OptimismBench: Forecasting Bias and the Alignment Effect in Language Model Judgment

2026-07-29

AI Summary

This study addresses the question of whether large language models exhibit systematic directional biases—such as undue optimism or pessimism—in probabilistic judgments, noting the absence of methods to detect such biases without ground-truth probability labels. To this end, the authors introduce OptimismBench, a novel benchmark that leverages paired inverse framing to quantify directional bias without requiring true probabilities. Evaluating 16 models across 10 languages, the experiments reveal that 14 models display significant optimism, while only Anthropic’s frontier model shows pessimism. Alignment training is found to predominantly shape bias direction, with model identity exerting a far stronger influence than linguistic variation. The work releases a dataset comprising 3,870 evaluations and validates its approach through multidimensional ablation studies and variance decomposition.

Abstract

Large language models are increasingly used as decision aids whose probability judgments shape downstream choices. Whether those judgments carry a systematic directional tilt has been hard to detect: calibration metrics aggregate unsigned errors, and naturalistic uncertainty offers no ground-truth probability. When an LLM rates a startup's success at 70% but its failure at 15%, the missing 15 points expose a distortion no aggregate score flags. We introduce OptimismBench, which detects directional bias with inverted pairs: each scenario elicits both P(success) and P(failure), and asymmetry between the two framings yields a signed bias score without ground truth. Across 16 models from 8 providers, fourteen are optimistic; pessimism appears only in Anthropic's frontier tier. Eleven matched base-versus-chat pairs across four families show post-training sets the sign of the bias, with opposite shifts in different families. The pattern survives prompt, temperature, perspective, and self-debiasing ablations. A seventeen-model six-language comparison further shows model identity dominates language, with inter-model variance at 4.7x inter-language variance. We release 3,870 items across 10 languages for per-model directional-bias auditing. When alignment makes a model more helpful, it also tilts its probabilities; downstream pipelines inherit the tilt by default.

Problem

• forecasting bias
• directional bias
• language model judgment
• probability calibration
• alignment effect

Innovation

• directional bias
• inverted pairs
• OptimismBench
• probability calibration
• post-training alignment