Technical Summary: Aligning LLMs with Human Uncertainty via a Beta-Bernoulli Calibrator

Problem Statement

Probabilistic forecasting is critical for decision-making in domains ranging from economics to national security. While Large Language Models (LLMs) possess broad knowledge and reasoning capabilities, they often struggle to outperform skilled human forecasters. Existing methods to improve LLM forecasting typically rely on supervised fine-tuning (SFT) or reinforcement learning (RL) using binary outcomes (e.g., whether an event occurred or not). However, these approaches face significant limitations: they are resource-intensive, often inapplicable to black-box models, and fail to leverage the rich information contained in human forecasts. Specifically, human forecasts provide not only an aggregate probability estimate but also a degree of consensus (agreement) among forecasters, which encodes valuable signals about human sentiment and epistemic uncertainty. Current methods largely ignore this distributional information, focusing instead on point estimates or verbalized confidence scores, which have been shown to be overconfident and noisy.

Methodology: The Beta-Bernoulli Calibrator (BBC)

The authors propose the Beta-Bernoulli Calibrator (BBC), a lightweight, model-agnostic post-hoc calibration framework. Unlike methods that attempt to fine-tune the underlying LLM, BBC acts as a separate, smaller model that refines the initial verbalized forecast from any input LLM.

Core Framework

The framework models the latent event probability $p$ as a random variable drawn from a Beta distribution, $p \sim \text{Beta}(\alpha, \beta)$, and the observed binary outcome $y$ as a realization from a Bernoulli trial, $y \sim \text{Bernoulli}(p)$.

• Input: The calibrator takes a forecasting question ($x_i$) and an initial verbalized point forecast ($\hat{p}^{init}_i$) from an input LLM.
• Output: Instead of outputting a single probability, the calibrator predicts the parameters of a Beta distribution (or a mixture of $K$ Beta distributions for flexibility).
  • The mean of the predicted distribution serves as the calibrated point forecast: $\hat{p}_i = \mathbb{E}[p_i] = \frac{\hat{\alpha}_i}{\hat{\alpha}_i + \hat{\beta}_i}$.
  • The variance serves as a measure of epistemic uncertainty: $\hat{u}_i = \text{Var}[p_i]$.

Training Objective

The calibrator is trained using a composite loss function $\mathcal{L}_{total} = \mathcal{L}_{binary} + \mathcal{L}_{human}$:

1. Binary Outcome Supervision ($\mathcal{L}_{binary}$): The model learns from the realized binary outcome ($y_i \in \{0, 1\}$). The authors demonstrate that maximizing the likelihood of the binary outcome under the Beta-Bernoulli model is mathematically equivalent to minimizing the Binary Cross-Entropy (BCE) loss with respect to the mean of the distribution. This ensures the model learns to predict the true probability.
2. Human Forecast Supervision ($\mathcal{L}_{human}$): To address the identifiability problem (where different Beta distributions can yield the same mean and BCE loss) and to capture distributional shape, the model is supervised by human forecasts. The authors treat human forecasts as noisy samples from the true latent distribution. They construct a histogram of human forecasts and minimize the Kullback-Leibler (KL) Divergence between this histogram and the predicted Beta (or mixture of Betas) distribution. This allows the model to learn the "consensus" and uncertainty inherent in the crowd.

Architecture

The calibrator is implemented as a small language model (e.g., Llama-3.2-1B) with a Multi-Layer Perceptron (MLP) head. It encodes the input text and the initial forecast, then predicts the Beta parameters. This design ensures the method is computationally efficient and can be applied to any black-box LLM without accessing its internal weights or representations.

Key Contributions

1. Beta-Bernoulli Calibrator: A lightweight, model-agnostic method that converts point forecasts into distributions over event likelihood, effectively capturing both aleatoric (inherent randomness) and epistemic (model uncertainty) uncertainty.
2. Humans as Distributional Supervision: The paper introduces a novel approach of using human forecast distributions (aggregated from platforms like Metaculus and Polymarket) as a supervisory signal. This provides information beyond binary outcomes, specifically regarding the shape of the probability distribution and the degree of consensus among forecasters.
3. Epistemic Uncertainty as a Predictor: The authors show that the variance of the predicted Beta distribution is a more reliable predictor of forecasting error than verbalized confidence scores provided by LLMs.

Experimental Results

The framework was evaluated on datasets from Metaculus and Polymarket, with out-of-distribution (OOD) testing on Kalshi.

• Performance vs. Baselines: BBC consistently outperformed traditional post-hoc calibration methods (Platt Scaling, Isotonic Regression) and models specifically fine-tuned for forecasting (e.g., OpenForecaster-8B, Future-as-a-label-32B).
  • On the test set, BBC reduced the Brier score by approximately 14.4% compared to verbalized baselines (e.g., from 0.146 to 0.125 for Claude-Sonnet-4) and improved AUC from 72.3% to 74.2%.
  • BBC achieved better calibration (lower Expected Calibration Error) and stronger discrimination (higher AUC) than fine-tuned models, suggesting that a lightweight calibrator is a more efficient alternative to full model fine-tuning.
• Impact of Human Supervision: Comparing "BBC (binary only)" with "BBC (binary + human)" revealed that incorporating human forecast distributions consistently improved Brier scores and AUC. The KL divergence between the predicted distribution and human forecasts decreased, indicating the model better aligned with human beliefs.
• Epistemic Uncertainty: Analysis showed that the variance of the Beta distribution correlated strongly with prediction errors (Brier score). In contrast, verbalized confidence scores were noisy and often uncorrelated with actual performance.
• Generalization: The calibrator demonstrated robust generalization to the OOD Kalshi dataset, maintaining strong performance where traditional calibration methods failed.
• Efficiency: Experiments confirmed that a 1-billion parameter calibrator is sufficient to effectively refine forecasts from much larger models (e.g., 70B parameters), highlighting the computational efficiency of the approach.

Significance and Claims

The paper claims that the Beta-Bernoulli Calibrator offers a significant advancement in LLM forecasting by:

1. Bridging the Gap: It effectively utilizes the "wisdom of the crowd" not just as a point estimate, but as a distributional signal to model uncertainty, addressing a gap in current LLM forecasting literature.
2. Model Agnosticism: By operating as a post-hoc calibrator, it allows the calibration of any LLM (including proprietary black-box models) without the need for expensive fine-tuning or access to internal states.
3. Reliable Uncertainty Quantification: It provides a principled way to estimate epistemic uncertainty that is predictive of error, moving beyond the unreliable self-reported confidence of LLMs.
4. Complementarity: The method is presented as complementary to existing techniques; it can be applied on top of models already fine-tuned for forecasting to yield further performance gains.

The authors conclude that while human forecasters remain superior to current LLMs, the BBC framework significantly narrows this gap by leveraging human consensus signals to calibrate model beliefs, offering a scalable and efficient path toward more reliable AI forecasting.