📖 Qiyao Wang's Blog
Total Blogs
2026
-
BEACON：基于里程碑锚定的长程语言 Agent 策略学习方法
-
DV-World：面向真实场景的全面数据可视化 Agent 评测基准
-
OpenSeeker-v2：以高质量高难度轨迹推动搜索Agent极限 —— 仅用SFT超越工业级CPT+SFT+RL训练管线
-
RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning
-
Turn-PPO：基于 Turn 级别优势估计的 PPO 改进 —— 面向多轮 Agent LLM 训练的稳定强化学习方法
-
Workspace-Bench：面向大规模文件依赖的AI Agent工作空间学习评测基准
-
LRAT：从Agent轨迹中学习检索 —— 迈向Agentic Search时代的检索训练新范式
-
OpenSeeker：完全开源训练数据，学术团队实现前沿级搜索 Agent
-
Claw-Eval：面向可信自主 Agent 评估的全轨迹审计与多维评分基准
-
ClawGym: Part II - ClawGym-Agents and Bench
-
GDPval: Measuring AI's Economic Impact
-
ClawGym: Part I - ClawGym-SynData
-
GRPO: Group Relative Policy Optimization
-
Relearning PPO
-
RL in LLM: An Introduction
-
Prompt-OIRL: Query-Dependent Prompt Optimization with Offline Inverse RL
-
CS336 Assignment 1: Transformers Language Model Architecture
2025
-
CS336 Assignment 1: BPE Tokenizer's Detailed Implementation
-
Three Sampling Methods: Temperature, Top K and Top P
-
Greedy Search and Beam Search
-
Autoregressive Decoding: Basic Manner of Decoder-Only LLMs
-
Proximal Policy Optimization (PPO) and RLHF
-
Basic Knowledge of Reinforcement Learning before PPO
-
CMU DLSys Course Homework 1: Implementation and Reflection (Part1)
-
Chain-of-Thought Reasoning without Prompting
-
CMU DLSys Course Homework 0: Implementation and Reflection