DPS (Dynamic Prompt Specialization via Semantic Micro-Model Routing)

Abstract

WHITE PAPER Dynamic Prompt Specialization via Semantic Micro-Model Routing STRATEGIC TECHNICAL MEMORANDUM: DPS Author: Valerii Khalif (VALEO), AVA LIVE Status: Draft for DOI / Zenodo submission Date: August 31, 2026 License: Open / free to use (attribution requested; voluntary donations welcome, not required) Abstract Modern development of large language models (LLMs) based on the Transformer architecture relies heavily on scaling: increasing parameter counts, training data volumes, and computational resources, alongside regular retraining and model updates. This document proposes an alternative, additive adaptation layer that does not require modifying the weights of the core LLM. The proposed approach is designated as Dynamic Prompt Specialization (DPS). DPS employs a small, high-speed auxiliary model—a semantic micro-model—that continuously analyzes the semantic and communicative structure of the current user input, extracts concise semantic fragments, and maps them to a predefined code table. Each code corresponds to a pre-engineered and validated prompt fragment, optionally accompanied by a confidence level or signal intensity score. Based on the active set of signals, DPS constructs a dynamic specialized context transmitted to the unmodified primary LLM. The primary goal of DPS is user interaction personalization: the system can adapt to stable individual communication patterns, ongoing conversational context, professional expertise level, task domain, feedback to previous answers, and other observable signals. Furthermore, DPS is not restricted to user communication. The same architecture can be applied to domain-specific pre-specialization of queries prior to reaching the primary LLM. This pre-structures multi-faceted queries and potentially reduces the necessity for dedicated specialist agents and complex orchestration pipelines. DPS operates without altering main model weights or requiring retraining. In the absence of a sufficiently confident signal, the system defaults to the baseline behavior of the main LLM. Consequently, DPS represents an orthogonal adaptation layer that can be integrated into existing LLM infrastructure without altering the underlying foundation model. 1. Problem Statement The proposed approach addresses two interconnected constraints in current LLM-based systems. 1.1. Scaling Does Not Solve Individual Specialization The main vector of foundation model advancement focuses on increasing: Parameter counts; Training data volume; Compute resources; Training duration and complexity; Model update frequency. However, expanding overall model capabilities does not yield a proportional enhancement in interaction quality for a specific individual user. In many practical scenarios, a user operates within a relatively bounded domain of knowledge and exhibits consistent individual patterns of task formulation, communication, verification, clarification, and correction. Thus, incremental general model knowledge often carries less practical utility than the system's ability to precisely align with a specific user's traits and the current interaction state. DPS shifts part of the specialization process from model parameter modification to dynamic context adaptation. 1.2. Static System Prompt Most LLM systems rely on a single primary system prompt or a narrow set of predefined instructions. As a result, interactions are typically treated as static: $$\text{System Instructions} \longrightarrow \text{User Request} \longrightarrow \text{Response}$$ DPS frames interaction as a dynamic process where conversational state evolves continuously: $$\text{Current User Signal} \longrightarrow \text{Analysis} \longrightarrow \text{Active Context Update} \longrightarrow \text{Response} \longrightarrow \text{New User Signal} \longrightarrow \text{State Update}$$ Thus, specialization occurs continuously rather than as a single static setup. 2. Core DPS Architecture DPS comprises three functionally separated layers, each holding distinct responsibilities. 2.1. Semantic Micro-Model – Fast Semantic Layer The first layer consists of a lightweight, ultra-fast auxiliary model. It is not designed for full reasoning or final output generation. Its scope is narrowly constrained to: Analyzing input utterances; Extracting short semantic fragments; Identifying corresponding semantic signals; Mapping signals to designated codes; Computing confidence or intensity levels for each signal. A typical semantic fragment may consist of a few words, though the architecture does not impose fixed-length constraints. Fragment size depends on input signal characteristics and implementation design. A single user turn can simultaneously trigger multiple independent codes. For instance, a query might concurrently express: Professional technical context; Financial domain context; Request for conciseness; Requirement for high precision; Emotional tone/signal; Affirmation of a previously established rule. Consequently, DPS is not forced to select a single category or agent, but instead builds a multidimensional set of co-active signals. 2.2. Code-Prompt Correspondence Table The second layer operates as a mapping repository: Code (e.g., PFA-0281) $\rightarrow$ Pre-engineered Prompt Fragment. A key principle of DPS is that prompt fragments are not automatically generated by the system on the fly. They are authored by engineers and undergo a controlled lifecycle: $$\text{Design} \longrightarrow \text{Testing} \longrightarrow \text{Verification} \longrightarrow \text{Approval} \longrightarrow \text{Deployment}$$ This establishes prompt engineering as a structured software engineering discipline. Codes map to behavioral, domain, or contextual specializations, such as: Technical engineering; Software programming; Financial analysis; Scientific discussion; Conversational interaction; Concise messaging; In-depth explanation; Critical evaluation; Re-affirmation of established context. 2.3. Dynamic Composite Prompt At the third layer, active codes are compiled into a unified composite context. Each signal maintains an intensity score, denoted internally as a temperature signal—a metric reflecting the activation strength of the given signal. Note: This term is distinct from the generation temperature parameter used in LLM sampling. DPS temperature measures signal activation strength rather than generative randomness. Based on these active signals, the system constructs a Composite Prompt passed to the primary LLM. Crucially, the primary LLM: Undergoes no retraining; Retains its original weights; Requires no specialized model variant; Executes a standard single generation pass. 3. Dynamic Temperature Profile – Breadcrumb State A core feature of DPS is that user profiling is not a one-off classification. It exists as a dynamic interaction state termed the Breadcrumb State. Profiling begins during the first exchanges of a conversation, eliminating the need to aggregate multiple historical sessions before initial activation. As dialogue progresses, individual signals can strengthen ($\uparrow$), weaken ($\downarrow$), disappear, be superseded, or interact with other active signals. Signal Dimension Trend Professional context $\uparrow$ Conciseness $\uparrow$ Need for detailed explanation $\downarrow$ Resistance to suggestion $\uparrow$ Confirmation of understanding $\uparrow$ This enables tracking not only topic content, but the evolving communicative dynamics of the session. 4. User Personalization The main purpose of DPS is adapting system behavior to an individual user. Across recurring sessions, a profile can persist (stored on the user side as a table of code mappings and signal temperatures) and evolve. Over time, the system continuously adapts to: Consistent professional contexts; Preferred task formulation styles; Desired response depth; Typical modes of confirmation or disagreement; Receptivity to suggestions; Preferred level of directness; Recurring communication patterns. This does not alter core LLM weights; instead, the interaction context shifts, allowing a single foundation model to serve diverse users through distinct dynamic adaptation profiles. 5. Professional Specialization Personalization does not degrade technical rigor. DPS can serve as a domain-specialization pre-layer between user requests and the primary LLM. When a user simultaneously addresses a technical problem, financial constraints, demands rigorous analysis, cites prior rules, and signals skepticism, traditional agentic architectures route the query through multiple stages: $$\text{Router} \longrightarrow \text{Technical Agent} \longrightarrow \text{Financial Agent} \longrightarrow \text{Context Aggregator} \longrightarrow \text{Main LLM}$$ DPS streamlines this pathway: $$\text{User Input} \longrightarrow \text{Semantic Micro-Model} \longrightarrow \text{Active Codes Selection} \longrightarrow \text{Composite Prompt} \longrightarrow \text{Main LLM}$$ Multiple context facets are integrated into a single pre-generation context, avoiding multi-agent pipeline overhead. 6. Reducing Agentic Complexity Modern AI implementations frequently rely on multi-agent orchestration: $$\text{Router} \longrightarrow \text{Specialist Agents (A, B, C)} \longrightarrow \text{Aggregator} \longrightarrow \text{Final LLM}$$ While effective, this paradigm increases inference latency, model call counts, compute costs, context size, and orchestration complexity. DPS offloads specialization that can be expressed via engineered prompt fragments to a lightweight micro-model prior to