Skip to content
AI Knowledge MapSearch / Text directory

Which fragments, typos or speech-transcription errors change the task's meaning?

A019 · A Expressing intent · Sufficient description and language form

Key answer

Prioritize errors changing the object, action, negation, quantity or condition. A fragment may be clear, while a complete sentence may specify the wrong task.

Full explanation

PromptRobust generated 4,788 adversarial prompts with character-, word-, sentence- and semantic-level perturbations across eight task types and 13 datasets, finding insufficient robustness. It deliberately sought failing inputs; its declines cannot be treated as the normal effect of one accidental typo. Study

The 2026 MulTypo study simulated multilingual typing errors on 18 open models, observing declines varying by reasoning/generation task and language. Its keyboard errors mainly cover alphabetic scripts, not Chinese pinyin input or Chinese speech transcription. It does not provide a formula linking Chinese typo rates and answer accuracy. Paper

Check first:

· Whether actions or negations such as “needed/not needed” and “keep/delete” were transcribed correctly.

· Whether numbers and units changed, such as “15/50” or “milligrams/grams.”

· Whether names, model numbers, filenames and variables now identify a different object.

· Whether “only if,” “unless” or “at least” was omitted.

“This passage, shorter, keep the numbers” can be meaningful when its object is known. Conversely, transcribing “do not delete” as “delete” changes the task even with correct grammar.

This is speech-transcribed text. Flag contradictions in amounts, dates, names and negations rather than guessing a version. Execute after confirmation.

This establishes a checking requirement; it cannot recover words you never supplied and the context cannot determine.

References

1. Zhu et al. (2023; revised 2024), PromptRobust: Towards Evaluating the Robustness of Large Language Models on Adversarial Prompts, technical report

2. Zhao et al. (2026), Evaluating Robustness of Large Language Models Against Multilingual Typographical Errors, ACL