Therapeutic LLMs face steep energy cost for clinical safety gains

The paper quantifies directionality and magnitude of the safety-efficiency trade-off, but doesn't claim it's avoidable. The real finding is that practitioners can no longer assume safety improvements come 'for free' at scale; they must now budget for the cost.

This directly constrains the deployment strategies outlined in the SLM trustworthiness benchmarking work from August 12. That paper examined whether compressed models retain safety properties during edge deployment; this new work suggests that even if compression preserves safety, the baseline safety-efficiency frontier itself is steep. The tension also echoes the ToolHazard and LODESTAR papers, which both assume safety evaluation happens before production, but don't address the operational cost of implementing those safeguards. For healthcare specifically, the VITA corpus-specific RAG system demonstrated that specialized, lower-compute approaches can match frontier models on clinical tasks, suggesting one workaround: domain-specific systems may sidestep the exponential cost curve entirely by avoiding the need for maximum safety margins in the first place.

If a healthcare vendor deploys a therapeutic LLM with measurably lower safety scores than competitors but documents 40+ percent lower energy per token in production, that signals practitioners are accepting the trade-off explicitly. Conversely, if major EHR vendors (Epic, Cerner) announce therapeutic LLM integrations in the next 12 months without disclosing energy budgets, that indicates the market is ignoring the constraint rather than pricing it into procurement decisions.