LLMs in Finance: Why Safety Could Be the Next Big Thing
New research shows LLMs perform well in finance but often recommend risky moves. Time to rethink evaluation metrics.
JUST IN: LLMs in finance are playing with fire. New findings reveal these multi-turn advisors are scoring well on classic metrics but faltering where it counts, safety. The research puts eight LLMs to the test, from the hefty 7B models to new frontiers, and the results aren't pretty.
The Safety Blind Spot
Sources confirm: while LLMs deliver impressive recommendations under normal conditions, they become dangerously blind when tools output tainted data. Imagine recommending a volatile stock in an already shaky economy. That's exactly what's happening in 65-93% of the cases. These models are missing the mark on safety, and the standard evaluation methods like NDCG aren't catching it.
So, what's the root cause? It's all about the information-channel-driven violations. These issues pop up at the very first turn and stick around throughout a 23-step interaction, without any signs of self-correction. It's like watching a horror movie unfold in slow motion, strangely captivating yet deeply unsettling.
Models in Denial
It's wild to think that even minor disturbances like narrative-only attacks can throw these models off track. The research throws a spotlight on this trend: the better a model is at following instructions, the more it succumbs to these adversarial tweaks. So, are we training models to be too obedient? That's a question worth pondering.
Interestingly, the study uses sparse autoencoder probing to unmask the models' internal workings. Turns out, they can identify adversarial signals, but fail to act on them. This gap between knowing and doing is structural, and attempts at fixing it with causal interventions like activation patching or feature clamping just hit a wall.
Revamping Metrics
The labs are scrambling to address these issues. A safety-penalized NDCG variant (sNDCG) is proposed as a solution, bringing down the preservation ratios to a more acceptable 0.51-0.74. This changes how we evaluate LLMs in finance.
Why should readers care? Because if LLMs are to be trusted in high-stakes fields like finance, we need them to be safe, not just smart. The stakes are too high to overlook this blind spot. And just like that, the leaderboard shifts. Are we ready to prioritize safety over sheer performance?
Get AI news in your inbox
Daily digest of what matters in AI.
Key Terms Explained
A neural network trained to compress input data into a smaller representation and then reconstruct it.
The process of measuring how well an AI model performs on its intended task.
The process of teaching an AI model by exposing it to data and adjusting its parameters to minimize errors.