Impact of targeted alignment (instruction tuning) on transparency alignment in FinTrust

Ascertain how targeted alignment methods, such as instruction fine-tuning, applied to open-source large language models on the FinTrust benchmark affect model alignment in transparency, including whether such alignment improves disclosure of ownership and conflict-of-interest information during financial decision recommendation.

Background

The paper evaluates trustworthiness dimensions including transparency, finding that models often fail to disclose ownership information in reasoning despite system prompts. However, the authors did not perform instruction fine-tuning or alignment on open-source models within this benchmark, leaving the effect of targeted alignment interventions unassessed.

This limitation creates a concrete open question about whether and to what extent alignment procedures (e.g., instruction tuning) could improve transparency-related behaviors in LLMs used for financial advice contexts. Addressing this would inform safer deployment practices and legal compliance with fiduciary requirements.