New Client Exits mechanism aims to increase LLM transparency
Researchers have introduced Client Exits, a method to separate AI semantic intent from presentation. Learn how this could improve multimodal responses.
Source: Grafy Research LabLarge language models (LLMs) often struggle to communicate complex information in ways that are easily inspectable or visually rich without being tied to a specific interface. A new research paper introduces "Client Exits," a lightweight mechanism designed to decouple an AI's semantic intent from how that information is actually presented to a user.
What are Client Exits?
According to the research paper "Client Exits - A Way to Increase LLM Transparency on the Client Side," the core concept involves the model inserting predefined, machine-interpretable markers into its textual output. Instead of the model trying to dictate exactly how a response should look—such as generating complex HTML or specific UI layouts—it simply signals the presence of specific types of information.
The receiving client (the application or interface the user is interacting with) then takes these markers and decides how to realize them. This separation allows the model to focus on the underlying meaning while the client handles the presentation layer, whether that be through text, graphics, speech, or other modalities.
How does this improve transparency?
One of the primary challenges with current LLM outputs is that the "reasoning" or the supplementary data is often buried within a wall of text, making it difficult for users or automated systems to inspect the model's underlying logic or the specific data it is referencing.
By using Client Exits, the semantic intent is explicitly marked. This allows the client-side application to pull out specific information and present it in a way that is more transparent to the user. For example, if a model suggests a specific scientific concept, the marker could trigger the client to display a diagram or a citation, making it clear exactly what the model is attempting to communicate without the model needing to "draw" the diagram itself.
What is the prototype implementation?
The researchers implemented a prototype called "Marc Jansen" to demonstrate the feasibility of this approach. This prototype serves as a proof of concept for how these markers can be used to enrich LLM responses in real-time. By testing the mechanism, the authors aim to show that separating the "what" (the semantic content) from the "how" (the visual or auditory presentation) creates a more flexible and robust interaction model for multimodal AI.
Why does this matter for multimodal AI?
As AI moves toward becoming truly multimodal, the ability to coordinate between text, image, video, and audio becomes critical. Currently, much of this coordination is handled by complex, end-to-end models that can be "black boxes," where it is difficult to tell why a model chose a specific visual representation for a piece of text.
Client Exits offer a way to standardize how models signal for multimodal assistance. If a model can signal "[IMAGE_REQUEST: nebula]" via a marker, the client can use a dedicated image generation engine to fulfill that request. This modularity ensures that the model remains a generator of intent, while specialized tools handle the heavy lifting of media creation, leading to more predictable and inspectable workflows.
How to explore these concepts in Grafy
While Client Exits is a research-level concept, the idea of separating AI intent from specialized output is central to how multimodal workspaces function. For researchers looking to dive deeper into the technical details of this paper, you can use Grafy Reader to upload the PDF and ask specific questions about the implementation of the Marc Jansen prototype or the mathematical implications of the markers.
For those interested in seeing how different AI models handle complex, multi-step reasoning and different output formats, Grafy Chat allows you to compare various models side-by-side to observe how they structure their responses.
Frequently asked questions
What is the primary goal of Client Exits?
The primary goal is to increase transparency in LLM responses by separating the semantic intent of the model from the specific way that information is presented to the user. This allows the model to use machine-interpretable markers that a client-side application can then turn into various formats like text, speech, or graphics.
Does the model decide how the information looks?
No, the model does not decide the final visual or auditory presentation. Instead, it inserts predefined markers into its text output that signal specific information is present. The receiving client application is responsible for deciding whether and how to transform those markers into a specific interface or modality.
How does the Marc Jansen prototype work?
The Marc Jansen prototype is an implementation used by the researchers to demonstrate the Client Exits mechanism. It acts as a proof of concept, showing how a model can successfully insert markers into its output and how a client can interpret those markers to enrich the user experience with supplementary information.
Source: Grafy Research Lab — Client Exits - A Way to Increase LLM Transparency on the Client Side


