AI Paper Intake

AI 논문을 읽을 때는 모델 이름보다 claim의 위치를 먼저 고정합니다. 같은 Transformer 논문이라도 architecture 논문, learning objective 논문, generative model 논문, evaluation 논문은 서로 다른 축으로 읽어야 합니다.

$$\text{paper claim}=(\text{input},\text{output},\text{model},\text{objective},\text{evidence})$$

Intake Fields

Field	Question	Route
Input and output	What object enters the model, and what object is predicted?	Modalities, Tasks
Claim pattern	Is the strongest claim architecture, learning method, generation, evaluation, scaling, systems, or agent workflow?	AI paper claim patterns
Prediction type	Is it classification, regression, ranking, retrieval, generation, or action selection?	Machine Learning
Architecture	What inductive bias, parameter sharing, and complexity does the model use?	Architectures
Learning signal	Is the signal label, mask, contrast, denoising target, preference, reward, or synthetic target?	Learning Methods
Distribution model	Does the method model $p_{\theta }(x)$, $p_{\theta }(y\mid x)$, a score, or a vector field?	Generative Models, Math
Energy or constraint	Does the method use an energy, force, constraint, projection, repair, or filter?	Energy-based model, Constrained optimization
Evidence	Which metric, split, baseline, ablation, and uncertainty support the claim?	Evaluation
Benchmark claim	What data, task, split, metric, allowed information, and reporting rule define the score?	Benchmark intake
System boundary	Does the contribution depend on data scale, compute, serving, tools, or reproducibility?	Systems, Infra, Agents
Scaling claim	What quality, data, model size, compute, and runtime variables changed?	Scaling claim contract

Claim Shapes

Claim Shape	Minimum Evidence
Better architecture	matched objective, matched training budget, baseline architecture, ablation, complexity or throughput note
Better learning method	same architecture where possible, pretraining data description, downstream transfer protocol, frozen or fine-tuned evaluation
Better generative model	likelihood or surrogate objective, sampling procedure, validity/diversity/novelty/utility metrics, failure examples
Better energy or score model	energy/score target, sampler or negative process, validity and calibration boundary
Better constrained model	hard/soft constraint, projection/repair/filter rule, invalid-output accounting
Better benchmark result	exact split, model-selection rule, primary metric, uncertainty or seed variation, leakage check
Better agent workflow	task suite, tool boundary, verifier, success/failure definition, human review boundary

Formula Checks

When a paper introduces an objective, rewrite it into these parts:

$$\hat{\theta }=\arg \underset{\theta }{\min }{\mathbb{E}}_{(x,y)\sim p_{\mathrm{train}}}\left[{\mathcal{L}}_{\theta }(x,y)\right]$$

• Distribution: what does the expectation sample from?
• Target: what is observed, masked, generated, preferred, or rewarded?
• Constraint: what is hard, penalized, projected, repaired, or filtered?
• Parameter: which module is optimized?
• Decision: how does the trained score or probability become an output?
• Evaluation: is the final metric the same object as the training loss?

Update Targets

After reading, update the smallest durable note that captures the reusable idea.

• Architecture idea: Architectures or Architecture concepts
• Learning signal: Learning Methods or Learning concepts
• Generative objective: Generative Models or Generative model concepts
• Energy or constraint objective: Energy-based model or Constrained optimization
• Evaluation issue: Evaluation or Evaluation concepts
• Benchmark issue: Benchmark intake or Benchmark card
• Paper-specific claim: Claim extraction and Evidence table
• Scaling or efficiency claim: Scaling claim contract
• Multi-axis paper: Claim routing
• Repeated AI paper type: AI paper claim patterns

Related

• AI
• Paper review workflow
• Claim routing
• AI paper claim patterns
• Formula intake
• Computational Biology paper intake