Foundation Models Do Not Understand Biology

2026-05-20 · Source: Towards AI - Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Internet of Things (IoT) & Connected Devices · Depth: Advanced, long

Summary

Foundation models, particularly multimodal LLMs, lack true biological understanding, leading to dangerous "clinical hallucinations" in medical diagnostics. These models generate text based on word prediction, not biological reasoning, which is especially risky in microscopy where tiny, specific features are critical. The article introduces NTD-Assist, a hybrid system designed to diagnose Neglected Tropical Diseases. It combines a multimodal core LLM (like MedGemma 4B or Qwen2-VL-2B) with deterministic morphology guardrails to validate biological plausibility before any output reaches a clinician. This prevents errors such as identifying blood-borne parasites in skin biopsies or incorrectly sized intracellular structures. The system also addresses edge deployment challenges through 4-bit NF4 quantization for memory efficiency and runtime model routing for hardware adaptability, ensuring functionality in resource-constrained environments.

Key takeaway

For AI Engineers developing clinical diagnostic tools, you must integrate deterministic biological guardrails into your pipelines. Relying solely on foundation models for medical reasoning risks generating confident, yet biologically impossible, "hallucinations" that can harm patients. Implement a hybrid architecture, like NTD-Assist, to validate model outputs against established medical guidelines, ensuring safety and reliability in resource-constrained edge environments. This approach shifts the LLM's role from unconstrained authority to a validated aggregator.

Key insights

Foundation models lack biological understanding, necessitating guardrails for safe clinical AI applications.

Principles

• LLMs predict tokens, not biological truth.
• Clinical hallucinations are not harmless.
• Decouple LLM synthesis from biological validation.

Method

NTD-Assist uses a hybrid architecture: a multimodal LLM generates structured JSON, which is then intercepted by a deterministic Guardrails Engine. This engine applies medical guidelines to validate biological plausibility before output.

In practice

• Implement morphology guardrails for diagnostics.
• Use 4-bit NF4 quantization for edge deployment.
• Pre-cache model files for offline functionality.

Topics

• Clinical AI Safety
• Medical Hallucinations
• Multimodal LLMs
• NTD-Assist
• Edge AI Deployment
• Biological Guardrails
• Neglected Tropical Diseases

Best for: Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• Can AI Chatbots Reason Like Doctors? — Large language models demonstrate potential in clinical reasoning, but require careful integration and standardized evaluation in medical workflows.
• MedBench v5: A Dynamic, Process-Oriented, and Hallucination-Aware Benchmark for Clinical Multimodal Models — MedBench v5 evaluates clinical AI models dynamically, focusing on process stability and hallucination propagation under various information stressors.
• MedPanel AI - Multi-Agent Clinical AI — Multi-agent systems, particularly with adversarial agents, enhance diagnostic safety by simulating human clinical reasoning and challenging initial conclusions.
• CuraView: A Multi-Agent Framework for Medical Hallucination Detection with GraphRAG-Enhanced Knowledge Verification — CuraView uses GraphRAG and multi-agent verification to detect and explain medical LLM hallucinations, improving factual reliability.
• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.
• Spoon-feeding a Monster — Autonomous security testing requires separating LLM reasoning from deterministic execution to ensure ground truth and prevent hallucinations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards AI - Medium.