reasoningcritiquebearishLLM initial outputs for SPARQL generation remain unreliable - generated queries may be executable yet semantically misaligned with input questionsComputation and Language01 Aug 2026http://arxiv.org/abs/2607.28082v1