The Future of Facts: Tracing the Factual Generation-Verification Gap

2026-05-26

AI Summary

This study addresses the significant gap between language models’ generative and veridical capabilities—termed the GV-gap—with respect to factual knowledge, whose dynamic evolution remains poorly understood. Distinguishing factual GV-gap from computational and aesthetic discrepancies, this work conducts a longitudinal analysis across four open-source model families at two scales, systematically tracing the co-evolution of generation and verification abilities through three phases: knowledge acquisition, continual learning, and knowledge updating. The findings reveal that verification consistently precedes and surpasses generation in performance; after knowledge updates, models often concurrently accept both old and new facts, resulting in a “multiverse of facts”; and even for high-frequency facts, persistent verification biases remain. These results uncover universal patterns underlying the GV-gap, offering theoretical foundations for improving factual consistency in language models.

Abstract

Language models are becoming the default interface to factual knowledge, yet they often verify outputs more reliably than they generate them. This generation-verification gap (GV-gap) underlies many recent advances in self-improvement and reasoning, but its dynamics on factual knowledge specifically remain poorly understood. We focus on the training mechanisms underlying factual GV-gaps, distinguishing them from their computational and aesthetic counterparts. We trace generation and verification capabilities through three training phases (acquisition, continual learning, and updating) across four open-source model families at two scales each. Three findings recur across models: (i) verification is consistently learned before generation; (ii) verification is more robust to continual learning than generation; and (iii) factual updates can leave models in a "multi-verse" state, simultaneously verifying both old and new answers as correct. Natural experiments on frontier models reproduce these dynamics at scale and reveal residual verification biases on well-covered facts.

Problem

• generation-verification gap
• factual knowledge
• language models
• training dynamics
• model updating

Innovation

• generation-verification gap
• factual knowledge
• continual learning
• model updating
• verification bias