Won's Blog

공부 및 실험 공유

A.X K1 Technical Report

A.X K1 논문 리뷰 — 519B MoE 모델의 아키텍처, 데이터 파이프라인, Think-Fusion 학습 전략

9 min read · 2026

TelAgentBench: A Multi-faceted Benchmark for Evaluating LLM-based Agents in Telecommunications

TelAgentBench 논문 리뷰 - 통신 도메인 LLM 에이전트의 5가지 핵심 역량 평가 벤치마크

24 min read · 2026

TelBench: A Benchmark for Evaluating Telco-Specific Large Language Models

TelBench 논문 리뷰 — 통신 도메인 특화 LLM 벤치마크의 설계, 구축, 평가

22 min read · 2026

FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling

FlashAttention-4 논문 리뷰 — Blackwell GPU의 비대칭 스케일링에 맞춘 파이프라인 재설계와 소프트웨어 지수함수

11 min read · 2026

FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

FlashAttention-3 논문 리뷰 — Hopper GPU의 비동기 실행과 FP8을 활용한 Attention 최적화

19 min read · 2026

Triton 07: Flash Attention 3 — Triton으로 어디까지 가능한가

Hopper 전용인 Flash Attention 3를 Triton으로 어디까지 따라잡을 수 있는가 — 확장 autotune·persistent kernel·실패한 실험까지

10 min read · 2026

Triton 06: Flash Attention 2 — FA1 대비 5가지 최적화

Flash Attention 2를 Triton으로 구현한다 — un-scaled 누적, exp2 트릭, Causal 2-stage, tl.dot accumulator, autotune

12 min read · 2026

Triton 05: Flash Attention — 종합 프로젝트

Flash Attention을 Triton으로 구현한다 — Forward/Backward 전체 구현과 RTX 4080·A100·H100·B200 아키텍처별 최적화 포인트

19 min read · 2026

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

FlashAttention-2 논문 리뷰 — non-matmul FLOPs 감소, 병렬화, warp partitioning 개선

12 min read · 2023

SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

SWE-bench 논문 리뷰 — 실 레포의 실 GitHub 이슈로 LLM agent를 평가하는 sandboxed execution benchmark

13 min read · May 25, 2026

2026 · llm agent benchmark evaluation coding paper · paper
TravelPlanner: A Benchmark for Real-World Planning with Language Agents

TravelPlanner 논문 리뷰 — multi-constraint planning에서 GPT-4도 1% 미만, agent planning의 한계 노출

12 min read · May 25, 2026

2026 · llm agent benchmark evaluation planning paper · paper
MedAgentBench: A Realistic Virtual EHR Environment to Benchmark Medical LLM Agents

MedAgentBench 논문 리뷰 — Stanford EHR 데이터 + FHIR 환경에서 의료 LLM agent를 평가하는 도메인 특화 벤치마크

13 min read · May 25, 2026

2026 · llm agent benchmark evaluation medical healthcare paper · paper
OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments

OSWorld 논문 리뷰 — 실제 OS 위에서 마우스·키보드로 작업하는 GUI 에이전트를 execution-based로 평가하는 벤치마크. 인간 72% vs 최고 모델 12%

12 min read · May 25, 2026

2026 · llm agent benchmark evaluation gui computer-use paper · paper
Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations

Red-Teaming 시리즈 #26 (마지막) — Llama-2-7B를 input/output safety classifier로 fine-tune, OpenAI Moderation API를 능가하는 공개 가드레일 (Inan et al., Meta, 2023)

20 min read · May 18, 2026

2026 · llm red-teaming safety paper defense classifier guardrail · paper

Won's Blog

공부 및 실험 공유

A.X K1 Technical Report

TelAgentBench: A Multi-faceted Benchmark for Evaluating LLM-based Agents in Telecommunications

TelBench: A Benchmark for Evaluating Telco-Specific Large Language Models

FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling

FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision

Triton 07: Flash Attention 3 — Triton으로 어디까지 가능한가

Triton 06: Flash Attention 2 — FA1 대비 5가지 최적화

Triton 05: Flash Attention — 종합 프로젝트

FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

TravelPlanner: A Benchmark for Real-World Planning with Language Agents

MedAgentBench: A Realistic Virtual EHR Environment to Benchmark Medical LLM Agents

OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments

Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations