论文解读
漫谈 NLP 精选
来自公众号「漫谈NLP」的 LLM 论文解读，点击卡片前往公众号原文。
fai-bench：LLM的Infra工程能力究竟如何？
The Infra layer is critical to AI's self-evolution, since better infrastructure means more efficient training and inference. Φ-Bench/fai-bench aims to measure LLMs' Infra engineering capability.
阅读全文 ↗（跳转微信公众号）2026年7月28日· 论文速览用下采样来加速GRPO
Paper: Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning
阅读全文 ↗（跳转微信公众号）2026年6月25日· 论文速览从LLM中抽取它看过的文章：记录一次简单的实验
LLMs are pretrained on internet-scale text data. Do they store an entire article — or even a whole book — in their parameters?
阅读全文 ↗（跳转微信公众号）2026年4月8日· 论文速览简单的CoT更能Scale，提示词工程也有“Bitter Lesson”
Prompt engineering can improve task performance, but not necessarily in test-time scaling: plain CoT (Chain-of-Thought) actually scales better.
阅读全文 ↗（跳转微信公众号）2026年3月13日· 论文速览从RL的视角理解SFT
SFT and RL are both common post-training techniques, and interestingly they are closely related: from the gradient perspective, SFT performs a special kind of policy gradient; from the loss perspective, the SFT loss is a lower bound of the RL loss.
阅读全文 ↗（跳转微信公众号）2025年12月24日· 论文速览CSP：LLM可以在协作中学会诚实
The COLM 2025 paper 'Don't lie to your friends: Learning what you know from collaborative self-play' proposes Collaborative Self-Play (CSP): build a multi-agent setup and apply Reinforced Self-Training on output rewards, and the LLM becomes more honest — knowing when it can trust its parametric knowledge and when it should search.
阅读全文 ↗（跳转微信公众号）2025年6月25日· 论文速览Agent基础篇：Tool-Use的定义、实现方式和效果优化
By 2025, LLMs keep getting smarter, and expectations have shifted from chatbots to agents that can actually get work done. Tool-use (also known as function calling or tool calling) is a crucial agent concept. This post introduces the fundamentals of tool-use to help readers build a fairly complete picture of it.
阅读全文 ↗（跳转微信公众号）2025年4月11日· 论文速览用AI把诗经、唐诗/宋词三百首译成了现代文
I once happened to try DeepSeek-R1 for translating classical Chinese poetry into modern prose, and the results were surprisingly good: fluent yet literary translations that made the poems easier to grasp while keeping the pleasure of reading. That gave me the idea to batch-translate classical poetry with AI.
阅读全文 ↗（跳转微信公众号）2025年3月27日· 论文速览漫谈Text-to-SQL：技术框架、论文细节以及落地讨论
In the past, a company's business data all landed in databases, where SQL engineers and data analysts wrote queries by hand. Now you can hand the simple tasks to a Text-to-SQL Agent.
阅读全文 ↗（跳转微信公众号）2025年2月17日· 论文速览Agent基础篇：Reflection的介绍、适用场景和应用要点
Reflection is a common and practical agent design pattern, and one of the four agent design paradigms championed by Andrew Ng. This post covers: what Reflection is, where it applies, and the key points of using it well.
阅读全文 ↗（跳转微信公众号）2025年1月14日· 论文速览自动写提示词：DSPy.MIPROv2的介绍与实践（附代码）
I previously shared papers on automatic prompt writing; this time I put it into practice, using DSPy.MIPROv2 to write prompts automatically.
阅读全文 ↗（跳转微信公众号）2024年12月9日· 论文速览怎么说 ≠ 怎么想？LLM推理中的Faithfulness问题 ｜ 大模型推理之旅（3）
LLM reasoning generally follows this form: question -> reasoning process (rationale) + answer. A good reasoning model must be not only smart but also faithful: what it says equals what it thinks, and the rationale fully reflects how the model arrived at the answer — so people can trust the chain from reasoning to answer. This property makes models easier to debug and improve, and increases the credibility of AI decisions in high-stakes tasks such as healthcare and law.
阅读全文 ↗（跳转微信公众号）2024年11月12日· 论文速览用思维树框架ToT，解决复杂推理问题 ｜ 大模型推理之旅（2）
When it comes to LLM reasoning methods, the first that comes to mind is Chain of Thoughts[1] (CoT for short): the model reasons step by step, then gives its answer.
阅读全文 ↗（跳转微信公众号）2024年9月29日· 论文速览初探base LLM的推理能力｜ 大模型推理之旅（1）
I believe reasoning ability is the key to AI generalization. Take humans: rote memorization isn't real intelligence — being able to draw inferences beyond what you were taught is. That's why I began this journey into LLM reasoning. The mountain road is rugged; readers are welcome to explore it with me.
阅读全文 ↗（跳转微信公众号）2024年8月15日· 论文速览Structured Generation（3）：如何让大模型100%输出符合json schema的结果
Structured generation is a new way of using LLMs.
阅读全文 ↗（跳转微信公众号）2024年7月25日· 论文速览Structured Generation with LLMs(2)：Function Calling，不止于agent
Structured generation is a new way of using LLMs.
阅读全文 ↗（跳转微信公众号）2024年7月18日· 论文速览Structured Generation with LLM（1）：介绍Kor，并用免费的LLM API做点练习
Structured Generation with LLMs means having the LLM output structured results that conform to a predefined schema.
阅读全文 ↗（跳转微信公众号）2024年6月9日· 论文速览作为标注员的LLM（六）：用prompt engineering，解锁LLM的专业领域能力
It has been a while since the last entry in the 'LLM as Annotator' series. This post introduces a prompt engineering method that improves LLM annotation quality in specialized domains, without relying on much expert labor.
阅读全文 ↗（跳转微信公众号）2024年5月22日· 论文速览用Coze做一个基于知识库的QA bot，TA熟读了我所有的文章
In this post, we continue building AI applications hands-on.
阅读全文 ↗（跳转微信公众号）2024年4月25日· 论文速览微软的SLMs之路：对Phi-3的看法&使用体验
Microsoft recently released the Phi-3 paper and open-sourced Phi-3-mini (3.8B parameters), which comes in 4K-context[2] and 128K-context versions.
阅读全文 ↗（跳转微信公众号）2024年4月12日· 论文速览RAD-Phi2：在垂直领域微调SLMs，来自微软的示范
In the previous post I introduced Microsoft's SLMs, the Phi series. This time I look at RAD-Phi2, a paper on fine-tuning Phi-2 in radiology. It can be seen as Microsoft's demonstration of fine-tuning SLMs for a vertical domain, and offers useful inspiration for readers with similar needs.
阅读全文 ↗（跳转微信公众号）2024年4月2日· 论文速览微软的SLMs之路：从phi-1到phi-2
The phi series are Microsoft's Small Language Models (SLMs). Three versions have been open-sourced so far: phi-1 (1.3B), phi-1.5 (1.3B), and phi-2 (2.7B) — none of them has gone through alignment.
阅读全文 ↗（跳转微信公众号）2024年3月20日· 论文速览DITTO：角色扮演的self-alignment方法
To scale training data, many role-play models rely on ChatGPT-like models to synthesize data. But this approach carries hallucination risks.
阅读全文 ↗（跳转微信公众号）2024年3月5日· 论文速览Large-Scale Weakly Supervision (Noisy) Data，何以成为大模型的数据基石？
Adding more diverse data during training often yields more robust and generalizable models, and at web scale a foundation model can emerge.
阅读全文 ↗（跳转微信公众号）2024年2月19日· 论文速览LabelFast v0.2版更新：加入SeqGPT模型、新增NER任务
In the previous post I released LabelFast, an open-source NLP annotation tool for the Chinese world. LabelFast's core idea: annotate the easy samples at top speed, so humans can focus on the critical hard ones.
阅读全文 ↗（跳转微信公众号）2024年1月25日· 论文速览用GLM4做一款文字冒险游戏：「带娃路上的打怪升级」
People often say raising a child is a level-grinding game. Inspired by a GPTs competition, I set out to use GLM4 — Zhipu's strongest LLM — to build a parenting-themed text adventure game called 'Leveling Up on the Parenting Road', hoping the game format helps people learn parenting knowledge.
阅读全文 ↗（跳转微信公众号）2023年12月21日· 论文速览自动Prompt Engineering（2）：PE2、LongPO以及BPO
Not long ago I shared three pieces of work on automatic prompt engineering: APE, APO, and OPRO. More than two months have passed and new work has emerged, so this post gives the field an update.
阅读全文 ↗（跳转微信公众号）2023年12月15日· 论文速览Embedding技术的两个小应用：鲁迅全集检索 & 新闻早报聚类
In the previous post I introduced BGE, a general-purpose embedding model for Chinese. This post covers common applications of embedding technology.
阅读全文 ↗（跳转微信公众号）2023年11月30日· 论文速览BGE论文解读：如何炼成中文世界的通用Embedding Model
Foundation models have two representatives: the Large Language Model and the Embedding Model.
阅读全文 ↗（跳转微信公众号）2023年11月22日· 论文速览角色扮演论文浅读：RoleLLM、Character-LLM以及ChatHaruhi
I recently read three role-play papers in one go: RoleLLM, Character-LLM, and ChatHaruhi. This post summarizes the three, mainly covering their technical approaches, dialogue data synthesis methods, and evaluation methods.
阅读全文 ↗（跳转微信公众号）2023年11月8日· 论文速览用AI来帮助AI：Detect Label Error with Cross Validation
In deploying NLP algorithms, engineers come to deeply appreciate this saying: garbage in, garbage out. Data quality matters enormously to model performance.
阅读全文 ↗（跳转微信公众号）2023年11月3日· 论文速览LabelFast：基于LLM的NLP任务自动标注开源工具，Demo发布「AI小作坊」
Introducing LabelFast, an open-source tool that uses LLM techniques to identify and quickly annotate easy text data. With LabelFast, human annotators only need to focus on the small but critical set of hard samples, cutting costs and improving efficiency.
阅读全文 ↗（跳转微信公众号）2023年10月25日· 论文速览AttrPrompt：让LLM生成的数据更加多样、更少Bias的一种思路
Getting an LLM to generate data isn't hard — generating high-quality data is. What does high-quality data mean?
阅读全文 ↗（跳转微信公众号）2023年9月27日· 论文速览自动优化Prompt：Automatic Prompt Engineering的3种方法
Remember 'Let's think step by step' from Zero-Shot CoT? Google recently found a better one via OPRO: 'Take a deep breath and work on this problem step-by-step' — lifting GSM8K straight from 71.8% to 80.2%.
阅读全文 ↗（跳转微信公众号）2023年9月18日· 论文速览无需标注即可Prompt Engineering：3种评估Prompt效果的无监督方法
The prompt engineering process is the process of choosing prompts that perform better, and its key lies in measuring prompt effectiveness.
阅读全文 ↗（跳转微信公众号）2023年9月11日· 论文速览Targeted Data Generation：半自动化的数据迭代框架，NLP模型迭代的新思路（ACL 2023）
Do you know where your model's weaknesses are? Do you know how to improve them?
阅读全文 ↗（跳转微信公众号）2023年9月4日· 论文速览作为标注员的LLM（四）：开源大语言模型标注效果优化之Bias Calibration篇
In the third post of the series, I concluded that on the TNEWS dataset, ChatGLM2-6B and Qwen-7B-Chat annotate worse than zero-shot classification models — Siamese-UniNLU and paddleNLP.
阅读全文 ↗（跳转微信公众号）2023年8月23日· 论文速览作为标注员的LLM（三）：中文开源LLM的标注效果初体验
In the first two posts of the 'LLM as Annotator' series, I introduced two ways of bringing LLMs into annotation: the distillation method, and multi-perspective annotation + weak supervision.
阅读全文 ↗（跳转微信公众号）2023年8月11日· 论文速览ACL 2023 ｜ 对验证集的一场重新审视 - 弱监督学习比你想象的更“弱”
Weak supervision set out to reduce reliance on labeled samples, but an award-winning ACL 2023 paper points out that weak supervision depends too heavily on clean validation sets — contrary to its original intent — and uses validation sets inefficiently, falling behind fine-tuning. The warning: earlier weak-supervision research has drifted from reality and needs timely correction.
阅读全文 ↗（跳转微信公众号）2023年8月7日· 论文速览作为标注员的LLM（二）：在弱监督学习的视角下
In the first post of the 'LLM as Annotator' series, I explained the simplest way to bring LLMs into an annotation system: distillation.
阅读全文 ↗（跳转微信公众号）2023年8月2日· 论文速览作为标注员的LLM（一）-- 蒸馏法
The LLM era has arrived. How will NLP develop? This is a question worth studying.
阅读全文 ↗（跳转微信公众号）