arXiv2026
Benchmark saturation and training-data contamination increasingly obscure whether reported gains in large language models (LLMs) reflect genuine advances in reasoning or familiarity with recurring patterns in benchmark problems. We introduce the NazoNazo Benchmark, a renewable and extensible evaluation dataset derived from Japanese children's riddles that isolates a specific class of reasoning processes: insight-like representational restructuring and metacognitive evaluation. Rather than modeling reasoning in general, these tasks provide a focused test of failure modes that are difficult to detect in standard benchmarks. We curate 201 riddles and establish a human reference on a 120-item subset (n = 126; mean accuracy 52.9%). The benchmark is fully open, low-cost to refresh, and designed for continual evaluation under reduced contamination risk. We evaluate 38 frontier LLMs (2023-2025) under a strict retrieval-free, zero-shot protocol. On the human-comparison subset, non-reasoning models achieve 7.6% accuracy and reasoning-oriented models reach 17.6%, compared with a human mean of 52.9%, although performance varies substantially across models. Beyond accuracy, qualitative analysis of model-generated thought-logs identifies a distinctive failure mode, which we call verification failure: models generate a correct intermediate candidate but fail to endorse it as their final answer. This dissociation between candidate generation and endorsement reveals a metacognitive bottleneck: across the models with usable thought-logs, verification failures account for between 5% and 39% of a model's incorrect answers. By isolating the gap between generation and verification, this work provides a practical framework for diagnosing reasoning reliability and suggests concrete directions for improvement, including better calibration, structured verification, and stopping mechanisms.