Cai-Jianfeng

Cai-Jianfeng

Second-Year Master Student, School of Information Science and Technology, USTC

Sitemap

A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.

Pages

Posts

The Basic Knowledge of Torch Train Pipeline

14 minute read

Published:

Update:

这篇博客主要讲解 PyTorch 训练模型的整个流程的具体细节， 包括如何在前向过程中构建计算图；后向传播过程中如何计算并保存梯度；优化器如何根据梯度更新模型参数。(建议先阅读我之前关于 torch.autograd 的博客 The Basic Knowledge of PyTorch Autograd)

The Basic Knowledge of MoE

109 minute read

Published:

这篇博客主要讲解了使用 Mixture of Experts (MoE) 将多个模型进行组合的原理。

The Basic Knowledge of TorchScript

15 minute read

Published:

这篇博客主要讲解了使用 TorchScript 将 Python 模型代码转化为其他语言代码(如 C++)的原理和具体实现。

The Basic Knowledge of Gradient Penalty

22 minute read

Published:

这篇博客主要讲解了使用梯度惩罚(gradient penalty)作为正则化项来促进模型学习的数学原理和具体实现。

The Basic Knowledge of PyTorch Distributed

49 minute read

Published:

Update:

这篇博客主要介绍了 LLM 分布式并行的训练方式，并着重讲解了 PyTorch 代码的实现 DDP 的方式。

PyTorch 随笔

16 minute read

Published:

torch.backends.cudnn.deterministic: 固定 cuda 的随机种子，使得每次返回的卷积算法都是确定的，即默认算法

The Basic Knowledge of LLM

9 minute read

Published:

这篇博客主要介绍了 Large Language Model 的基础知识，包括常见的 LLM，微调方式等。

The Basic Knowledge of NLP

43 minute read

Published:

这篇博客主要介绍了 NLP 任务中的基础知识，包括性能评价指标(metrics)，分词算法(tokenization)等。

Emu series (Emu & Emu Edit & Emu Video)

49 minute read

Published:

本文主要对近期 Meta 发表的三篇关于视觉处理的文章(Emu 系列)进行论文解读(按照它们的发布顺序)： 首先是 SOTA 的 text-to-image 生成模型 Emu；接着以它为 baseline，进行 image edit 的研究改进，提出了一个大一统的图像编辑模型 Emu Edit， 这基本上就把图像领域主流的任务都刷了个遍。最后又提出了 Emu Video 模型，利用 Emu 完成了对 text-to-video 生成模型的改进，也获得了 SOTA。 (ps：我猜下一步应该就是 video edit 的研究改进了🙂)

The Basic Knowledge of Scored-based Generative Model

89 minute read

Published:

这篇博客参考了 Generative Modeling by Estimating Gradients of the Data Distribution， 详细讲述了最近大火的 Diffusion Model 的另一个理解/推理角度: Score-based Generative Model 的数学原理及编程。 (ps：建议先看完上述的 Generative Modeling by Estimating Gradients of the Data Distribution 博客，虽然是全英文的，但是写的十分详细，且简单易懂，真的非常良心)

笔记本电脑固态硬盘更换扩容

9 minute read

Published:

这篇博客主要讲解只有一个固态硬盘(SSD)槽的笔记本电脑在需要扩容的简便操作(无需重装系统)。

latex 问题合集

8 minute read

Published:

这篇博客主要记录我在使用 latex 的过程中所遇到的问题和解决的方法(注：有些问题可能我自己也不知道原理，但是所有的解决方法都是亲测有效)。

The Basic Knowledge Of Latex

24 minute read

Published:

这篇博客主要记录了 latex 的基本知识和用法，适用于第一次听到 latex 这个写作排版工具的小白。

portfolio

publications

Mitigating Hallucination in VideoLLMs via Temporal-Aware Activation Engineering

in Neural Information Processing Systems, 2025

It is the first to systematically investigate the effectiveness and underlying mechanisms of activation engineering for mitigating hallucinations in VideoLLMs. And it proposes a temporal-aware activation engineering framework for VideoLLMs, which adaptively identifies and manipulates hallucination-sensitive modules based on the temporal variation characteristic, substantially mitigating hallucinations without additional LLM fine-tuning.

Download here

Bias Fitting to Mitigate Length Bias of Reward Model in RLHF

under review in Neural Information Processing Systems, 2025

To accurately model the intricate nature of length bias and facilitate more effective bias mitigation, it proposes FiMi-RM (Bias Fitting to Mitigate Length Bias of Reward Model in RLHF), a framework that autonomously learns and corrects underlying bias patterns.

Download here

Multi-Level Aware Preference Learning: Enhancing RLHF for Complex Multi-Instruction Tasks

under review in Neural Information Processing Systems, 2025

The research identifies a critical oversight in existing techniques, which predominantly focus on comparing responses while neglecting valuable latent signals embedded within prompt inputs, and which only focus on preference disparities at the intra-sample level, while neglecting to account for the inter-sample level preference differentials that exist among preference data. To leverage these previously neglected indicators, it proposes a novel Multi-level Aware Preference Learning (MAPL) framework, capable of enhancing multi-instruction capabilities.

Download here

Disentangling Length Bias In Preference Learning Via Response-Conditioned Modeling

in International Conference on Learning Representations, 2026

It introduces a Response-conditioned Bradley-Terry (Rc-BT) model that enhances the model’s capability in length bias mitigating and length instruction following, through training on the augmented dataset. Furthermore, it proposes the Rc-RM and Rc-DPO algorithm to leverage the Rc-BT model for reward modeling and direct policy optimization (DPO) of LLMs.

Download here

teaching

Study Experience-Undergraduate

Undergraduate, Artifical Intelligence Turing Class, the School of Artifical Intelligence, Xidian University, 2020

I completed my undergraduate studies in Xidian University from 2020 to 2024.