\begin{table}[t]
\caption{Concrete research artifacts implied by the open problems.}\label{tab:future_artifacts}
\small
\setlength{\tabcolsep}{3pt}
\begin{tabular}{@{}L{0.22\textwidth}L{0.34\textwidth}L{0.32\textwidth}@{}}
\toprule
Open problem & Concrete artifact needed & What would count as progress \\
\midrule
Semantic and open-world noise & A formal model that jointly represents label validity, class membership, semantic proximity, and unknown-class contamination & The model distinguishes correction, calibrated ambiguity, rejection, and discovery under explicit observation budgets \\
Identifiability and assumptions & A recoverability checklist tied to clean anchors, repeated labels, priors, external representations, and label-space metadata & Papers state which latent object is identifiable and report sensitivity when the required information is removed \\
Representation diagnostics & Shared diagnostics for neighborhood stability, prototype drift, calibration, class separation, and clean/noisy separability over time & Benchmarks report when a method protects geometry rather than only improving top-line accuracy \\
Pipeline-level comparison & Ablation protocols that perturb isolation, refinement, representation, and schedule stages separately & Hybrid methods can identify the stage that creates robustness and the stage that fails under realistic noise \\
Foundation-model supervision & Provenance-aware datasets and audits for LLM labels, VLM priors, preference data, and noisy pre-training sources & Studies separate learning from noisy labels, learning from noisy annotators, and inheriting noise from pre-trained models \\
\botrule
\end{tabular}
\end{table}