AWS and Artificial Genius show a way to cut LLM hallucinations in finance and healthcare

AWS, together with its partner Artificial Genius, has shown how to adapt large language models for tasks where an error is unacceptable. The approach is built on Amazon Nova and SageMaker, but the key idea is not a new model size. It is to use the model's language understanding without letting it freely generate the answer.

Why this is a problem

For financial services, healthcare, insurance and legal processes, ordinary LLMs still look risky. They write, summarize and explain well, but by nature they remain probabilistic systems: the model predicts the next token rather than extracting a guaranteed-correct fact. This produces hallucinations: answers that sound convincing but are not supported by the source data. Where audit, reproducibility and accountability matter, this mode of operation fits poorly with production.

The authors of the article suggest viewing the evolution of AI in three steps. The first wave was built on symbolic logic and rigid rules: those systems were deterministic but too inflexible. The second wave, which includes modern transformers, brought a huge leap in fluency and language understanding, but also brought unpredictability. Artificial Genius calls its approach the third generation: the model still understands natural language like a modern LLM, but the final answer passes through deterministic logic and must not go beyond what is actually present in the input context.

How the setup works

The central thesis of AWS and Artificial Genius is this: a generative model can be used in a strictly non-generative way. That is, it does not "fill in" an answer based on next-token probability. It checks whether the answer can be extracted from the document, and if it cannot, it refuses to answer. This mode is especially useful for questions about dates, amounts, names, excerpts from a report, or confirmation of a specific fact.

The article puts it very directly:

"If the question cannot be answered from the document, the model must

answer: \"Unknown\"."

• Amazon Nova Lite was chosen as the base model because it suits short, precise answers without excess chattiness;
• fine-tuning runs in SageMaker via supervised fine-tuning, so that the model follows a single system rule: invent nothing;
• training uses a synthetic set of questions and answers that includes both answerable and deliberately unanswerable queries;
• instead of classic RAG, which is still generative, the emphasis is on a tighter link between the document text and the specific question;
• on top of this sits an agent platform where a free-form request can be translated into a stricter specification, and the only manual check remains at the stage of that translation.

An important detail: the authors explicitly contrast their method with the familiar advice to "set temperature to zero." In their view, this does not solve the root problem, because the model keeps generating anyway. In their variant, what changes is not just the degree of randomness but the very logic of how the model is used: probabilistic understanding is kept at the input, while at the output the system aims for a binary mode: answer only with what the text confirms, or honestly say there is no answer.

What the tests showed

Technically, the setup looks fairly down-to-earth, which is what makes it interesting. Training data is stored in Amazon S3, fine-tuning of the base Nova model runs in SageMaker Training Jobs, and then the custom version is imported into Amazon Bedrock and served to the application through the standard inference path. For enterprise teams this matters not only for convenience but also for a transparent data lineage: it is easier to see what the model was trained on, where it was modified and how it was later deployed to production.

The team also disclosed several engineering findings. For fine-tuning they used LoRA so as not to break the model's base language understanding. In earlier experiments on a different model, they even had to forcibly switch off chain-of-thought via the service token `</think>`, because extended reasoning got in the way of short deterministic answers. For the Nova Lite version, the authors combined LoRA dropout at 50%, manual early stopping and an expansion of the synthetic dataset to 30 thousand examples. According to them, this cut the hallucination rate from percents and fractions of a percent in early configurations to 0.03% in the best variant.

What this means

The story matters not only for AWS users. It points to a broader shift: the market is starting to look not simply for the "smartest" LLMs but for models with engineered behavioral boundaries. For banks, insurers, clinics and legal tech, it is a signal that AI adoption will increasingly be built around verifiability, refusal to answer and controlled workflows, rather than around impressive generation at any cost.