Model Reading Map

AI 논문을 읽을 때는 모델 이름보다 먼저 입력, 표현, 구조, 학습 신호, 평가 claim을 분리합니다. 이 페이지는 한 편의 논문을 wiki note로 옮길 때 쓰는 절차형 지도입니다.

paper claim = object + representation + architecture + objective + evidence

논문을 읽을 때 가장 먼저 해야 할 일은 “새 모델 이름”을 외우는 것이 아니라 claim의 단위를 고정하는 것입니다. 같은 architecture라도 input object, split, objective, evaluation이 달라지면 완전히 다른 주장입니다.

Reading Order

Step	Ask	Route
1. Object	입력 단위가 sample, token, image, graph, molecule, protein, pocket, agent state 중 무엇인가?	Modalities, Entities
2. Representation	raw object가 token, embedding, graph, coordinate, conformer, feature로 어떻게 바뀌는가?	Representation contract
3. Architecture	어떤 축을 섞고 어떤 inductive bias를 쓰는가?	Architectures, Architecture-objective fit
4. Objective	label, mask, contrast, likelihood, denoising, flow, preference, reward 중 무엇을 최적화하는가?	Learning Methods, Objective taxonomy
5. Training	data source, split, optimizer, schedule, selection rule이 claim과 맞는가?	Machine Learning, Systems
6. Evaluation	metric, baseline, uncertainty, leakage check가 주장을 지지하는가?	Evaluation
7. Runtime	inference cost, serving boundary, reproducibility가 실제 사용 조건과 맞는가?	Systems, Infra

논문 note를 작성할 때는 최소한 아래 기호를 먼저 고정합니다.

Symbol	Meaning
$u_{i}$	sample unit: molecule, sequence, image, protein-ligand pair, user task
$x_{i}$	model input after preprocessing
$r_{i}$	representation: token, graph, coordinate, embedding
$y_{i}$	label, target, reference, preference, reward
$f_{θ}$	model or predictor
$L$	training objective
$\overset{m}{^}$	finite-sample metric estimate

Claim	먼저 볼 것	흔한 함정
Better architecture	같은 objective와 data에서 architecture만 바뀌었는가?	objective, augmentation, compute 차이가 architecture 효과처럼 보임
Better pretraining	downstream split과 pretraining corpus가 분리되는가?	duplicate, homolog, scaffold, source leakage
Better generation	sample quality와 likelihood/score/evaluator가 같은 claim인가?	filtered denominator를 숨김
Better representation	frozen probe, kNN, retrieval, fine-tuning 중 무엇으로 검증했는가?	probe budget이 representation 차이를 가림
Better system	latency, memory, throughput, reproducibility를 같은 조건에서 비교했는가?	hardware, precision, batch size 차이

Field	Write
Problem	task unit, input $x$ , target/output $y$
Representation	token/graph/coordinate/embedding contract
Architecture	family, key operation, inductive bias, scaling cost
Objective	exact loss or training signal with symbols defined
Evidence	split, metric, baseline, uncertainty, ablation
Boundary	AI, Math, Computational Biology, Infra, Agents 중 어디로 넘길지

내용	둘 곳
model-internal structure, objective, training/evaluation claim	AI
formula shape, estimator, probability, gradient, geometry primitive	Math
protein, molecule, ligand, pocket, docking, assay, structure workflow	Computational Biology
GPU, Slurm, storage, deployment environment	Infra
tool-using LLM workflow, paper brief automation, wiki maintenance	Agents