uai-topic-selection

UAI Topic Selection

Use this before writing a line. UAI — the AUAI's annual conference, running since the

1980s — is the venue where uncertainty is the subject, not the seasoning. The 2026 CFP

invited novel theory, methodology, and applications spanning AI, machine learning, and

statistics, but the reviewer pool and the accepted-paper record reward a specific

shape: papers whose contribution is a probabilistic representation, an inference

procedure, a causal identification result, or a decision rule under uncertainty.

The one-question filter

If you deleted every probabilistic element from this paper, would anything remain?

• Remains a strong paper → the probability is decoration; route to a general ML venue.
• Nothing remains → you are holding a UAI candidate; now check the evidence shape.
• A weaker but real paper remains → the venue call depends on which half is novel;

interrogate with the table below.

Signal table

| Project signal | UAI reading |

|---|---|

| New inference algorithm (MCMC, SMC, variational, belief propagation) with analysis | Home territory since the venue's founding |

| Identifiability or discovery result for causal or graphical structure | Core — UAI is a first-choice causality venue |

| Calibration, conformal, or coverage guarantee for predictive models | Strong fit; growing lane in recent volumes |

| Decision making / planning under uncertainty, Bayesian experimental design | Core, especially with formal treatment |

| Probabilistic programming semantics or inference | Distinctive UAI lane, rare elsewhere |

| Deep architecture, uncertainty used only as an evaluation metric | Re-route: NeurIPS / ICML / ICLR |

| General estimation theory, uncertainty incidental | Often better at AISTATS |

| Pure regret/sample-complexity theory, no probabilistic-modeling core | COLT or ALT |

| Causal ML with an applied-community audience | Compare CLeaR before defaulting to UAI |

| Journal-depth asymptotics needing 40 pages | JMLR or a statistics journal |

Sibling-venue geometry

AISTATS and UAI are the commonly confused pair — both single-deadline, PMLR-published,

statistics-adjacent, and of comparable scale. A working separation: AISTATS emphasizes

the statistics–ML interface broadly (estimators, rates, high-dimensional methods);

UAI concentrates on the representation and use of uncertainty itself — graphical

models, causality, Bayesian reasoning, decisions. A debiased estimator with a

convergence rate leans AISTATS; an identifiability theorem over MAGs leans UAI; a

calibrated-prediction method with finite-sample coverage plays at either, so decide by

which reviewer conversation helps the work more.

Timing is a legitimate tiebreaker between honest fits: UAI's cycle (submission ~February,

decision ~June) interleaves with AISTATS (~October submission) and NeurIPS/ICML, and a

paper genuinely at home in two venues may reasonably pick the calendar that meets it

ready. Never let timing overrule fit — a misrouted paper burns a cycle anyway.

Sharpening a genuine fit

1. Name the probabilistic object the paper contributes: a posterior approximation, a

graph class, an identification condition, an interval, a policy. No object, no UAI

framing.

1. Name the guarantee or diagnostic attached to it: consistency, coverage, ESS

efficiency, SHD improvement, regret bound. This becomes the page-1 promise.

1. Check the evidence shape fits an 8-page main part with appendices — one decisive

exact-truth experiment beats six benchmark sweeps here.

1. Scan the current CFP's topic list (2026 spanned inference families, optimization,

probabilistic programming, and application areas from computational biology to

robotics) to phrase your subject areas in the venue's own vocabulary.

Borderline vignettes

Three recurring hard calls, with the reasoning that resolves them:

• "We add a Bayesian layer to a transformer and report accuracy plus ECE."

The deletion test partially survives (the architecture stands alone), and the

uncertainty content is evaluation-grade, not contribution-grade. As described:

NeurIPS/ICML material. It becomes UAI-shaped only if the paper's center of gravity

moves — e.g., a posterior-approximation scheme with analyzed bias, or a calibration

guarantee — and the architecture becomes the test bed.

• "We prove a regret bound for Thompson sampling in a structured bandit."

Probabilistic object (posterior sampling policy) and guarantee (regret) both

present. The split: if the novelty is the analysis technique, COLT/ALT readers

value it most; if the novelty is the modeling of uncertainty in the structure

(priors over graphs, correlated arms), UAI is the natural room.

• **"We identify treatment effects under a new missingness mechanism, with an

estimator and simulations."** Identification result + estimator + coverage

simulations is simultaneously UAI-core and AISTATS-core. Decide by conversation:

UAI if the paper leans on graphical criteria and reasoning about the mechanism;

AISTATS if it leans on estimation theory and rates. Either way, do not split the

difference in the framing — pick one audience and write for it.

Calendar map for re-routes

If the verdict is "route elsewhere", the practical question becomes when. Approximate

rhythm of the neighbors (always verify each venue's current dates):

• AISTATS: submissions historically in early autumn, conference in spring — the

natural next stop after a February UAI miss.

• NeurIPS: late-spring deadline; ICML: winter deadline close to UAI's own — an ICML

reject can often turn around for UAI in the same season, and vice versa.

• COLT: winter deadline, summer conference; CLeaR: autumn-ish deadline for the

causality community.

• JMLR / statistics journals: rolling — the release valve when the contribution

outgrew 8 pages, not a consolation prize.

Keep the re-route target written down before the UAI decision arrives; deciding

while disappointed produces prestige-chasing, not fit-chasing.

Anti-signals worth trusting

• The draft's introduction cites no UAI, AISTATS, or CLeaR papers at all — the

community you are addressing does not include the one you are submitting to.

• The strongest section is a systems/scaling result and the probabilistic section

defends design choices rather than proving properties.

• The reviewers you would wish for are all from one distant subfield.
• You cannot fill the [Attached guarantee] line below without the word "hope".

Output format

[Deletion test] survives without probability? yes / partially / no
[Probabilistic object] <posterior / graph / interval / policy / condition>
[Attached guarantee] <the formal or diagnostic promise>
[Verdict] UAI-first / UAI-viable / route to <AISTATS | NeurIPS/ICML | COLT | CLeaR | journal>
[Next step] <framing fix, missing experiment, or venue switch>