Vaincre les hallucinations des LLM dans les secteurs réglementés : les modèles déterministes d'Artificial Genius sur Amazon Nova

This post is cowritten by Paul Burchard and Igor Halperin from Artificial Genius. The proliferation of large language models (LLMs) presents a significant paradox for highly regulated industries like financial services and healthcare. The ability of these models to process complex, unstructured information offers transformative potential for analytics, compliance, and risk management. However, their inherent probabilistic nature leads to hallucinations , plausible but factually incorrect information. In sectors governed by stringent requirements for auditability and accuracy, the non-deterministic behavior of standard generative AI is a barrier to adoption in mission-critical systems. For a bank or a hospital, determinism isn’t only a goal; the outcomes must be accurate, relevant, and reproducible. In this post, we’re excited to showcase how AWS ISV Partner Artificial Genius is using Amazon SageMaker AI and Amazon Nova to solve this challenge. By introducing a third generation of language models, they’re delivering a solution that is probabilistic on input but deterministic on output, helping to enable safe, enterprise-grade adoption. To understand the solution, let’s look at how AI has evolved: First generation (1950s): Researchers used symbolic logic to build deterministic, rule-based models. While safe, these models lacked fluency and could not scale. Second generation (1980s–present): The shift to probabilistic models (culminating in the Transformer architecture) unlocked incredible fluency. However, because these models predict the next token based on probability, they suffer from unbounded failure modes (hallucinations) that are difficult to engineer away. Third generation (the Artificial Genius approach): Rather than a new generation that replaces the old, we’re moving from the rigidity of symbolic logic and the unpredictability of probabilistic models toward a hybrid architecture. This approach uses the generative power of Amazon Nova to understand context but applies a deterministic layer to verify and produce output. It’s the convergence of fluency and factuality. The solution: A paradoxical approach to generation It’s mathematically difficult to prevent standard generative models from hallucinating because the extrapolative, generative process itself causes errors. Artificial Genius addresses this by using the model strictly non-generatively. In this paradigm, the vast probability information learned by the model is used only interpolatively on the input. This allows the model to comprehend the innumerable ways a piece of information or a question can be expressed without relying on probability to generate the answer. To create this third-generation capability, Artificial Genius uses SageMaker AI to perform a specific form of instruction tuning on Amazon Nova base models. This patented method effectively removes the output probabilities. While standard solutions attempt to ensure determinism by lowering the temperature to zero (which often fails to address the core hallucination issue), Artificial Genius post-trains the model to tilt log-probabilities of next-token predictions toward absolute ones or zeros. This fine-tuning forces the model to follow a single system instruction: don’t make up answers that don’t exist. This creates a mathematical loophole where the model retains its genius-level understanding of data but operates with the safety profile required for finance and healthcare. Going beyond RAG Retrieval Augmented Generation (RAG) is frequently cited as the solution to accuracy, but it remains a generative process and creates fixed vector embeddings that might not be relevant to subsequent queries. The third-generation approach improves upon RAG by effectively embedding the input text and the user query into a unified embedding. This helps ensure that the data processing is inherently relevant to the specific question asked, delivering higher fidelity and relevance than standard vector retrieval methods. Delivering value using agentic workflows To help enterprises maximize the value of their unstructured data, Artificial Genius packages this model into an industry-standard agentic client-server platform, available through AWS Marketplace . Unlike second-generation agents, which risk compounding errors when strung together in workflows, the inherent reliability of this third-generation model allows for complex, high-fidelity automation. The prompts used to create these workflows follow the structure of a product requirements document (PRD). Through this structure, domain experts—who might not be AI engineers—can formulate queries in natural language while maintaining strict control over the output. The product additionally offers free-form prompting of the workflow specification. For this purpose, the Amazon Nova Premier model, which is especially capable of translating free-form prompts into PRD format, is used. Although Nova Premier is a generative model, which requires a