Auditing Asset-Specific Preferences in Financial Large Language Models: Evidence from Bitcoin Representations and Portfolio Allocation

Battle Count: 62

The paper is moderately relevant to quantitative trading. It does not develop trading strategies or directly address market microstructure, but it is highly relevant to the growing use of LLMs as autonomous trading agents and robo-advisors. The findings on frame-dependent asset preferences, bounded behavioral leverage, and portfolio allocation effects (5.2pp Bitcoin shift) directly impact how LLM-powered trading systems should be audited and governed. The KYA framework and activation steering methodology could inform bias detection in algorithmic trading systems that use LLMs for asset selection or portfolio construction. However, the paper focuses on monetary asset ranking rather than price prediction or execution optimization.

High implementation complexity. Requires: (1) access to open-weight LLMs with hook points for residual stream intervention; (2) trained sparse autoencoders (GemmaScope 2) with 16,384 features per layer; (3) differential activation search across thousands of features; (4) careful experimental design with 9 models × 8 instruments × 8 frames × 5 conditions; (5) 324 investor profiles for portfolio allocation; (6) 320 random controls for specificity validation; (7) transcoder attribution graph construction. The behavioral audit is more accessible (API-only), but the representation and decision audits require deep model internals access and specialized SAE infrastructure.

The paper provides a publicly accessible replication package at https://anonymous.4open.science/r/asset-preference-audit-6F03/ containing prompt specifications, collection and analysis code, saved analysis tables, derived graph summaries, and figure-generation code. The verification script reproduces 581 archived-output consistency checks. However, the anonymous link may limit long-term accessibility, and the analysis requires specific SAE models (GemmaScope 2) and model weights (Gemma 3 family) that may not always be available.