Reading Between the Tokens: Improving Preference Predictions through Mechanistic Forecasting

TL;DR

This paper argues that large language models contain structured, latent information about human preferences that is not reliably expressed in surface outputs. It introduces mechanistic forecasting, which identifies party-aligned value vectors in MLPs and aggregates persona-induced activations to produce distributional predictions comparable to real-world surveys. Across seven model families and six national elections, the approach yields improvements over probability-based predictions in many settings, with demographic attributes and high-entropy attributes benefiting most. The work highlights a new path for social-science prediction tasks by leveraging internal model representations, while emphasizing that latent signals should complement, not replace, traditional surveys. Practical guidance includes using entropy-based gating to decide when to apply latent-based estimators, and acknowledges limitations such as the need for white-box access and substantial compute.

Abstract

Large language models are increasingly used to predict human preferences in both scientific and business endeavors, yet current approaches rely exclusively on analyzing model outputs without considering the underlying mechanisms. Using election forecasting as a test case, we introduce mechanistic forecasting, a method that demonstrates that probing internal model representations offers a fundamentally different - and sometimes more effective - approach to preference prediction. Examining over 24 million configurations across 7 models, 6 national elections, multiple persona attributes, and prompt variations, we systematically analyze how demographic and ideological information activates latent party-encoding components within the respective models. We find that leveraging this internal knowledge via mechanistic forecasting (opposed to solely relying on surface-level predictions) can improve prediction accuracy. The effects vary across demographic versus opinion-based attributes, political parties, national contexts, and models. Our findings demonstrate that the latent representational structure of LLMs contains systematic, exploitable information about human preferences, establishing a new path for using language models in social science prediction tasks.

Figures (7)

• Figure 1: Overview of the proposed mechanistic forecasting method. Persona prompts are mapped to party-aligned latent value-vector activations inside the LLM. These activations are aggregated across personas into party-level preference distributions and compared against real-world survey outcomes.
• Figure 2: Win-rates by model and country comparing mechanistic forecasting ($\Psi_g^{\text{latent}}$) and probability-based ($\Psi_g^{\text{prob}}$) preference distributions against survey data ($\Psi_g^{\text{survey}}$).
• Figure 3: Party-level estimation error for estimating conditional vote shares. Each point shows the median absolute error in estimating $P(\text{party}\mid\text{category})$ relative to survey benchmarks, with offsets indicating different models. Green points highlight the potential gains achievable by choosing mechanistic forecasting estimates when they outperform probability-based estimations.
• Figure 4: Latent win-rates by persona attribute and country, aggregated across models. Each bar reports the fraction of cases in which mechanistic forecasting is closer to survey estimates of category shares conditional on party than probability-based estimations.
• Figure 5: Median probability-based estimation error and corresponding improvement from substituting mechanistic forecasting estimates for category–given–party probabilities, evaluated on high-entropy attributes ($>0.85$). Green indicates error reduction, red indicates increased error, and gray indicates no change.