The Participation Problem: Why Hallucination Persists When Language Models Are Required to Always Generate Output

Hallucination in large language models (LLMs) is commonly addressed through improved grounding, retrieval, and verification mechanisms. While these approaches reduce some classes of factual error, hallucination persists in extended, interactive use. Prior work has shown that hallucination often appears late in an interaction, following a period of gradual degradation in reasoning integrity. 

This paper argues that a structural requirement contributes to this persistence: the absence of a firstclass abstention or halt state. Autoregressive systems are optimized to continue generating output given prior context. As a result, they continue even as uncertainty rises or internal reasoning integrity weakens, transforming degradation into fluent but weakly supported output.

This paper presents an architectural and evaluative analysis of this participation requirement and explains how it transforms upstream degradation into downstream hallucination. The contribution is explanatory rather than prescriptive. No new model, benchmark, or algorithm is proposed. The goal is to clarify why hallucination remains difficult to eliminate despite improvements in grounding, and why current evaluation practices struggle to detect the conditions under which a system should not continue
generating output.