Banning filler words improves AI reasoning more than removing grammatical structures

Researchers tested whether removing specific words or grammatical patterns from prompts makes AI systems reason better. It turns out that the shallowest constraint — banning meaningless filler words like 'very' and 'just' — produced the largest improvement in reasoning accuracy, while removing more linguistically complex elements produced smaller gains. This suggests AI reasoning breaks down because of fluent but sloppy default patterns, not because of logical structure problems.

For years, linguists and AI researchers assumed that reasoning failures in language models were tied to deep linguistic structures — that removing certain verbs or grammatical features would force the model to think more carefully. This paper disconfirms that entire theory. What actually works is forcing the model off its autopilot: any constraint that makes it slow down and reconsider helps, and the dumber the constraint, the better. That's a different problem to solve than anyone thought it was.