Multimodal LLMs fail majority of autonomous aerial missions in new benchmark

The zero-shot constraint is the actual finding. MissionBench doesn't allow domain-specific fine-tuning, which means the 65% failure rate reflects a reasoning architecture problem, not a training data problem. That distinction matters for what fixes are actually possible.

This joins a cluster of high-stakes domain benchmarks published this week. DBA-Bench (PostgreSQL operations) and the NRC reactor licensing study both target safety-critical infrastructure where agent hallucination or incomplete reasoning directly breaks production systems. MissionBench extends that pattern to autonomous aerial navigation, another domain where incomplete multi-step planning creates real consequences. The common thread: benchmarks are shifting from isolated task performance to embodied, sequential decision-making under uncertainty, where current MLLMs hit a wall regardless of scale.

If researchers can close the human-model gap on MissionBench by adding architectural changes (like explicit spatial tokenization or planning modules) without fine-tuning, that confirms the problem is reasoning structure, not data. If the gap only closes with domain-specific training, the benchmark becomes a tool for measuring fine-tuning efficiency rather than a fundamental capability gap, which changes what the result actually tells us about MLLM readiness for embodied AI.