Logical Judgments Under Pressure: Diagnosing Syllogistic Stability with Learned Soft Prefixes

En 30 secondes

Ce qui s'est passé

Researchers used learned soft prefixes to override correct logical reasoning in large language models, causing 37 to 99 percentage point increases in wrong answers on syllogistic tasks.

Pourquoi ça compte

Engineers building or deploying LLMs for logical reasoning, formal verification, or safety-critical applications need to understand this vulnerability.

Vigilance

The study tested only three models and used a specific syllogistic benchmark; results may not generalize to other reasoning tasks or model architectures.