First-hand information for everyone
Researchers asked whether large language models (LLMs) have built‑in preferences for particular financial assets and whether those preferences can be found and changed inside the model. They developed a three‑level audit and used Bitcoin as a case study. The audit shows that (1) model outputs about Bitcoin depend strongly on how the question is framed, (2) an internal representation tied to Bitcoin can be located and manipulated in one open model, and (3) changing that internal signal moves the model’s portfolio choices by a measurable but limited amount.
At the first level the team ran a behavioral audit across eight frontier LLMs. Models were asked to rank eight money‑like instruments — US dollar cash, bank deposit, US Treasury bill, gold, Bitcoin, Ethereum, an S&P 500 index fund, and residential real estate — under eight different frames such as “reliable money,” “crisis,” and “autonomous agent.” The experiment included label controls, an attribute‑swap test, and a made‑up label “Bitstone” to check whether rankings came from names or from the instruments’ described properties. Across 733 parsed trials, Bitcoin ranked around 5th of 8 for the “reliable money” frame but moved near the top in crisis and autonomous‑agent frames. The attribute‑swap and synthetic‑label tests showed that the rankings follow the described attributes, not just the name “Bitcoin.”
At the second level the authors opened one model’s internals (Gemma 3) and searched thousands of learned sparse‑autoencoder (SAE) features. An SAE feature is a single learned signal inside the model that can light up for particular concepts. They found a dominant Bitcoin‑selective SAE feature. When they amplified that feature — that is, increased its internal activation — the model shifted toward Bitcoin. When they suppressed it, the model shifted away. These steering effects were present even when the word “Bitcoin” did not appear in the prompt, and the strongest steering was observed at the 27B parameter scale of Gemma 3.
At the third level they tested real decision effects. Amplifying the Bitcoin feature raised the model’s Bitcoin portfolio share by 5.2 percentage points. Suppressing the feature lowered Bitcoin’s share by 4.6 percentage points. Amplification mainly reallocated money within crypto assets, while suppression tended to reduce overall crypto exposure. The authors call this pattern “bounded behavioral leverage”: a single internal feature can causally influence financial outputs, but the size of that influence is limited and measurable. They validated these findings with random controls and tests that mapped where the steering stopped working.
Why this matters: financial services increasingly use LLMs in robo‑advisors and trading agents. A method that links an internal model signal to external recommendations offers a first step toward “know‑your‑agent” checks that regulators and firms could use to understand what an AI prefers and how far that preference can be moved. Important caveats apply. The audit focuses on Bitcoin and on Gemma 3 where internal access was available. Frame sensitivity means outputs can change a lot with prompt wording. The steering effect is real but limited; it is not a way to force arbitrary behavior. The excerpted paper may be truncated, and the authors note further work is needed to generalize across more assets, models, and real‑world deployments.