AI Paper Claim Patterns

Use this page when a new AI paper looks important but the main contribution is unclear. A paper should be routed by the strongest claim, not by the most recognizable model name.

The minimum AI claim is:

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

If the paper also depends on molecules, proteins, structures, or formulas, route it through AI Computational Biology Math contract after choosing the AI claim pattern.

Pattern Map

Pattern	Main Claim	Minimum Evidence	Route
Architecture claim	a model structure improves quality, cost, or scaling	matched training setup, baseline architecture, ablation, complexity	Architectures
Learning-method claim	a supervision or pretraining signal improves representation or transfer	same architecture where possible, signal definition, transfer protocol	Learning methods
Generative-model claim	a distribution model or sampler improves samples	objective, sampler, validity/diversity/utility metrics	Generative models
Evaluation claim	a benchmark, metric, split, or protocol changes what can be trusted	dataset card, split rule, metric definition, leakage check	Evaluation
Scaling claim	more data, parameters, compute, memory, or inference budget changes performance	scaling variable, controlled baseline, cost boundary	Scaling claim contract
Systems claim	implementation, serving, reproducibility, or artifacts make a method practical	runtime evidence, artifact status, reproducibility contract	Systems, Infra
Agent claim	tool use, planning, memory, or verification improves task completion	task suite, tool boundary, verifier, failure definition	Agents

Architecture Claim

Architecture papers change the function family:

$$f_{\theta }\in {\mathcal{F}}_{\mathrm{new}}\text{vs}{f}_{\varphi }\in {\mathcal{F}}_{\mathrm{base}}$$

The claim is not just “new block works.” It should identify what bias changed.

Required	Why
Input structure	sequence, graph, set, image, coordinate, and multimodal inputs need different biases
Complexity	quality claims often trade off with memory, latency, or sequence length
Ablation	the new block, connection, normalization, or routing decision must be isolated
Matched objective	a different loss can masquerade as an architecture improvement
Baseline family	compare against the right CNN, RNN, Transformer, GNN, SSM, MoE, or geometric model

Useful links:

• Architecture selection
• Computational complexity
• Scaling claim contract

Learning-Method Claim

Learning-method papers change the training signal:

$$\mathcal{J}(\theta )={\mathbb{E}}_{u\sim q(u)}\left[{\ell }_{\theta }^{\mathrm{signal}}(u)\right]$$

The paper should say whether the signal is label, mask, contrast, denoising target, preference, reward, pseudo-label, or synthetic target.

Required	Why
Signal source	labels, masks, augmentations, preferences, and rewards carry different noise
Representation unit	token, sequence, graph, image patch, residue, molecule, or trajectory
Adaptation protocol	frozen probe, fine-tuning, retrieval, and full retraining test different claims
Transfer split	transfer evidence needs target-domain validation, not only pretraining loss
Objective-metric alignment	pretraining objective may not optimize downstream utility

Useful links:

• Learning methods
• Representation evaluation
• Objective-metric alignment

Generative-Model Claim

Generative papers change a distribution, path, or sampler:

$$x\sim p_{\theta }(x\mid c),\text{sample quality}\ne \text{training loss}$$

The important distinction is learned quantity: likelihood, latent variable, score, noise, velocity, discriminator signal, or consistency map.

Required	Why
Learned quantity	score, noise, velocity, likelihood, and reward define different methods
Sampling procedure	solver, steps, guidance, filtering, and rejection can own performance
Validity rule	syntactic validity, physical validity, and task utility differ
Diversity/novelty	high average score can hide mode collapse or near-duplicates
Utility metric	final task success may come from a downstream evaluator or filter

Useful links:

• Generative models
• Formula pattern catalog
• Generation evaluation

Evaluation Claim

Evaluation papers change what evidence means:

$$\hat{M}=\frac{1}{m}\sum _{j=1}^{m}M({\hat{y}}_j,y_j)$$

This only supports a claim after the split, model-selection rule, denominator, and uncertainty are known.

Required	Why
Example unit	examples, prompts, molecules, proteins, targets, and tasks aggregate differently
Split unit	random split, scaffold split, family split, temporal split, and source split test different generalization
Selection rule	checkpoint, threshold, prompt, sampler, and filter selection can overfit validation
Baseline	a weak or mismatched baseline inflates the contribution
Uncertainty	seed variance, confidence interval, and paired comparison determine reliability

Useful links:

• Evaluation
• Benchmark claim contract
• Evaluation math

Scaling Claim

Scaling papers make performance a function of resource variables:

$$Q=F(D,N,C,B,L)$$

• $Q$: quality, loss, accuracy, utility, or reward.
• $D$: data size or data quality.
• $N$: parameter count or active parameters.
• $C$: training compute.
• $B$: inference budget, samples, search width, or tool calls.
• $L$: latency, context length, sequence length, or memory budget.

Checks:

• Which variable changed and which variables were controlled?
• Is quality normalized against compute, memory, latency, or cost?
• Does the reported scaling transfer to the target domain or only the benchmark?
• Are inference-time budgets included when comparing methods?

Systems and Artifact Claim

Systems claims are about making a method usable, reproducible, or cheaper:

$$\text{method value}=\text{quality}\times \text{reliability}\times \text{deployability}$$

Minimum evidence:

Required	Why
Artifact availability	code, data, splits, configs, weights, and environment must be public or marked missing
Runtime boundary	throughput, latency, memory, batch size, hardware class, and precision matter
Reproducibility	setup steps and deterministic enough evaluation are part of the claim
Failure mode	practical methods need explicit error behavior, not only average speed

Agent Claim

Agent papers should not be treated as ordinary model papers if the contribution depends on tools, memory, planning, or verification:

$$\text{agent}=\text{model}+\text{state}+\text{tools}+\text{policy}+\text{verifier}$$

Minimum evidence:

Required	Why
Task suite	agents can overfit narrow demonstrations
Tool contract	tool permissions, side effects, and outputs define the action space
Memory boundary	context, scratchpad, retrieved memory, and persisted memory differ
Verification loop	success must be checked by evidence, tests, or human review
Failure taxonomy	tool misuse, hallucination, prompt injection, and partial completion need separate handling

Final Routing Checklist

• What is the strongest claim: architecture, learning signal, generation, evaluation, scaling, systems, or agent workflow?
• Is the input/output object clear before naming the model?
• Is the objective separate from the final metric?
• Is the evaluation protocol strong enough for the headline claim?
• Does the paper require a domain-specific contract such as Computational Biology?
• Does the paper require a formula-pattern rewrite before summary?
• Are missing metadata, artifacts, or metrics marked to verify?

Related

• AI paper intake
• Claim routing
• Formula pattern catalog
• Coverage matrix
• AI Computational Biology Math contract