Instruction Tuning Alters Confidence and Rationale Diversity in Language Models

A recent study published on arXiv (2608.13430) examines the impact of instruction tuning on the confidence levels and lexical variety of language models in question-answering scenarios. The research assesses three pairs of base and instruction-tuned models against various question-answering benchmarks. Results indicate that instruction tuning consistently modifies answer confidence, even though predictive accuracy shows minimal changes and likelihood-based calibration declines. Furthermore, the influence on rationale diversity is inconsistent: cross-rationale diversity tends to decrease, while surface-level lexical diversity fluctuates in both directions. This paper was authored by a team of researchers and is noted as a cross submission on arXiv.