Abstract
Current conversational AI systems rely on monotonic streaming, requiring users to wait for complete responses before correcting misinterpretations. This disclosure describes a bifurcated dual-dispatch architecture that steers generative AI outputs mid-stream. By pairing a localized Small Language Model (SLM) shadow parser with a high-capacity frontier LLM, the system dynamically analyzes real-time client-side behavioral telemetry - including keystroke dwell times, backspace mutation ratios, and paste metadata to predict user intent and cognitive friction. When volatile semantic nodes or potential intent mismatches are detected, the SLM executes pre-emptive mitigations by tuning inference hyperparameters (such as temperature) or surfacing non-blocking, inline "intent chips" anchored directly to ambiguous output tokens. User interaction with an intent modifier triggers a localized Abstract Syntax Tree (AST) rollback and semantic client-side diffing, enabling the primary model to pivot structural formats mid-stream (e.g., morphing a Markdown table into JSON) without clearing the canvas or disrupting saccadic reading flow. For data governance, the client-side SLM functions as an on-device semantic firewall, performing Named Entity Recognition to cryptographically mask sensitive strings prior to frontier LLM dispatch. The architecture extends across voice-first systems via acoustic bridging, spatial gaze tracking in mixed reality, and coordinate-locked localized inpainting with latent-space freezing in generative media.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Mathur, Rachit and Shah, Payal, "Steering Generative AI Responses by Predicting User Intentions Using SLMs and Behavioral Telemetry", Technical Disclosure Commons, (September 16, 2026)
https://www.tdcommons.org/dpubs_series/11734