoretical analysis of ICL primarily exhibits two limitations: (a) Limited i.i.d. Setting. Most studies focus on supervised function learning tasks where prompts are constructed with i.i.d. input-label pairs. This i.i.d. assumption diverges significantly from real language learning scenarios where prompt tokens are interdependent. (b) Lack of Emergence Explanation. Most literature answers what ICL does from an implicit optimization perspective but falls short in elucidating how ICL emerges and the impact of pre-training phase on ICL. In our paper, to extend (a), we adopt a more practical paradigm, auto-regressive next-token prediction (AR-NTP), which closely aligns with the actual training of language models. Specifically, within AR-NTP, we emphasize prompt token-dependency, which involves predicting each subsequent token based on the preceding sequence. To address (b), we formalize a systematic pre-training and ICL framework, highlighting the layer-wise structure of sequences and topics, alongside a two-level expectation. In conclusion, we present data-dependent, topic-dependent and optimization-dependent PAC-Bayesian generalization bounds for pre-trained LLMs, investigating that ICL emerges from the generalization of sequences and topics. Our theory is supported by experiments on numerical linear dynamic systems, synthetic GINC and real-world language datasets.

cs.CL↗

arXiv2024

MindTuner: Cross-Subject Visual Decoding with Visual Fingerprint and Semantic Correction

Decoding natural visual scenes from brain activity has flourished, with extensive research in single-subject tasks and, however, less in cross-subject tasks. Reconstructing high-quality images in cross-subject tasks is a challenging problem due to profound individual differences between subjects and the scarcity of data annotation. In this work, we proposed MindTuner for cross-subject visual decoding, which achieves high-quality and rich semantic reconstructions using only 1 hour of fMRI training data benefiting from the phenomena of visual fingerprint in the human visual system and a novel fMRI-to-text alignment paradigm. Firstly, we pre-train a multi-subject model among 7 subjects and fine-tune it with scarce data on new subjects, where LoRAs with Skip-LoRAs are utilized to learn the visual fingerprint. Then, we take the image modality as the intermediate pivot modality to achieve fMRI-to-text alignment, which achieves impressive fMRI-to-text retrieval performance and corrects fMRI-to-image reconstruction with fine-tuned semantics. The results of both qualitative and quantitative analyses demonstrate that MindTuner surpasses state-of-the-art cross-subject visual decoding models on the Natural Scenes Dataset (NSD), whether using training data of 1 hour or 40 hours.

cs.CV↗

arXiv2024

Wills Aligner: Multi-Subject Collaborative Brain Visual Decoding

Decoding visual information from human brain activity has seen remarkable advancements in recent research. However, the diversity in cortical parcellation and fMRI patterns across individuals has prompted the development of deep learning models tailored to each subject. The personalization limits the broader applicability of brain visual decoding in real-world scenarios. To address this issue, we introduce Wills Aligner, a novel approach designed to achieve multi-subject collaborative brain visual decoding. Wills Aligner begins by aligning the fMRI data from different subjects at the anatomical level. It then employs delicate mixture-of-brain-expert adapters and a meta-learning strategy to account for individual fMRI pattern differences. Additionally, Wills Aligner leverages the semantic relation of visual stimuli to guide the learning of inter-subject commonality, enabling visual decoding for each subject to draw insights from other subjects' data. We rigorously evaluate our Wills Aligner across various visual decoding tasks, including classification, cross-modal retrieval, and image reconstruction. The experimental results demonstrate that Wills Aligner achieves promising performance.

cs.CV↗

arXiv2024

NeuroClips: Towards High-fidelity and Smooth fMRI-to-Video Reconstruction

Reconstruction of static visual stimuli from non-invasion brain activity fMRI achieves great success, owning to advanced deep learning models such as CLIP and Stable Diffusion. However, the research on fMRI-to-video reconstruction remains limited since decoding the spatiotemporal perception of continuous visual experiences is formidably challenging. We contend that the key to addressing these challenges lies in accurately decoding both high-level semantics and low-level perception flows, as perceived by the brain in response to video stimuli. To the end, we propose NeuroClips, an innovative framework to decode high-fidelity and smooth video from fMRI. NeuroClips utilizes a semantics reconstructor to reconstruct video keyframes, guiding semantic accuracy and consistency, and employs a perception reconstructor to capture low-level perceptual details, ensuring video smoothness. During inference, it adopts a pre-trained T2V diffusion model injected with both keyframes and low-level perception flows for video reconstruction. Evaluated on a publicly available fMRI-video dataset, NeuroClips achieves smooth high-fidelity video reconstruction of up to 6s at 8FPS, gaining significant improvements over state-of-the-art models in various metrics, e.g., a 128% improvement in SSIM and an 81% improvement in spatiotemporal metrics. Our project is available at https://github.com/gongzix/NeuroClips.

eess.IV↗

arXiv2024

Lite-Mind: Towards Efficient and Robust Brain Representation Network

The limited data availability and the low signal-to-noise ratio of fMRI signals lead to the challenging task of fMRI-to-image retrieval. State-of-the-art MindEye remarkably improves fMRI-to-image retrieval performance by leveraging a large model, i.e., a 996M MLP Backbone per subject, to align fMRI embeddings to the final hidden layer of CLIP's Vision Transformer (ViT). However, significant individual variations exist among subjects, even under identical experimental setups, mandating the training of large subject-specific models. The substantial parameters pose significant challenges in deploying fMRI decoding on practical devices. To this end, we propose Lite-Mind, a lightweight, efficient, and robust brain representation learning paradigm based on Discrete Fourier Transform (DFT), which efficiently aligns fMRI voxels to fine-grained information of CLIP. We elaborately design a DFT backbone with Spectrum Compression and Frequency Projector modules to learn informative and robust voxel embeddings. Our experiments demonstrate that Lite-Mind achieves an impressive 94.6% fMRI-to-image retrieval accuracy on the NSD dataset for Subject 1, with 98.7% fewer parameters than MindEye. Lite-Mind is also proven to be able to be migrated to smaller fMRI datasets and establishes a new state-of-the-art for zero-shot classification on the GOD dataset.

cs.CV↗

arXiv2024

MLIP: Efficient Multi-Perspective Language-Image Pretraining with Exhaustive Data Utilization

Contrastive Language-Image Pretraining (CLIP) has achieved remarkable success, leading to rapid advancements in multimodal studies. However, CLIP faces a notable challenge in terms of inefficient data utilization. It relies on a single contrastive supervision for each image-text pair during representation learning, disregarding a substantial amount of valuable information that could offer richer supervision. Additionally, the retention of non-informative tokens leads to increased computational demands and time costs, particularly in CLIP's ViT image encoder. To address these issues, we propose Multi-Perspective Language-Image Pretraining (MLIP). In MLIP, we leverage the frequency transform's sensitivity to both high and low-frequency variations, which complements the spatial domain's sensitivity limited to low-frequency variations only. By incorporating frequency transforms and token-level alignment, we expand CILP's single supervision into multi-domain and multi-level supervision, enabling a more thorough exploration of informative image features. Additionally, we introduce a token merging method guided by comprehensive semantics from the frequency and spatial domains. This allows us to merge tokens to multi-granularity tokens with a controllable compression rate to accelerate CLIP. Extensive experiments validate the effectiveness of our design.

cs.CV↗

arXiv2024

Multimodal Federated Learning with Missing Modality via Prototype Mask and Contrast

In real-world scenarios, multimodal federated learning often faces the practical challenge of intricate modality missing, which poses constraints on building federated frameworks and significantly degrades model inference accuracy. Existing solutions for addressing missing modalities generally involve developing modality-specific encoders on clients and training modality fusion modules on servers. However, these methods are primarily constrained to specific scenarios with either unimodal clients or complete multimodal clients, struggling to generalize effectively in the intricate modality missing scenarios. In this paper, we introduce a prototype library into the FedAvg-based Federated Learning framework, thereby empowering the framework with the capability to alleviate the global model performance degradation resulting from modality missing during both training and testing. The proposed method utilizes prototypes as masks representing missing modalities to formulate a task-calibrated training loss and a model-agnostic uni-modality inference strategy. In addition, a proximal term based on prototypes is constructed to enhance local training. Experimental results demonstrate the state-of-the-art performance of our approach. Compared to the baselines, our method improved inference accuracy by 3.7\% with 50\% modality missing during training and by 23.8\% during uni-modality inference. Code is available at https://github.com/BaoGuangYin/PmcmFL.

cs.LG↗