Reasoning and Tools for Human-Level Forecasting

TL;DR

The paper introduces RTF, a reasoning-and-acting framework for language models that enhances their forecasting abilities by integrating retrieval and simulation tools, enabling models to outperform humans in some forecasting tasks.

Contribution

The paper presents RTF, a novel framework combining reasoning, retrieval, and simulation tools to improve language models' forecasting capabilities beyond pattern mimicry.

Findings

RTF outperforms baseline models on forecasting tasks.

Models with RTF match or surpass human predictions.

Demonstrates potential for AI to reason and adapt like humans.

Abstract

Language models (LMs) trained on web-scale datasets are largely successful due to their ability to memorize large amounts of training data, even if only present in a few examples. These capabilities are often desirable in evaluation on tasks such as question answering but raise questions about whether these models can exhibit genuine reasoning or succeed only at mimicking patterns from the training data. This distinction is particularly salient in forecasting tasks, where the answer is not present in the training data, and the model must reason to make logical deductions. We present Reasoning and Tools for Forecasting (RTF), a framework of reasoning-and-acting (ReAct) agents that can dynamically retrieve updated information and run numerical simulation with equipped tools. We evaluate our model with questions from competitive forecasting platforms and demonstrate that our method is competitive with and can outperform human predictions. This suggests that LMs, with the right tools, can indeed think and adapt like humans, offering valuable insights for real-world decision-making.