LLM Agents Exhibit Semantic Fragility Across Variations

A March 13, 2026 arXiv preprint by J. De Curtò presents a metamorphic testing framework assessing robustness of LLM reasoning agents under eight semantic-preserving transformations across seven foundation models from four architectural families. The authors evaluate 19 multi-step reasoning problems in eight scientific domains and find that model scale does not predict robustness: Qwen3-30B achieves 79.6% invariant responses with semantic similarity 0.91, while larger models exhibit greater fragility.