Fabricator or dynamic translator?

arXiv cs.CL / 4/17/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper investigates how LLMs can perform machine translation, but highlights that their generative behavior can sometimes lead to different kinds of overgeneration beyond typical NMT artifacts.
These overgenerations can range from harmless self-explanations and accurate clarifications to risky or misleading confabulations.
The authors argue that identifying the exact nature of each overgeneration is difficult, because the outputs may be appropriate in some contexts and problematic in others.
They describe multiple strategies tested in a commercial setting and report results aimed at distinguishing and managing these behaviors for better audience comprehension.
Overall, the work frames LLM translation as either “fabrication” or “dynamic translation,” depending on whether the extra content is warranted and safe.

Abstract

LLMs are proving to be adept at machine translation although due to their generative nature they may at times overgenerate in various ways. These overgenerations are different from the neurobabble seen in NMT and range from LLM self-explanations, to risky confabulations, to appropriate explanations, where the LLM is able to act as a human translator would, enabling greater comprehension for the target audience. Detecting and determining the exact nature of the overgenerations is a challenging task. We detail different strategies we have explored for our work in a commercial setting, and present our results.