Disordered Systems & Neural Networks

Henry Hunt, Mason Kamb, Surya Ganguli

For the machine learning community, this provides a theoretical framework explaining generalization in diffusion models, addressing a fundamental mystery in generative AI.

Lorenzo Bardone, Claudia Merger, Sebastian Goldt

This provides foundational insights into how diffusion models learn complex distributions, which is incremental but clarifies a key mechanism for the AI/ML community.

Xin Du, Kumiko Tanaka-Ishii

For practitioners of LLM generation, this provides a practical method to maintain diversity and quality even at very low entropy settings.

Kaito Takanami, Cengiz Pehlevan

Provides a unified theoretical understanding of how test-time reasoning depth affects generalization in LLMs, addressing a poorly understood scaling behavior.

Fiona Y. Wang, Lee Marom, Subhadeep Pal et al.

This addresses the challenge of scalable, decentralized scientific discovery for researchers, though it appears incremental as an extension of existing agent-based and provenance-aware systems.

Luca Maria Del Bono, Giulio Biroli, Patrick Charbonneau et al.

Provides theoretical insight into the limitations of diffusion models near criticality and demonstrates how architectural design can overcome these bottlenecks, relevant for statistical physics and generative modeling.

Dayal Singh Kalra, Maissam Barkeshli

For practitioners training large language models, this work clarifies the mechanism behind μP's success and provides a simple fix (adjusting embedding learning rate) to improve hyperparameter transfer in standard parameterization.

Brice Huang, Mark Sellke

This work provides theoretical evidence for computational hardness in spin glass optimization, impacting physics and algorithm design, though it is incremental on prior conjectures.

Ali Hussaini Umar, Alessandro Laio

This work provides insights for practitioners into how data quality and quantity influence neural network representations, decoupling alignment from generalization.

Amrut Nadgir, Vijay Balasubramanian, Pratik Chaudhari

This provides theoretical insights into CoT methods for improving LLM performance, but it is incremental as it builds on existing observations without introducing new practical techniques.

Fabiola Ricci, Claudia Merger, Sebastian Goldt

For researchers studying learning dynamics and generalization in neural networks, this work provides a theoretical and experimental framework linking Fourier properties of data to sample complexity and training speed.

Clarissa Lauditi, Cengiz Pehlevan, Blake Bordelon

Provides a theoretical framework for understanding feature learning and hyperparameter transfer in wide neural networks, relevant to practitioners scaling up models.

Antoine Maillard, Sebastian Goldt

This work clarifies the fundamental distinction between convergence and latent factor recovery in generative models, providing theoretical insights for practitioners regarding data requirements and evaluation metrics.

Hyunmo Kang, Noam Itzhak Levi, Corinna Elena Wegner et al.

For researchers in generative modeling and sampling, this work provides theoretical and empirical insights into the mixing behavior of diffusion-based samplers, though the findings are primarily diagnostic rather than offering a practical improvement.

Xiao-Liang Qi

For the scientific community, this article provides a conceptual framework for understanding AI's transformative role in research, though it is a perspective piece without empirical results.

Hidenori Tanaka

This provides a framework for understanding social representation formation in multi-agent systems, with implications for AI deployment in decision-making, though it is incremental as it builds on prior naming-game studies.

Enrico Ventura, Beatrice Achilli, Luca Ambrogioni et al.

For practitioners using CFG in diffusion models, this work provides a theoretical understanding of diversity loss and a practical fix, though the analysis is limited to Gaussian mixtures.

Jack T. Parley, Francesco Cagnetta, Matthieu Wyart

For researchers in cognitive science and machine learning, this work provides a theoretical and empirical framework to understand how deep networks learn syntax from local statistics, though it is incremental as it builds on known PCFG testbeds.

Louie Hong Yao, Yuhao Li, Shengchao Liu

For researchers studying self-supervised learning, this work provides a theoretical understanding of collapse mechanisms and prevention, though the model is highly simplified.

Ejaaz Merali, Mohamed Hibat-Allah, Mohammad Kohandel et al.

This work makes recurrent neural network quantum states practical for large-scale quantum many-body simulations, addressing a scalability bottleneck.