AI reasoning benchmarks are easier to game than they look.
Researchers introduced LGMT, a testing framework that evaluates large language models not on whether they get answers right, but on whether they get them right consistently. The approach draws on first-order logic to generate semantically equivalent variations of the same problem — different surface forms, same underlying truth. Running six state-of-the-art LLMs through this gauntlet revealed defects that standard reference-based benchmarks missed entirely. Models proved especially brittle when symbols or conclusions were reworded, even when the logical content stayed identical.
The finding matters because benchmark scores drive purchasing decisions, research funding, and deployment confidence. If a model aces a test only because it pattern-matches the exact phrasing, that score is measuring memorization as much as reasoning — a gap that matters most in high-stakes applications like legal analysis, medical triage, or code verification. LGMT's oracle-free design also means it doesn't require human-labeled ground-truth answers, making it cheaper to scale than conventional evaluation pipelines.
Advanced prompting techniques like few-shot chain-of-thought helped somewhat but didn't close the gap — which is the kind of result that should make any vendor's benchmark-touting press release worth reading twice.