LLMs, Bayesian optimization, uncertainty-aware ML, probabilistic search, and related ML-based decision tools.LLM, 베이지안 최적화, 불확실성 인식 머신러닝, 확률적 탐색, 관련 ML 기반 의사결정 도구.
Diffusion models need the marginal score to reverse a noising process. This note explains why a tractable conditional Gaussian target learns that score and in what sense the two losses are equivalent.Diffusion model이 noising process를 역전할 때 필요한 marginal score를 조건부 Gaussian target으로 학습할 수 있는 이유와 두 손실함수의 정확한 등가성을 정리한다.
Stanford CME296 Diffusion & Large Vision Models | Spring 2026 | Lecture 2 - Score matchingStanford CME296 Diffusion & Large Vision Models | Spring 2026 | Lecture 2 - Score matching
A critical note on ChatP&ID: P&IDs are better treated as structured engineering knowledge graphs than as raw images or raw XML, but the benchmark mainly validates context engineering rather than a new GraphRAG algorithm.ChatP&ID에 대한 비판적 노트. P&ID는 이미지나 raw XML보다 구조화된 engineering knowledge graph로 다루는 편이 설득력 있지만, 이 논문의 핵심은 새로운 GraphRAG algorithm보다 context engineering과 retrieval 구조 설계에 가깝다.
GraphRAG for Engineering Diagrams: ChatP&ID Enables LLM Interaction with P&IDsGraphRAG for Engineering Diagrams: ChatP&ID Enables LLM Interaction with P&IDs
A critical note on Tolerance Ball acquisition: a clean probability-of-feasibility objective for specification-driven inverse design, but not a direct optimizer of diversity, boundary coverage, or global feasible-set recovery.Specification-driven inverse design에서 Tolerance Ball acquisition을 비판적으로 읽는다. TB는 valid hit probability에는 잘 정렬되어 있지만 diversity, boundary coverage, global feasible-set recovery를 직접 최적화하지는 않는다.
Range-aware Bayesian optimization for discovering diverse designs within target property windowsRange-aware Bayesian optimization for discovering diverse designs within target property windows
A critical note on a BOHB-MILP framework for international liquid hydrogen supply-chain design under hourly renewable variability, weekly shipping, lead time, and sampled demand-weather scenarios.시간별 재생에너지 변동, 주별 선박운항, 운송 lead time, 수요-기상 시나리오를 결합한 국제 액화수소 공급망 BOHB-MILP 프레임워크에 대한 비판적 노트.
Techno-economic analysis for design and management of international green hydrogen supply chain under uncertainty: An integrated temporal planning approachTechno-economic analysis for design and management of international green hydrogen supply chain under uncertainty: An integrated temporal planning approach
A critical note on Q/K/V projection sharing in Transformer attention, where K=V preserves query-key directionality while cutting KV-cache memory in half.Transformer attention에서 Q/K/V projection을 모두 독립적으로 둘 필요가 있는지 검토하고, K=V sharing이 query-key 방향성을 유지하면서 KV cache를 절반으로 줄이는 압축 축임을 비판적으로 정리한다.
Do Transformers Need Three Projections? Systematic Study of QKV VariantsDo Transformers Need Three Projections? Systematic Study of QKV Variants
LoRA reduces fine-tuning cost by learning low-rank task-specific corrections on top of frozen pretrained weights, while QLoRA adds 4-bit quantization so larger base models can be adapted under tighter GPU memory constraints.LoRA는 고정된 pretrained weight 위에 저랭크 task-specific correction만 학습해 fine-tuning 비용을 줄이고, QLoRA는 여기에 4-bit quantization을 결합해 더 큰 base model을 제한된 GPU memory에서 적응시킬 수 있게 한다.
FlashAttention is not an approximation to attention. Its core idea is to avoid materializing the N by N attention matrix in HBM by computing tiled attention in SRAM and maintaining online softmax statistics.FlashAttention은 어텐션의 근사가 아니다. 핵심은 SRAM에서 타일 단위 어텐션을 계산하고 online softmax 통계를 유지함으로써 HBM에 N by N 어텐션 행렬을 만들지 않는 것이다.