Your AI, On a Dial: Controlling Investment Bias in LLMs with a Single Neuron

Battle Count: 97

Highly relevant to quantitative trading as it addresses systematic biases in LLM-based investment decision-making. The paper demonstrates control over buy/sell priors, propagation to security rankings and portfolio composition, and effects on agentic research workflows. The backtest with weekly rebalanced top-100 portfolios directly connects to portfolio construction. However, it does not propose a trading strategy or establish alpha, and the intervention requires white-box model access.

The core intervention is simple (additive shift to one MLP coordinate via pre-hook), but the selection procedure requires gradient computation over all MLP coordinates, held-out calibration with security universe splitting, and model-specific tuning. Requires access to internal activations (not feasible for API-only models). The agentic retrieval and backtest evaluations add significant engineering complexity.

The paper provides detailed algorithm descriptions (Algorithm 1), prompt templates, evaluation protocols, and compute environment specifications. However, no code repository is linked. The method requires access to internal model activations and specific calibration procedures. Five open-weight models are used, making replication feasible but requiring significant compute resources (8x NVIDIA RTX PRO 6000 GPUs).