这里主要记录我在大语言模型算法方向上的学习与研究笔记，重点关注训练目标、推理能力、偏好优化、检索增强以及高效推理等相关主题。
这里整理近期的大模型算法笔记、论文阅读总结，以及围绕具体问题展开的研究思考。
A working note on how DPO, IPO, and related objective variants differ in gradient behavior, implicit reward modeling, and data sensitivity. I focus on what actually changes during training instead of only comparing final benchmark numbers.
This post collects ideas on why extra inference-time computation sometimes helps reasoning, when self-consistency breaks down, and how verifier-guided decoding can improve answer quality without retraining the whole model from scratch.
A practical overview of synthetic data generation for SFT: prompt construction, filtering, diversity control, and the tradeoff between scale and response quality when building instruction datasets for domain adaptation.
Notes on serving long-context LLMs efficiently, including KV cache reuse, chunk scheduling, retrieval granularity, and where the practical bottleneck shifts from model quality to system design.
按主题对内容进行归档，方便后续持续整理，也让这个网站更像一个真实在更新的研究笔记站。
Preference optimization, supervised fine-tuning, reward modeling, and objective design.
Serving efficiency, long-context behavior, memory tradeoffs, batching, and latency analysis.
Verifier models, self-consistency, evaluation protocol design, and failure case analysis.
Retrieval pipelines, reranking, context compression, and grounding strategies for LLMs.
This section can later be connected to a real site search. For now, it highlights the main themes that appear across the notes and articles on this homepage.