AI
LLMs misinterpret verb tenses through false completion bias
New research reveals that large language models systematically misconstrue the logic of progressive verb forms, confusing ongoing actions with completed goals.
A new study probes a fundamental weakness in how leading open-weight language models handle the logic of English grammar. Rather than understanding the compositional meaning of verb tenses, these systems appear to rely on learned patterns that frequently mislead them.
The Imperfective Paradox#
The research focuses on the Imperfective Paradox—a linguistic distinction between two types of actions. In activities like running, a statement such as "she was running" logically guarantees that "she ran" is true. But with goal-oriented accomplishments like building, "she was building a house" does not necessarily mean "she built a house"—the project may have gone unfinished.
This subtle but crucial difference in how language encodes meaning separates actions that inherently conclude from those that may be interrupted or abandoned.
The Teleological Bias Problem#
Researchers created ImperfectiveNLI, a specialized dataset to test whether LLMs could distinguish these cases. What they found was systematic failure: models exhibited a "Teleological Bias," automatically hallucinating successful completion for goal-oriented events. The models would claim that "was building" entails "built," even when the text explicitly canceled the accomplishment—directly contradicting the logical structure of the language.
What Happens When Models Are Corrected#
Attempts to adjust model behavior through prompting only partially worked. More troublingly, corrective interventions introduced new errors: models began rejecting valid entailments for true activities, breaking cases that had previously worked. This "calibration crisis" shows that surface-level prompting cannot fix the underlying problem.
Why It Matters#
Deeper analysis of model internals revealed that while embeddings can distinguish progressive from simple past forms, the actual inference decisions are dominated by strong statistical priors about goal completion. The models function more like predictive narrative engines—guessing what happens next in a plausible story—rather than faithful logical reasoners.
Resolving this gap will require moving beyond prompting tricks toward structural alignment grounded in linguistic principles, rather than relying on learned probabilistic patterns alone.

