Is Capability a Liability? More Capable Language Models Make Worse Forecasts When It Matters Most

2026-05-21 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Research reveals an inverse scaling phenomenon in large language models (LLMs) when applied to forecasting problems. These problems are characterized by superlinear growth and tail risk, common in finance and epidemiology. On these critical tasks, more capable LLMs surprisingly generate worse distributional forecasts. This pattern appeared on the new ForecastBench-Sim (FBSim) benchmark and synthetic SIR epidemics. It also replicated in real-world datasets for COVID-19, measles, housing markets, and hyperinflation. The failure primarily occurs in the upper tail of forecasts. Advanced models aggressively extrapolate growth, shifting this tail upward while the lower tail remains unchanged. A Llama-3.1 analysis indicated that both model scale and post-training independently contribute to this effect. Crucially, standard single-threshold metrics often used in LLM forecasting benchmarks fail to detect this upper-tail cost. They sometimes even reverse the perceived capability-accuracy relationship. The study recommends using continuous and unbounded accuracy measures alongside binary threshold metrics for robust LLM forecasting evaluations.

Key takeaway

For AI Scientists and Research Scientists developing or deploying LLMs for forecasting, you must critically re-evaluate your model evaluation strategies. If your applications predict superlinear growth or tail risk, relying solely on single-threshold accuracy metrics will mask significant upper-tail failures. More capable models exhibit this issue. You should implement continuous and unbounded accuracy measures, alongside traditional binary thresholds. This will accurately assess LLM performance and mitigate the risk of poor, aggressively extrapolated forecasts in critical scenarios.

Key insights

More capable LLMs make worse forecasts for superlinear growth and tail risk, especially in the upper tail.

Principles

• Inverse scaling can occur with LLMs on specific forecasting tasks.
• Model scale and post-training independently worsen tail forecasts.
• Single-threshold metrics can mask critical forecasting failures.

Method

The study used ForecastBench-Sim (FBSim), synthetic SIR epidemics, and real-world data to identify inverse scaling, analyzing per-quantile decomposition to pinpoint upper-tail failure.

In practice

• Evaluate LLM forecasts using continuous, unbounded accuracy metrics.
• Supplement binary threshold metrics with tail-inclusive scoring.
• Be wary of LLM forecasts for superlinear growth with tail risk.

Topics

• LLM Forecasting
• Inverse Scaling
• Tail Risk
• ForecastBench-Sim
• Model Evaluation
• Distributional Forecasts

Best for: AI Engineer, Machine Learning Engineer, NLP Engineer, AI Scientist, Research Scientist

Related on AIssential

• Don’t Blame the Model — LLM reliability issues often stem from API limitations, not inherent model flaws, restricting developer control and visibility.
• UnpredictaBench: A Benchmark for Evaluating Distributional Randomness in LLMs — LLMs consistently struggle to faithfully sample from target probability distributions, hindering their utility in stochastic simulations.
• Safety and accuracy follow different scaling laws in clinical large language models — Clinical LLM safety is a deployment property, not a passive consequence of scaling model size or compute.
• Impact of modelling assumptions on time horizon results — LLM time horizon estimates are highly sensitive to modeling choices and data noise, especially for frontier models.
• Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations — LLM robustness to CoT perturbations varies significantly by type and model scale, challenging assumptions that size alone ensures reliability.
• Hidden Reliability Risks in Large Language Models: Systematic Identification of Precision-Induced Output Disagreements — Numerical precision variations in LLM inference can systematically create critical safety vulnerabilities, leading to jailbreak divergence.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.