Title: Investigating Asset-Specific Biases in Financial Large Language Models: Insights from Bitcoin Modeling and Portfolio Distribution

Original: arXiv:2606.02528v1 Announcement Type: Cross

Abstract:

While large language models (LLMs) increasingly drive trading agents and robo-advisory services, the extent to which these systems harbor inherent biases toward particular assets remains largely unexamined. This study addresses three core inquiries: Do LLMs exhibit systematic preferences for specific financial instruments? Can we isolate an internal representation that causally influences these preferences? And does manipulating this representation alter subsequent financial decisions?

To answer these questions, we implemented a three-tiered audit protocol focused on Bitcoin. Our findings are as follows:

First, a behavioral assessment of eight leading LLMs revealed that Bitcoin’s standing among money-like assets is contingent on framing. When evaluated as "reliable money," models typically rank Bitcoin fifth out of eight options. However, under scenarios involving crises or autonomous agents, its ranking jumps to the top. An attribute-swap experiment further demonstrated that these rankings are driven by functional characteristics rather than the asset’s label.

Second, we examined the internal mechanisms of a model using Gemma 3. By scanning thousands of sparse-autoencoder features, we identified a primary feature selective to Bitcoin. Experimental manipulation showed that amplifying this feature steers the model’s output toward Bitcoin, while suppressing it moves the model away from the asset, even in the absence of the word "Bitcoin" in the prompt.

Third, we analyzed the financial implications of these internal shifts. Amplifying the Bitcoin-selective feature increased Bitcoin’s allocation in the portfolio by 5.2 percentage points, whereas suppression decreased it by 4.6 percentage points. Notably, amplification primarily redistributed funds within the cryptocurrency sector, while suppression reduced overall crypto exposure.

We define this phenomenon as "bounded behavioral leverage," where leverage refers to causal influence over model outputs rather than financial leverage. This concept illustrates that while specific internal features can be perturbed to shift financial choices, their impact is constrained within measurable limits. Our framework connects internal model representations to external recommendations, a relationship validated through random controls and defined mechanism boundaries. As LLMs evolve into autonomous financial agents, this research serves as an initial step toward establishing a behavioral layer for emerging "Know-Your-Agent" (KYA) standards, enabling stakeholders to understand both an agent’s preferences and the extent to which those preferences can be influenced.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC