ICLR Topic Selection

Use this when a project is still movable. ICLR is broad, but the paper should teach the learning community something about representations, objectives, models, data, optimization, evaluation, or deployment.

Strong ICLR signals

• A clear representation-learning, model-behavior, optimization, generative modeling, RL, theory, or

evaluation contribution.

• Evidence that changes how researchers should build, analyze, or judge learning systems.
• A simple central claim that can be verified by focused theory, experiments, or artifacts.
• Interest beyond one dataset, product, or application vertical.
• Honest limitations and ethics treatment for high-impact model or data claims.

Weak ICLR signals

• Pure application paper with little learning insight.
• Incremental benchmark bump without mechanism, analysis, or robust evidence.
• Closed system claim that reviewers cannot inspect or reproduce.
• Dataset-only paper without a learning-representation or evaluation advance.
• Theory result disconnected from modern learning practice and not routed to a theory-focused venue.

Routing logic

• Prefer NeurIPS or ICML for broader ML method/theory work with less ICLR-specific representation

framing.

• Prefer AISTATS or UAI for statistics, uncertainty, causal, or probabilistic emphasis.
• Prefer ACL, CVPR, KDD, or robotics/HCI venues when the contribution is primarily domain-specific.
• Prefer workshops when the idea is timely but under-evidenced.

Fit-versus-route decision table

ICLR's center of gravity is deep representation learning: architectures, self-supervision, generative models, foundation models, RL with deep function approximation, optimization for deep nets, interpretability, and alignment. Score the project against that center before routing.

Worked vignette

A team has a method that improves recommendation click-through in production. As written it is an application paper. To make it ICLR-shaped, they extract the representation-learning claim: a new contrastive objective that yields embeddings transferring across catalogs, demonstrated with an ablation and a probe on a public dataset. The product result becomes one validation point, not the contribution. If that reframing fails to surface a learning insight, the honest route is KDD.

Reviewer-pushback patterns

• "No learning insight, just engineering." Reframe around the mechanism or route to a domain venue.
• "Dataset-only paper." Add an evaluation or representation advance, or target a datasets-and-

benchmarks track instead.

• "Theory disconnected from practice." Tie the result to an observed deep-learning phenomenon.

Output format

[ICLR fit] strong / plausible / weak / no
[Core learning insight] <one sentence>
[Evidence required] <theory, experiment, benchmark, artifact>
[Best venue route] ICLR / NeurIPS / ICML / AISTATS / UAI / domain venue / workshop
[Reframe] <how to make the paper more ICLR-shaped>