Paper Overview
- arXiv ID: arXiv:2610.06889 [cs.CL]
- Subject Areas: Computation and Language (cs.CL); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
- Title: Zero-Shot Visualization: Exploring Text Corpora with User-Prompted Axes
- Authors: Arnau Bueno Tricas, Jose A. Rodríguez-Serrano
- Submitted: 23 September 2026
- DOI: 10.48550/arXiv.2610.06889 (arXiv-issued via DataCite)
Abstract
This work investigates how large language models (LLMs) can support the visual exploration of textual corpora. The authors introduce zero-shot visualization (ZSV), a task in which users specify concepts in natural language and documents are mapped onto the corresponding concept axes for visualization. Building a ZSV system of practical value is non-trivial, as it requires decisions at the intersection of feature functions, efficient implementation trade-offs, and pre-/post-processing choices that affect visualization quality.
To address this, the authors establish a benchmark comparing methods that span embedding similarity, direct semantic judgments, and conditional likelihood estimation. Across multiple datasets and use cases, they evaluate the properties of different scoring methods and design choices along three dimensions: semantic faithfulness, score fidelity, and computational cost.
The results identify that scoring based on next-token probabilities offers the strongest practical trade-off among the evaluated methods. The authors further apply this approach to unlabeled corpora to examine its behavior in realistic exploratory settings. These experiments highlight additional design considerations, including the use of graded axes together with binary relevance filtering, and reveal a compositional sentiment bias in off-topic documents. Based on these findings, the paper provides practical guidelines for constructing end-to-end ZSV baselines.
Key Contributions
- Formalization of zero-shot visualization (ZSV): A task where natural-language user prompts define concept axes onto which documents are projected.
- A comparative benchmark: Covers embedding similarity, direct semantic judgments, and conditional likelihood estimation.
- Evaluation criteria: Semantic faithfulness, score fidelity, and computational cost, tested across multiple datasets and use cases.
- Findings on scoring methods: Next-token probability scoring provides the best practical trade-off.
- Realistic exploration experiments: Applied to unlabeled corpora, revealing design considerations such as graded axes with binary relevance filtering.
- Bias observation: Identifies a compositional sentiment bias in off-topic documents.
- Practical guidelines: Recommendations for constructing end-to-end ZSV baselines.
Significance in 2026 Context
As LLM-based tooling matures in 2026, the ability to explore large text corpora through natural-language queries without task-specific training has become increasingly important for research, journalism, and enterprise analytics. ZSV sits at the intersection of representation learning and interactive data visualization, offering a flexible alternative to fixed-label topic models. By benchmarking scoring methods and documenting practical pitfalls—including sentiment bias and axis design trade-offs—the paper provides actionable guidance for researchers and practitioners building next-generation corpus exploration tools.
via ArXiv CL+LG