AI agents arguing over an image now get to pick their own zoom level before they start fighting.
A new paper introduces DREAM, a framework for multimodal multi-agent debate, where several AI models look at the same image and argue toward a consensus answer. Today's systems force every agent to study one fixed resolution, even though some questions need a close-up and others need the full frame. DREAM adds a quick probe round where each agent tries multiple resolutions, measures its own uncertainty with a metric called Average Normalized Log-Likelihood, and settles on whichever resolution it handles best. A second piece, Uncertainty-Guided Rollback Aggregation, tracks each agent's confidence across rounds and restores an early correct answer if an agent later caved to a confident-sounding but wrong peer.
That second fix targets a real failure mode in these systems: agents that had the right answer talking themselves out of it because another agent sounded sure. Multi-agent debate was supposed to make models more reliable by making them check each other's work, so an architecture that explicitly detects and reverses peer-pressure flip-flops is a useful admission that debate alone doesn't fix hallucination, it can also spread it.
Across six multimodal datasets, DREAM improved the accuracy-token trade-off by 1.5 to 3.2 percentage points over existing debate baselines, with no per-dataset tuning. That's a real but modest gain, and it comes with the usual cost of multi-agent setups: more inference passes per answer. Worth watching, not worth a press release.