Daily AI Digest — 2026-05-11
Published:
Posts by Collection
daily
Daily AI Digest — 2026-05-12
Published:
Daily AI Digest — 2026-05-13
Published:
Daily AI Digest — 2026-05-14
Published:
Daily AI Digest — 2026-05-15
Published:
Daily AI Digest — 2026-05-16
Published:
Daily AI Digest — 2026-05-17
Published:
Daily AI Digest — 2026-05-18
Published:
Daily AI Digest — 2026-05-19
Published:
Daily AI Digest — 2026-05-20
Published:
Daily AI Digest — 2026-05-21
Published:
Daily AI Digest — 2026-05-22
Published:
Daily AI Digest — 2026-05-23
Published:
Daily AI Digest — 2026-05-24
Published:
Daily AI Digest — 2026-05-25
Published:
Daily AI Digest — 2026-05-26
Published:
Daily AI Digest — 2026-05-27
Published:
Daily AI Digest — 2026-05-28
Published:
Daily AI Digest — 2026-05-29
Published:
Daily AI Digest — 2026-05-30
Published:
Daily AI Digest — 2026-05-31
Published:
Daily AI Digest — 2026-06-01
Published:
Daily AI Digest — 2026-06-02
Published:
Daily AI Digest — 2026-06-03
Published:
Daily AI Digest — 2026-06-04
Published:
Daily AI Digest — 2026-06-05
Published:
Daily AI Digest — 2026-06-06
Published:
Daily AI Digest — 2026-06-07
Published:
Daily AI Digest — 2026-06-08
Published:
Daily AI Digest — 2026-06-09
Published:
Daily AI Digest — 2026-06-10
Published:
Daily AI Digest — 2026-06-11
Published:
Daily AI Digest — 2026-06-12
Published:
Daily AI Digest — 2026-06-13
Published:
Daily AI Digest — 2026-06-14
Published:
Daily AI Digest — 2026-06-15
Published:
Daily AI Digest — 2026-06-16
Published:
Daily AI Digest — 2026-06-17
Published:
Daily AI Digest — 2026-06-18
Published:
Daily AI Digest — 2026-06-19
Published:
Daily AI Digest — 2026-06-20
Published:
Daily AI Digest — 2026-06-21
Published:
Daily AI Digest — 2026-06-22
Published:
Daily AI Digest — 2026-06-23
Published:
llm-libs
PyTorch 深度学习框架
张量操作、autograd、nn.Module、优化器、分布式训练(DDP/FSDP)、混合精度(AMP)、CUDA
NumPy 科学计算库
ndarray、广播机制、矩阵运算、线性代数(SVD/特征值)、FFT、在LLM中的数据预处理
Pandas 数据处理库
DataFrame、数据读写、清洗、groupby聚合、merge/join、训练数据预处理
Pydantic 数据验证库
BaseModel、数据验证、序列化、Field约束、与LLM结构化输出配合
tiktoken 分词库
BPE分词、encoding_for_model、token计数与成本估算、cl100k_base/p50k_base编码
tokenizers 分词库
BPE/WordPiece/Unigram算法、训练自定义分词器、与transformers集成
Transformers 模型库（结合源码）
PreTrainedModel/AutoModel架构、from_pretrained加载流程、generate解码、Trainer训练循环、量化
PEFT 参数高效微调库（结合源码）
LoRA/AdaLora/QLoRA数学原理、LoraLayer注入与合并、get_peft_model
TRL 对齐训练库（结合源码）
SFTTrainer/DPOTrainer/GRPOTrainer、DPO/GRPO/KTO数学原理、对齐训练流程
Accelerate 分布式训练库
Accelerator、DDP/FSDP/DeepSpeed集成、device_map自动分片、混合精度
Datasets 数据集库
load_dataset、map批量处理、流式加载(streaming)、Arrow零拷贝架构
bitsandbytes 量化库
LLM.int8()混合精度分解、NF4量化(QLoRA)、4bit/8bit推理与训练
DeepSpeed 分布式训练框架
ZeRO-1/2/3显存分析、ZeRO-Offload、激活检查点、HF Trainer集成
Megatron-Core 大模型训练框架
张量并行(Column/Row)、流水线并行(1F1B)、序列并行、上下文并行(Ring Attention)
Weights & Biases 实验追踪
wandb.init/log、Sweep超参搜索、Artifact版本管理、HF集成
MLflow ML生命周期管理
Tracking/Models/Registry/Projects、autolog、HuggingFace集成
vLLM 高性能推理引擎
PagedAttention原理、连续批处理、LLM类/SamplingParams、OpenAI兼容服务器
SGLang 结构化生成引擎
RadixAttention基数树KV cache复用、编程原语、约束解码
TensorRT-LLM 推理优化
计算图优化、内核融合、INT8/FP8量化、张量并行/流水线并行
llama.cpp 推理引擎
GGUF格式、Q4_0~Q6_K量化、CPU/GPU推理、mmap内存映射
llama-cpp-python Python绑定
Llama类、create_completion/chat、GBNF语法约束、嵌入提取、OpenAI服务器
FlashAttention 高效注意力
IO-awareness原理、在线Softmax、O(N²)→O(N)内存、分块计算推导
xFormers Transformer组件库
memory_efficient_attention、SwiGLU、RoPE旋转位置编码、稀疏注意力
FastAPI 异步Web框架
路由/Pydantic集成、SSE流式响应、中间件、LLM推理API服务
Gradio 交互式UI框架
Interface/Blocks、Chatbot组件、流式输出、HF Spaces部署
LangGraph Agent编排框架
StateGraph、节点/边/条件路由、持久化、Human-in-the-loop、多Agent
LlamaIndex RAG框架
Index/Retriever/QueryEngine、RAG管道抽象、Document/Node、Agent工具调用
DSPy 声明式Prompt编程
Signature、Module(Predict/ChainOfThought/ReAct)、Optimizer自动优化
AutoGen 多Agent对话框架
ConversableAgent、GroupChat、代码执行(本地/Docker)、v0.4 AgentChat
CrewAI 角色驱动协作
Agent/Task/Crew、Sequential/Hierarchical流程、Memory记忆系统
instructor 结构化输出
response_model、流式模式、Mode(TOOLS/JSON/MD_JSON)、Pydantic验证
Outlines 约束解码
generate.text/choice/regex/json/cfg、FSM驱动采样、logit偏置原理
smolagents 轻量Agent框架
CodeAgent/ToolCallingAgent、@tool装饰器、HfApiModel、代码沙箱
Ragas RAG评估框架
evaluate()、Faithfulness/AnswerRelevancy/ContextPrecision等指标、TestsetGenerator
DeepEval 评估测试框架
LLMTestCase、Hallucination/Bias/Toxicity等指标、GEval通用评估
LangSmith SDK 可观测性
@traceable追踪、evaluate()评估、数据集管理、trace→span层级
Arize Phoenix 开源可观测性
OpenTelemetry追踪、嵌入UMAP可视化、评估器、本地优先架构
Chroma 轻量向量数据库
PersistentClient、Collection CRUD、嵌入函数、元数据过滤、RAG存储
Qdrant 高性能向量数据库
HNSW索引、Filter过滤系统、Payload索引、量化、FastEmbed集成
Milvus 分布式向量数据库
IVF/PQ/HNSW索引、hybrid_search、存算分离架构、MilvusClient
pgvector PostgreSQL向量扩展
vector/halfvec/sparsevec类型、HNSW/IVFFlat索引、SQL距离操作符
portfolio
ARH — AI Research Helper Permalink
Open-source CLI-first research assistant agent for AI researchers. Tool use, three-layer memory, plan–execute safety, skill self-learning, multi-LLM fallback.
publications
FinVault: Benchmarking Financial Agent Safety in Execution-Grounded Environments
Published in Submitted to EMNLP 2026, 2026
The first execution-grounded security benchmark for LLM-based financial agents — 31 regulatory sandbox scenarios, 107 real-world vulnerabilities, 963 test cases. Submitted to EMNLP 2026.
Recommended citation: Zhi Yang, Runguo Li, Qiqi Qiang, Jiashun Wang, Fangqi Lou, et al. (2026). "FinVault: Benchmarking Financial Agent Safety in Execution-Grounded Environments." arXiv preprint arXiv:2601.07853. Submitted to EMNLP 2026.
Download Paper