Implicit Bias in Large Language Models with Concept Learning Dataset (arxiv.org)

Researchers released a concept-learning task dataset designed to probe hidden biases in large language models (LLMs). By running in‑context concept learning experiments—where models learn a target concept from examples provided in the prompt—the authors found a consistent bias toward upward monotonicity in quantifiers: models preferentially generalize in ways that treat quantifiers as preserving truth under supersets (e.g., favoring inferences like “if some A are B, then some larger set is B”). This monotonicity bias surfaced strongly during concept learning but was much weaker when models were evaluated with direct prompting (no concept-learning component), showing that the probing method itself can reveal latent model tendencies. The finding matters for AI/ML because quantifier monotonicity underpins logical reasoning, natural-language inference, and fairness-sensitive decisions; a systematic upward bias can produce predictable errors in QA, rule induction, and downstream systems that rely on correct subset/superset reasoning. Technically, the paper provides a reusable benchmark and experimental protocol that leverage in‑context learning as a diagnostic tool, suggesting teams should include concept‑learning probes alongside standard zero‑shot and fine‑tuned tests when auditing LLM reasoning and bias. Code, data, and demos accompany the arXiv submission to help reproduce and extend the analyses.