AI systems analyzing medical images have historically struggled with hallucinations—generating plausible but incorrect morphological descriptions that can lead to diagnostic errors. PathoSage, a new framework developed by computational pathology researchers, directly addresses this by implementing an experience-aware agentic workflow that adjudicates between multiple evidence sources before rendering final assessments. Rather than relying on end-to-end multimodal large language models (MLLMs) that process entire pathology slides at once, PathoSage breaks analysis into patch-level reasoning tasks, where individual image regions are evaluated against documented evidence chains. This approach significantly reduces the confidence hallucinations that plague current pathology AI systems, which often misidentify cellular structures or tissue morphology due to insufficient contextual grounding.
Parallel to advances in medical imaging, OmniMem tackles a different computational bottleneck: enabling long-form video understanding on devices with limited memory. Developed for audio-visual LLMs, OmniMem implements perturbation-aware memory compression that controls the linear growth of video tokens and key-value caches that typically exhaust GPU memory during extended video analysis. By selectively compressing the KV cache—the memory structure storing attention information—OmniMem maintains reasoning quality while reducing memory overhead by up to 60 percent, allowing real-time analysis of multi-hour video streams. This efficiency gain opens deployment possibilities for video-based AI agents in resource-constrained environments, from edge devices to cost-sensitive cloud infrastructure.
Both systems exemplify a broader research trend: replacing monolithic AI models with structured agentic workflows that decompose complex tasks into verifiable subtasks. PathoSage's deployment pathway focuses on integration with existing digital pathology platforms in clinical labs, where its evidence adjudication approach can be validated against curated pathology datasets and expert annotations. OmniMem's memory efficiency enables immediate application in video surveillance, autonomous vehicle systems, and real-time medical monitoring where continuous processing currently requires prohibitive computational resources. These frameworks move beyond incremental model scaling toward architectural innovations that make AI systems more reliable, interpretable, and deployable in high-stakes domains.