Trustworthy clinical AI requires that performance gains reflect genuine evidence integration rather than surface-level artifacts, according to a study published on arXiv by researchers Vu, Doan Nam Long, and Balloccu. Their paper, "The Scaffold Effect: How Prompt Framing Drives Apparent Multimodal Gains in Clinical VLM Evaluation," evaluates 12 open-weight vision-language models (VLMs) on binary classification across two clinical neuroimaging cohorts: FOR2107 (affective disorders) and OASIS-3 (cognitive decline). Both datasets come with structural MRI data that carries no reliable individual-level diagnostic signal, providing a controlled test of multimodal reasoning.
The Study: Measuring Genuine Multimodal Reasoning
Under these conditions, smaller VLMs exhibited gains of up to 58% F1 upon introduction of neuroimaging context, with distilled models becoming competitive with counterparts an order of magnitude larger. However, a contrastive confidence analysis revealed that merely mentioning MRI availability in the task prompt accounts for 70-80% of this shift, independent of whether imaging data is present. The researchers term this domain-specific instance of modality collapse the scaffold effect. The table below summarizes key findings:
| Condition | Performance Impact | Underlying Cause |
|---|---|---|
| Prompt mentions MRI (with or without image) | 70-80% of F1 gain | Scaffold effect: prompt framing, not actual vision input |
| No MRI mention in prompt | Baseline performance | No prompt-driven bias |
| Small VLMs with neuroimaging context | Up to 58% F1 improvement | Apparent multimodal gain (largely scaffold effect) |
| Distilled vs. large models | Distilled competitive with 10x larger counterparts | Prompt framing equalizes performance |
Expert Evaluation Reveals Fabrication
Expert evaluation disclosed that VLMs fabricated neuroimaging-grounded justifications across all conditions. Additionally, preference alignment—a common technique to reduce unwanted behaviors—while eliminating MRI-referencing behavior, collapsed both conditions toward random baseline. This indicates that current evaluation methods are inadequate indicators of multimodal reasoning.
Implications for Enterprise AI
For technology leaders deploying AI in clinical or other high-stakes domains, the scaffold effect underscores the danger of surface evaluations. The findings demonstrate that performance gains attributed to multimodal fusion may instead stem from prompt engineering artifacts. As the researchers state, "surface evaluations are inadequate indicators of multimodal reasoning, with direct implications for the deployment of VLMs in clinical settings." The study serves as a cautionary tale for any enterprise relying on VLM benchmarks without probing the underlying reasoning mechanisms.