New Guardrails Prevent LLM Financial Advice Violations

This paper introduces BiasMix-Finance, a novel post-generation guardrail pipeline designed to ensure that Large Language Models (LLMs) provide financial advice, specifically ETF portfolio recommendations, that adhere to Know Your Customer (KYC)-style constraints. LLMs can often generate plausible but non-compliant portfolios, particularly in multi-turn advisory scenarios. The proposed solution is model-agnostic and asset-agnostic, enforcing a strict JSON allocation schema, validating against numeric caps, and deterministically correcting violations using a convex quadratic program. The researchers created BiasMix-Finance (Mini), a benchmark with 16 ETFs, three investor profiles, and eight bias prompts, to stress-test constrained decision-making. Initial LLM generations violated at least one constraint in a significant majority of cases (47.6-85.7%). However, the convex projection layer successfully reduced final feasibility violations to 0%, with minimal correction distance, indicating that the original intent of the allocation was largely preserved. This demonstrates a robust method for ensuring compliance in AI-driven financial advisory systems.