Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes
Score
77首次系统解构多模态预训练中的知识流与协同机制，提供了此前缺乏的实证规律。
受控实验与大规模验证相结合，方法严谨，结论清晰，论文结构完整。
给出了可直接落地的预训练配方与架构选择建议，能显著节省计算资源。
多模态统一预训练是当前热点，该研究提供了及时的理论与实验参考。
来自 arXiv 预印本，虽未经同行评审，但实验规模与作者背景可信度较高。
Base 85 × T2 Source weight ×0.9 = 77
Curation line 70Selected
Why selected
The study fills a critical gap in the mechanistic understanding of multimodal pretraining, combining rigorous controlled experiments with large-scale validation, offering high reference value for both researchers and engineers.
AI commentary
「This paper pushes multimodal pretraining from 'alchemy' toward 'physics', with four solid empirical findings. The 'vision laziness' phenomenon is a particularly useful warning that late fusion may lead models to lazily rely on language priors—a caution for teams seeking shortcuts. Highly recommended for algorithm engineers designing multimodal architectures and researchers pursuing efficiency.」
Quick overview
This paper systematically explores the design space and fundamental mechanisms of multimodal pretraining through controlled experiments on synthetic and large-scale real-world datasets. Four key insights emerge: knowledge flow across modalities shows distinct asymmetric influence patterns; data complexity determines whether modalities are synergistic or competitive, with shared attention and modality-specific feed-forward layers promoting synergy; early joint training outperforms late alignment or sequential training, revealing a 'vision laziness' phenomenon; and derived recipes achieve strong generative performance using only 5% of the compute budget. These findings are validated at scale by training multiple 13.5B MoE models on 2T tokens.
- Knowledge Flow: Language, visual understanding, and visual generation transfer knowledge across modalities with distinct asymmetric influence patterns.
- Synergy vs. Competition: Data complexity determines modality relationships; shared attention with modality-specific feed-forward layers promotes synergy, generalizing across visual tokenizer designs.
- Early Unification: Joint training from the very early stages outperforms late alignment or sequential training, uncovering a 'vision laziness' phenomenon.
- Efficient Recipes: Strong generative performance is achieved with only 5% of the compute budget, validated on 13.5B MoE models trained on 2T tokens.
Full text
This is an AI-generated Chinese digest; use "View original" for the source text.
Related items
85
FeaturedSetoka: A Benchmark for Hierarchical User Understanding in Personalized Agents over Heterogeneous Data
Setoka：面向异构数据个性化Agent的层级用户理解基准
This paper is well worth attention, especially for researchers and engineers working on personalized agents, dialogue systems, and memory-augmented AI. It not only provides a benchmark but defines levels of user understanding from psychology, revealing systematic failures of current systems at higher levels. If you are designing memory-based agents, this paper reminds you that memory systems need to support cross-temporal, cross-source abstract reasoning, not just retrieval.
85
FeaturedBAND: Bayesian Network Distribution Regression for High-Dimensional Distribution Estimation – Breaking the Curse
BAND：高维分布估计的贝叶斯网络回归方法——打破维度诅咒
This paper is a must-read for anyone working on high-dimensional statistics, probabilistic modeling, and generative AI. It provides a simple yet powerful theoretical answer: sparsity can break the curse of dimensionality, and Bayesian networks offer a natural framework. Compared to black-box methods like VAEs and GANs, BAND is interpretable and provably convergent, especially valuable for industry needs like reliable confidence intervals. If you research time series forecasting or risk management, this paper may change how you approach high-dimensional density estimation.
84
FeaturedCycle-Consistent and Uncertainty-Aware Neural Surrogates for Tokamak Edge Plasmas
循环一致与不确定性感知的托卡马克边缘等离子体神经代理模型
This paper is worth attention because it demonstrates how advanced AI techniques (cycle consistency, uncertainty quantification) directly address a core computational bottleneck in nuclear fusion. For fusion researchers and engineers, it offers a practical tool to speed up simulations by orders of magnitude without sacrificing accuracy; for AI practitioners, it provides an excellent case of self-supervised learning and inverse problem solving in scientific applications.
83
FeaturedEquivariant Learning of a Transferable Three-Dimensional Classical Density Functional
等变学习可迁移的三维经典密度泛函
The most striking aspect of this paper is its direct coupling of machine learning to a core problem in statistical mechanics, using equivariant networks to learn a transferable density functional without any labels. For researchers in liquid physics or soft matter, this could be a new paradigm: do we need to run separate simulations for every new system anymore? For AI researchers, it also shows how to embed physical symmetries and variational principles into models rather than just fitting data. I recommend focusing on the methods, especially how equivariance and variational consistency are implemented—that's the real essence.
83
FeaturedComputational and Statistical Guarantees of the c-Rectified Flow
c-Rectified Flow 的计算与统计保证
This paper moves rectified flow from 'works in practice' to 'theoretically controllable,' which is especially important for those working on the underlying algorithms of generative models. If you are tuning pipelines like FLUX or SD3, understanding that ordinary rectified flow can fail when covariances do not commute may help you avoid hidden training pitfalls.
83
FeaturedLearning Deployable Manipulation Policies from High-Fidelity UMI Data Alone
从高保真UMI数据学习可部署操作策略
The most striking point of this paper is proving that pure robot-free data can match real teleoperation performance, breaking the convention of needing a small real-robot anchor for post-training. For research groups looking to reduce hardware costs and industrial robot application developers, HiFi-UMI's data production system and open-source dataset are directly usable resources. Pay attention to the engineering details of head-mounted SLAM and microsecond synchronization—they are key to achieving the reported accuracy.
83
FeaturedCapital Markets LLM Reliability Score: From Plausible to Bankable
资本市场LLM可靠性评分：从“看似合理”到“可银行化”
This paper is valuable for both finance and AI communities: it provides an actionable, workflow-oriented evaluation framework that's closer to real needs than generic benchmarks. If you're deploying LLMs in fintech or responsible for explainability/auditability, this is a must-read. The finding on decision usefulness is particularly striking—judges agree it's critical, yet models vary the most on it.
83
FeaturedWhere Should Optimizer State Live? Tiered State Allocation for Memory-Efficient Mixture-of-Experts Training
优化器状态应放在哪里？面向内存高效MoE训练的分层状态分配
This paper is highly valuable for teams training MoE models: it demonstrates that with careful optimizer state allocation, one can dramatically reduce memory without sacrificing accuracy—and even improve performance. It is especially relevant for labs or enterprises constrained by 40 GB or 80 GB GPU memory. The tiered strategy and ablation on momentum retention are worth studying for optimizing training frameworks like DeepSpeed or Megatron.