技能 编程开发 机器学习理论可复现性审计

机器学习理论可复现性审计

v20260724
colt-reproducibility
本指南提供了理论机器学习论文的严格可复现性审计标准。它要求作者确保论文中的所有数学论证都是自洽、可追溯的,包括完整的假设声明、参数常量跟踪以及数值图表的生成流程。用于提升论文的严谨性和可信度,以应对顶会严格的审稿过程。
获取技能
81 次下载
概览

COLT Reproducibility

At COLT, reproducibility means a competent reader can re-derive every claim from the PDF alone. There is no reproducibility checklist form in the COLT 2026 CFP (checked 2026-07-08) — the venue enforces the property the hard way, through referees who attempt the re-derivation during review. This skill is the pre-submission audit that makes that attempt succeed.

The re-derivability standard

A theorem is reproducible when all of the following hold:

  • Its statement is formally complete: every symbol quantified, every assumption named in the statement itself or cited by an explicit assumption label, the probability space and adversary model unambiguous.
  • Its proof exists in full inside the submitted PDF — COLT's unlimited appendix removes every excuse for "omitted for lack of space."
  • Each proof step is locally checkable: an expert reading line k needs only lines 1..k-1, cited external results, and standard background — never your unpublished intuition.
  • Constants and parameter regimes survive the chain: if Theorem 1 needs n ≥ C·d·log(1/δ), the reader can trace what C is or where it was declared absolute.

Audit table: where re-derivation fails

Failure mode Typical symptom in the PDF Repair
Hypothesis smuggling Proof uses independence never assumed Add the assumption to the statement, or redo the step
External-result misuse "By [X], ..." where [X] needs bounded support you don't have Check [X]'s hypotheses; find the right variant or prove a lemma
Constant drift C doubles silently between two displays Number constants (C_1, C_2, ...) and track them in a ledger
Case leakage "The case d=1 is analogous" when it is not Write the case or prove the reduction
Notation overload Same symbol for a filtration and a function class One notation table, enforced globally
Silent regime switch Bound proved for T large, quoted for all T State the threshold explicitly in the theorem

Assumption bookkeeping pattern

Give assumptions their own numbered environment and cite them by label everywhere:

\newtheorem{assumption}{Assumption}

\begin{assumption}[Bounded losses]\label{ass:bounded}
For all $t$, the loss $\ell_t$ maps to $[0,1]$.
\end{assumption}

\begin{theorem}\label{thm:regret}
Under Assumptions~\ref{ass:bounded} and~\ref{ass:oblivious}, the algorithm's
regret satisfies $R_T \le 4\sqrt{T \log K}$ for all $T \ge 1$.
\end{theorem}

The payoff is auditable dependency: a reviewer can grep which theorems rely on obliviousness, and your rebuttal can answer "is Assumption 2 needed for Theorem 3?" with a pointer instead of an essay.

Numerical illustrations, when present

Some COLT papers plot a simulated regret curve or a phase transition to illustrate a bound. The theory community's floor for those figures:

  • Fixed seeds, stated replication counts, and error bars whose meaning the caption defines; a noisy single run "consistent with the theory" persuades no one here.
  • The simulated regime must be the theorem's regime — matching horizon, dimension, and noise assumptions — or the mismatch must be acknowledged as exploratory.
  • The generating script should be one dependency-light file, releasable after acceptance; during review, describe the procedure precisely in the appendix since no upload channel exists.
  • Never let an illustration silently extend the claim ("the bound appears to hold for heavy tails too") without labeling it as conjecture.

Vignette: the spine of a lower-bound paper

Consider a submission whose main result is a $\Omega(\sqrt{TK})$ lower bound via a new instance family. Its re-derivability spine, in the order a referee will attack it:

  • the instance family's construction, with every distribution parameter explicit and the randomization protocol (oblivious vs. adaptive) named;
  • the information-theoretic step (say, a KL-divergence calculation) with the exact divergence bound displayed, not cited as "standard";
  • the reduction from learner performance to the divergence quantity, where hypothesis smuggling most often hides (does the argument secretly assume deterministic learners? say so or generalize);
  • the final optimization over instance parameters, with the maximizing choice written out — "choosing ε appropriately" is where constants go to die.

A referee who can walk this spine without leaving the PDF marks correctness resolved; each externalized step converts into a review question, and three review questions into a reject.

Pre-submission re-derivation drill

  1. Print the numbered-statement list (all definitions, assumptions, lemmas, theorems).
  2. For each, a non-author coauthor answers: can I state precisely what this claims, including quantifiers, without reading the proof? Rewrite until yes.
  3. Verify proofs in dependency order, marking each line verified/unverified; the colt-artifact-evaluation skill's ledger format works here.
  4. Re-check every external citation against the cited source's actual hypotheses — allocate real time; this is where careful papers die.
  5. Reconcile body sketches against appendix proofs: a sketch that describes an older proof strategy than the appendix executes reads as a gap to a referee.

Cycle-volatility warnings

  • If a future COLT cycle adds any checklist, code policy, or supplementary channel, the current CFP announces it; the 2026 cycle had none (待核实 in later cycles).
  • The 12-page body and single-PDF rules that shape where proofs live are the 2026 formulation; re-read the live CFP before restructuring a paper around them.
  • Formatting of assumptions and environments is a house-style choice, not a CFP rule; the CFP-level constraints remain the 12-page body and the single PDF.

Output format

[Re-derivability verdict] re-derivable / gaps found
[Statement completeness] <theorems needing quantifier or assumption repair>
[Constant ledger] tracked / drift at <displays>
[External-results audit] <citations with unchecked hypotheses>
[Illustration floor] seeds+replications stated / absent / no numerics in paper
信息
Category 编程开发
Name colt-reproducibility
版本 v20260724
大小 6.41KB
更新时间 2026-07-28
语言