Effect of stopping generation at end-of-sequence tokens

Determine the effect of stopping generation at the first end-of-sequence token on identity-aware gray-box hallucination detection performance, rather than continuing generation to the token cap after an end token is emitted.

Background

The open-weight experiments generated exactly 100 tokens because end-of-sequence stopping was disabled, even when the models emitted end tokens earlier. Consequently, detector inputs included continuations after the first end token, and the automatic factuality labels were computed over those complete extended responses. The paper explicitly leaves unresolved whether terminating generation at the first end token would change detection behavior or performance.

References

In the open-weight experiments, detector inputs and labeling included continuation after end tokens; the effect of stopping at the first end token remains unevaluated (Appendix~\ref{app:prompts-generation}).

— Gray-Box Hallucination Detection in LLMs with Token Identities  (2610.19996 - Goldshtein et al., 14 Oct 2026) in Section 6, Discussion and Limitations; Appendix, Section A.2, Prompts and Generation