We investigate how large language models can support visual exploration of text corpora. We define the zero-shot visualization (ZSV) task: users describe concepts in natural language and documents are projected onto those concept axes for display. Building a practical ZSV system requires trade‑offs among feature functions, implementation efficiency, and pre/post‑processing choices. To study this, we create a benchmark that compares embedding similarity, direct semantic judgments, and conditional likelihood estimation. Across several datasets and use cases we measure semantic faithfulness, score fidelity, and computational cost. Results show that scoring with next‑token probabilities offers the best practical balance. We also apply the approach to unlabeled corpora to see how it behaves in realistic exploratory settings. Experiments highlight the need for graded axes together with binary relevance filtering and reveal a compositional sentiment bias in off‑topic documents. Finally we provide practical guidelines for building end‑to‑end ZSV baselines.
Review