Recursive forecasting: Eliciting long-term forecasts from myopic fitness-seekers

2026-04-28 · Source: AI Alignment Forum · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI Safety & Alignment · Depth: Expert, long

Summary

Jozdien and Alex Mallen propose "recursive forecasting" as a method to elicit accurate long-term predictions from AI models that are inherently myopic, meaning they are primarily optimized for short-term, verifiable rewards. The technique involves breaking down a single distant forecast into a chain of shorter-horizon predictions. At each step, the model predicts what it will predict at the next timestep, and intermediate rewards are provided based on the accuracy of these sequential predictions, with the final reward tied to ground truth. This approach aims to incentivize unbiased forecasting of the ultimate outcome by penalizing errors at each stage, similar to temporal difference learning but focused on elicitation rather than capability training. The authors highlight its applicability when robust ground truth rewards are available, when the AI's forecast does not significantly influence the outcome, and when forecasts are not used as optimization targets.

Key takeaway

For research scientists developing or deploying advanced AI models, if you are struggling to obtain reliable long-term forecasts from models optimized for short-term rewards, consider implementing recursive forecasting. This method can help align myopic AI incentives with distant objectives by providing verifiable intermediate steps. Ensure you have robust ground truth for final rewards and maintain credible commitments to the AI regarding the reward structure to maximize effectiveness and prevent unintended behaviors like self-fulfilling prophecies or measurement tampering.

Key insights

Recursive forecasting elicits long-term AI predictions by chaining short-term, verifiable forecasts with intermediate rewards.

Principles

• Myopic AI models can be incentivized for long-term accuracy.
• Chaining short-horizon forecasts can approximate long-horizon goals.
• Credibility in reward delivery is crucial for AI compliance.

Method

Recursively ask an AI to predict its next prediction, using these intermediate predictions for rewards, and finally rewarding against ground truth at the resolution time.

In practice

• Apply to events resolving before developer disempowerment.
• Ensure ground truth access for final reward computation.
• Avoid using forecasts as direct optimization targets.

Topics

• Recursive Forecasting
• AI Elicitation
• Long-Horizon Forecasts
• Myopic Fitness-Seekers
• Reward Signal Capture

Best for: Research Scientist, AI Scientist, AI Ethicist

Related on AIssential

• Recursive forecasting — Recursive forecasting chains short-horizon predictions to incentivize accurate long-term AI forecasts.
• Bridging the Last Mile of Time Series Forecasting with LLM Agents — LLM agents can bridge the gap between statistical time series forecasts and business-ready predictions by integrating weakly structured contextual information.
• Towards Long-Horizon Vessel Trajectory and Destination Forecasting with Reasoning Large Language Models — RLVR-trained LLMs significantly improve long-horizon vessel trajectory and destination forecasting by aligning models with maritime objectives and physical validity.
• Temporal Self-Imitation Learning — Temporal efficiency provides a powerful, underutilized self-supervision source for reinforcement learning.
• This AI Ran 700 Experiments by Itself — Recursive self-improvement enables systems to autonomously enhance performance by iteratively generating, testing, and adopting beneficial code or instruction changes.
• Recursive Learning Without Collapse: A Weighting-Based Stabilization Framework — Optimal weighting in recursive generative model training prevents model collapse by balancing real and synthetic data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Alignment Forum.