Automatic discovery of quality dimensions for abstract concepts in conceptual spaces

Develop data-driven methods to automatically discover interpretable quality dimensions for abstract concepts within the conceptual spaces framework, enabling dimension grounding without relying solely on heuristic expert specification.

Background

The paper’s chess proof-of-concept defines seven quality dimensions and corresponding domains using expert heuristics. While this enables initial region anchoring for strategies such as King Attack, Positional Sacrifice, and Space Domination, the authors emphasize that relying on expert-defined dimensions limits scalability and introduces subjectivity.

They explicitly identify the automatic discovery of quality dimensions for abstract, temporally evolving concepts as a significant open research question, highlighting the need for data-driven mechanisms to ground interpretable dimensions in conceptual spaces beyond perceptual domains.

References

The problem of automatically discovering quality dimensions for abstract concepts, learning concept boundaries from observational data, and validating geometric representations against human expert knowledge represents significant open research questions for both artificial intelligence and conceptual spaces theory.

— Abstract Concept Modelling in Conceptual Spaces: A Study on Chess Strategies  (2601.21771 - Banaee et al., 29 Jan 2026) in Discussion — Subsection “Challenges and Implications”

However, discovering the optimal semantic dimensions for a specific task without laborious hand-curation remains an open challenge.

— LLMs as Feature Engineers for Text-and-Tabular Prediction  (2609.21894 - Barlier et al., 18 Sep 2026) in Section 1, Introduction

Semantic dimension naming influences which features the agent explores---our churn experiments show this effect can be large---and principled dimension selection remains open.

— KnowFeat: Knowledge-Guided Feature Engineering via LLM Agents  (2609.03529 - You et al., 3 Sep 2026) in Limitations and Future Work, Section 5 (Conclusion)

All four coding agents share one frozen, mid-capability base model and are evaluated in a single run, so the study isolates scaffold effects under that model and supports the broad execution-versus-discovery profile, whose roughly 38--40-point gap between the strong and weak dimensions far exceeds any plausible run-level noise (Section~\ref{sec:breadth}), but not exact pairwise ordering, tight-margin equivalence, or whether identical gaps persist with frontier models; multi-seed and multi-model studies are the natural next step, and the planned protocol re-runs each system--task with fixed seeds where the configuration permits and reports the mean over $k$ repeated runs alongside the per-run spread.

— TruthInsightBench: An Evidence-Grounded Benchmark for Automated Evaluation of Open-Ended Scientific Discovery Agents  (2609.05079 - Yang et al., 4 Sep 2026) in Section 5.3, “Limitations and future work”