Trade-off between reducing overconfidence and rationale diversity

Determine whether reducing post-training overconfidence in instruction-tuned language models comes at the cost of further reducing the lexical diversity of generated rationales.

Background

The experiments show that instruction tuning increases model confidence while consistently reducing cross-rationale variability, without consistently improving accuracy or calibration. These findings motivate determining whether interventions that reduce post-training overconfidence produce an additional loss of rationale diversity, which would reveal a potential trade-off between confidence correction and diversity preservation.

References

Future work could further investigate whether reducing post-training overconfidence comes at the cost of further reducing rationale diversity, and whether the observed uncertainty-diversity patterns extend beyond lexical variation to semantic diversity.

— Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity  (2608.13430 - Proskurina et al., 13 Aug 2026) in Section 5, Conclusion

Whether the observed confidence and diversity patterns extend to safety-sensitive or demographic-sensitive prompts requires separate evaluation.

— Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity  (2608.13430 - Proskurina et al., 13 Aug 2026) in Section 6, Ethical Considerations