Yixiong's Blog

神经网络通过数据学习规律，会学习到什么规律？本文提出了新的神经网络MASS（Multi-physics AI Scalar Scientist）用以学习网络从数据中学到的模型信息。这篇工作属于PINN。

文献： * Do Two AI Scientists Agree? * 源代码GitHub * Hamiltonian Neural Networks * HNN GitHub * Lagrangian Neural Networks * LNN GitHUb

通过强化学习微调生成模型，使得生成的拓扑绝缘体和拓扑材料结构准确率上升。

Reference: * Design Topological Materials by Reinforcement Fine-Tuned Generative Model

提出一种通过最小化交叠矩阵逆的乘积，寻找多体波函数基态的算法。

Link: * A Unified Variational Framework for Quantum Excited States * Spectral Inference Networks: Unifying Deep and Spectral Learning * The Geometry of Algorithms with Orthogonality Constraints * Accurate computation of quantum excited states with neural networks

参加杨东平老师在之江实验室举办的大语言模型研讨会议，从物理、神经计算、生物等方面理解和看待语言模型以及神经网络的发展。

提出两种多智能体合作方案。

文献： * RECONCILE: Round-Table Conference Improves Reasoning via Consensus among Diverse LLMs * ReConcile code * ReAgent: Reversible Multi-Agent Reasoning for Knowledge-Enhanced Multi-Hop QA * []

多智能体（Multi-Agent System, MAS）合作处理问题的思路十分流行，但是在一些热门的batchmark上并没有明显的表现差距。本文针对MAS没有性能提升的问题进行探究，总结出以下三个方面：

1. specification and system design failures
2. inter-agent misalignment
3. task verification and termination.

文献： * Why Do Multi-Agent LLM Systems Fail? * 源代码GitHub

回顾DeepSeek模型发展过程，从最初的数据训练和模型搭建出发，为了在受限的硬件条件下创在出更加高效的模型，修改模型的架构，最后提出基于强化学习的模型微调方案。

Reference: * DeepSeek LLM Scaling Open-Source Language Models with Longtermism * DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models * DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model * DeepSeek-V3 Technical Report * DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning * DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models * Brief analysis of DeepSeek R1 and its implications for Generative AI * Auxiliary-Loss-Free Load Balancing Strategy for Mixture-of-Experts * Training language models to follow instructions with human feedback

贝叶斯优化（Bayesian Optimization）是一种基于贝叶斯定理:

$$ P(A|B) = \frac{P(B|A) \cdot P(A)}{P(B)} $$

的全局优化方法，通常用于在计算代价高昂的情况下优化黑箱函数。它主要用于高效地寻找目标函数的最优解，尤其在函数不可微、函数形状复杂、或者评估函数代价昂贵（如深度学习模型的超参数优化）时特别有效。

这篇文章研究机器学习过程中，隐藏层对数据的处理方式，发现数据在等几何定律的分离，并且可以观察到类别的出现，因此总结归纳了一个可以量化的规律。

Reference: * A law of data separation in deep learning * github地址

介绍Active Learning的基本概念与算法，以及相关python库——ALiPy的使用。

The key idea behind active learning is that a machine learning algorithm can achieve greater accuracy with fewer labeled training instances if it is allowed to choose the data from which is learns.

Reference: * Active Learning Literature Survey * ALiPy: Active Learning in Python * GitHub:ALiPy