Conference Agenda

We use sparse autoencoders (SAEs) to ``brain scan'' large language models, mapping internal activations to interpretable concepts and enabling targeted ``steering'' of specific features. We show that prompting models to adopt specific preferences or demographics causes severe concept contamination, invalidating causal inference in agent simulations. In contrast, SAEs permit smooth, isolated steering of target concepts without collateral activation. Furthermore, our scans reveal a stark disconnect between a model's true internal focus and its self-generated explanations, highlighting flawed introspection. Ultimately, SAEs provide a lightweight intervention for existing models, adding interpretability to embedding-based methods and surgically correcting economic biases.

Large language models (LLMs) cannot be trusted for economic forecasts during periods covered by their training data. Under black-box access, counterfactual forecasting ability is non-identified when the model has seen the realized values: any observed output is consistent with both genuine skill and memorization. Any evidence of memorization represents only a lower bound on encoded knowledge. We demonstrate LLMs have memorized economic and financial data, recalling exact values before their knowledge cutoff. Instructions to respect historical boundaries fail to prevent recall-level accuracy, and masking fails as LLMs reconstruct entities and dates from minimal context. Post-cutoff, we observe no recall. Memorization extends to embeddings.

We study how market returns shape news consumption using 700 million pageviews over 27 months from Australia’s largest newspaper, the Australian Financial Review. Aggregate news consumption intensifies after the market index decreases. After major market movements, news consumption of markets news increases while declining for firm news. In panel regressions, firm stock return movements increase firm-specific news consumption while index return movements dampen it. These findings imply aggregate and firm-specific news are substitutes, supporting limited attention theories. Highlighting a news demand mechanism, market returns drive consumption of stale news articles, even without new articles about the firm on that day.