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.
| Project shape |
ICLR fit |
Better route if not ICLR |
| New self-supervised objective with analysis |
Strong |
— |
| Theory explaining a deep-net phenomenon |
Strong |
AISTATS/UAI if purely statistical |
| LLM/foundation-model behavior study |
Strong |
ACL if narrowly language-specific |
| Benchmark bump, no mechanism |
Weak |
Domain venue or workshop |
| Causal/uncertainty emphasis |
Plausible |
AISTATS or UAI |
| Deployed application, little learning insight |
Weak |
KDD, CVPR, robotics/HCI venue |
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>
Source: brycewang-stanford/Awesome-Journal-Skills → ICLR-Skills/skills/iclr-topic-selection/SKILL.md
1---2name: iclr-topic-selection3description: Use when deciding whether a project is a strong ICLR submission, should be reframed for ICLR, or should be routed to NeurIPS, ICML, AAAI, AISTATS, ACL, CVPR, KDD, or another venue. Use when a project lacks a clear representation-learning insight, when an application result needs a learning contribution to fit ICLR, or when weighing ICLR's deep-learning center of gravity against a better-matched venue.4---567# ICLR Topic Selection89Use this when a project is still movable. ICLR is broad, but the paper should teach the learning10community something about representations, objectives, models, data, optimization, evaluation, or11deployment.1213## Strong ICLR signals1415- A clear representation-learning, model-behavior, optimization, generative modeling, RL, theory, or16 evaluation contribution.17- Evidence that changes how researchers should build, analyze, or judge learning systems.18- A simple central claim that can be verified by focused theory, experiments, or artifacts.19- Interest beyond one dataset, product, or application vertical.20- Honest limitations and ethics treatment for high-impact model or data claims.2122## Weak ICLR signals2324- Pure application paper with little learning insight.25- Incremental benchmark bump without mechanism, analysis, or robust evidence.26- Closed system claim that reviewers cannot inspect or reproduce.27- Dataset-only paper without a learning-representation or evaluation advance.28- Theory result disconnected from modern learning practice and not routed to a theory-focused venue.2930## Routing logic3132- Prefer NeurIPS or ICML for broader ML method/theory work with less ICLR-specific representation33 framing.34- Prefer AISTATS or UAI for statistics, uncertainty, causal, or probabilistic emphasis.35- Prefer ACL, CVPR, KDD, or robotics/HCI venues when the contribution is primarily domain-specific.36- Prefer workshops when the idea is timely but under-evidenced.3738## Fit-versus-route decision table3940ICLR's center of gravity is deep representation learning: architectures, self-supervision,41generative models, foundation models, RL with deep function approximation, optimization for deep42nets, interpretability, and alignment. Score the project against that center before routing.4344| Project shape | ICLR fit | Better route if not ICLR |45| --- | --- | --- |46| New self-supervised objective with analysis | Strong | — |47| Theory explaining a deep-net phenomenon | Strong | AISTATS/UAI if purely statistical |48| LLM/foundation-model behavior study | Strong | ACL if narrowly language-specific |49| Benchmark bump, no mechanism | Weak | Domain venue or workshop |50| Causal/uncertainty emphasis | Plausible | AISTATS or UAI |51| Deployed application, little learning insight | Weak | KDD, CVPR, robotics/HCI venue |5253## Worked vignette5455A team has a method that improves recommendation click-through in production. As written it is an56application paper. To make it ICLR-shaped, they extract the representation-learning claim: a new57contrastive objective that yields embeddings transferring across catalogs, demonstrated with an58ablation and a probe on a public dataset. The product result becomes one validation point, not the59contribution. If that reframing fails to surface a learning insight, the honest route is KDD.6061## Reviewer-pushback patterns6263- "No learning insight, just engineering." Reframe around the mechanism or route to a domain venue.64- "Dataset-only paper." Add an evaluation or representation advance, or target a datasets-and-65 benchmarks track instead.66- "Theory disconnected from practice." Tie the result to an observed deep-learning phenomenon.6768## Output format6970```text71[ICLR fit] strong / plausible / weak / no72[Core learning insight] <one sentence>73[Evidence required] <theory, experiment, benchmark, artifact>74[Best venue route] ICLR / NeurIPS / ICML / AISTATS / UAI / domain venue / workshop75[Reframe] <how to make the paper more ICLR-shaped>76```7778---7980**Source:** [`brycewang-stanford/Awesome-Journal-Skills`](https://github.com/brycewang-stanford/Awesome-Journal-Skills) → `ICLR-Skills/skills/iclr-topic-selection/SKILL.md`
thedixitjain