Rephrase a problem. Reorder the facts. Switch from academic to business context. The answer should stay the same — the question hasn't changed. But for most large language models, it doesn't.
De Zarzà and colleagues apply eight semantic-preserving transformations to 19 multi-step reasoning problems across seven foundation models. The transformations are carefully designed to change nothing about the problem's content: paraphrase, fact reordering, expansion, contraction, register shift. The same question, asked differently.
The result inverts the scaling hypothesis: the smallest model tested — Qwen3-30B-A3B — achieves the highest stability at 79.6% invariant responses and 0.91 semantic similarity. Larger models exhibit greater fragility. Scale buys capability but sells robustness.
The mechanism is suggestive. Larger models have richer internal representations, which means more surface area for input variations to perturb. A paraphrase that leaves a small model's representation unchanged may activate a different cluster of representations in a larger model — not because the larger model misunderstands, but because it has more ways to understand. The additional capacity creates additional sensitivity.
This has a direct design implication for agentic AI systems. The industry assumption is that more capable models make better agents. But an agent that gives different answers to the same question asked differently is unreliable regardless of how capable it is on any single formulation. Stability and capability are different axes. Optimizing for one doesn't optimize for the other. The agent that doesn't change its mind when you rephrase the question may be more useful than the agent that answers more impressively.