Knowledge-Graph-Gated Defactualization for Style-Controllable and Fact-Preserving Generation in Agentic Conversational AI
This paper introduces Defactualize-Steer-Rehydrate (DSR), a fine-tuning-free framework that combines activation steering with a salience-weighted knowledge graph to enable style-controllable generation in agentic LLMs while systematically preserving factual accuracy through entity extraction, placeholder substitution, and deterministic rehydration.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of artificial intelligence, large language models have become powerful tools for conversation, capable of writing emails, answering questions, and offering advice. However, when these systems are deployed in real-world settings like customer support, they face a difficult balancing act. They must adopt a specific tone, such as being empathetic to a frustrated caller or formal during a serious complaint, while simultaneously holding onto exact, verifiable facts like order numbers, product names, and customer identities. If the model gets too caught up in the desired tone, it might accidentally change the facts, a problem researchers call semantic leakage. For instance, a model trying to sound very sympathetic might inadvertently swap a customer's order number for a different one, or invent a product that was never mentioned. This creates a risk where the AI sounds helpful but provides incorrect information. To solve this, scientists have developed methods to steer the model's behavior without retraining it, essentially nudging its internal thought process toward a specific style. Yet, these nudges often blur the line between style and substance, making it hard to keep the facts pure while changing the mood.
A team of researchers at the Indian Institute of Technology Bhilai has introduced a new framework called Defactualize-Steer-Rehydrate, or DSR, which treats this problem as a matter of careful information management rather than just mathematical adjustment. Instead of trying to teach the model to be perfect, the system acts as a gatekeeper that separates the facts from the feelings before the model even begins to speak. The process begins by analyzing a customer's message and extracting key pieces of information, such as names and order details, into a structured list. The system then replaces these specific facts with generic placeholders, like "customer name" or "order ID," effectively scrubbing the real data from the text. This cleaned-up version is then fed to the artificial intelligence, which is instructed to generate a response in a specific style, such as being warm and understanding or cold and professional. Because the real facts have been removed, the model cannot accidentally alter them; it can only focus on the tone and the flow of the conversation.
Once the model generates its response, the system performs the final, crucial step: rehydration. It looks back at the original list of facts it extracted and deterministically swaps the generic placeholders back with the exact, verified values from the customer's message. This means the final output has the desired emotional tone but contains the precise facts the customer provided, with no room for the model to hallucinate or change them. The researchers tested this approach on six different versions of a popular AI model family, ranging from smaller to larger sizes, using six hundred simulated customer support cases. They found that this method significantly improved the rate at which correct facts were preserved in the final response compared to using style-steering alone. While the improvement was statistically clear, the absolute number of facts recovered was still modest, suggesting that the system helps but does not yet guarantee perfect accuracy in every single instance. Crucially, the study showed that this process did not degrade the quality of the style; the responses remained just as empathetic or formal as intended, proving that the two goals could be pursued independently without one ruining the other.
The research also revealed that the success of this method does not depend heavily on the size of the AI model. Whether the system used a smaller model with one billion parameters or a larger one with thirteen billion, the ability to preserve facts remained consistent, and the style control worked equally well across the board. This suggests that the value comes from the structured process of separating and recombining information, rather than from the raw power of the model itself. The team also discovered that the system is robust against errors; in over a thousand generated responses, the mechanism never failed to restore the facts, and it never left a placeholder token visible in the final text. This reliability indicates that the method is stable enough for practical use, even if it does not solve every possible error a model might make.
One of the most interesting findings from the study concerns the relationship between how strongly the system pushes for a specific style and how often it makes mistakes. The researchers observed that for some models, increasing the strength of the style instruction did not lead to a steady increase in errors. Instead, there was a specific range where the system could become more expressive in its tone without immediately sacrificing its grip on the facts. This discovery points to a previously unknown operating zone where style and accuracy can coexist more comfortably than previously thought. By mapping out these interactions, the researchers provided a clearer picture of how to tune these systems for real-world use.
The work demonstrates that we do not need to fundamentally retrain massive AI models to make them more reliable. Instead, by adding a layer of structured knowledge engineering that acts as a filter and a restorer, we can guide these systems to be both polite and precise. The approach relies on a simple but powerful idea: if you separate the facts from the feelings before the thinking begins, you can control the feelings without losing the facts. This method offers a path forward for building AI assistants that are not only capable of understanding human emotion but are also trustworthy enough to handle the critical details of our daily lives. The researchers have made their code and data available to the public, allowing others to verify these results and build upon this new way of thinking about how machines can talk to us.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.