Navigating the Synchrony-Stability Frontier in Adaptive Chatbots

TL;DR

The paper establishes a formal synchrony–stability frontier for adaptive chatbots, showing that constraining turn-by-turn stylistic changes via bounded policies (Cap, EMA, Dead-Band, and Hybrids) yields substantial gains in persona stability and prompt legibility with only modest losses in immediate synchrony. It introduces an 8-dimensional style vector and a base+delta prompting framework within a closed-loop control system, and provides extensive validation across a 162-participant human-log dataset and three public corpora, plus LLM-in-the-loop experiments with GPT-4.1 nano and Claude Sonnet 4. The work demonstrates a robust, generalizable trade-off: Pareto-efficient policies on the frontier improve coherence and maintainability (prompt legibility, reduced register flips) while preserving user experience, offering concrete guidelines for deploying adaptive conversational agents. The accompanying reproducible artifact enables researchers to reproduce results and explore policy choices in diverse domains, informing principled design of trustworthy, adaptive AI systems.

Abstract

Adaptive chatbots that mimic a user's linguistic style can build rapport and engagement, yet unconstrained mimicry risks an agent that feels unstable or sycophantic. We present a computational evaluation framework that makes the core design tension explicit: balancing moment-to-moment linguistic synchrony against long-term persona stability. Using an 8-dimensional style vector and a closed-loop "base+delta" prompting architecture, we simulate and compare explicit adaptation policies - Uncapped, Cap, Exponential Moving Average (EMA), Dead-Band, and Hybrids - on a human-log dataset. Our analysis maps a clear Pareto frontier: bounded policies achieve substantial gains in stability at a modest cost to synchrony. For example, a Hybrid (EMA+Cap) raises stability from 0.542 to 0.878 (+62%) while reducing synchrony by only 17%. We confirm this trade-off through large-scale replications on three public corpora (DailyDialog, Persona-Chat, EmpatheticDialogues) and LLM-in-the-loop validation across two model families. Furthermore, we quantify "prompt legibility," showing that frontier policies reduce instruction churn and cut jarring register flips (major tone changes) from 0.254 to 0.092, yielding systems that are easier to reason about and maintain. Taken together, our framework provides a general evaluation harness for style adaptation; a systematic ablation that identifies Pareto-efficient policies; robust validation across diverse datasets and models; and novel legibility metrics linking policy choices to system maintainability.

Figures (7)

• Figure 1: The Synchrony--Stability Trade-off Generalizes Across LLMs. Each point represents a policy's mean performance. The lines illustrate the trade-off: Uncapped anchors the high-synchrony, low-stability corner, while bounded policies like Hybrid (EMA+Cap) achieve higher stability at a cost to synchrony. The same qualitative pattern holds for both OpenAI and Anthropic model families, and the per-participant deltas are statistically significant.
• Figure 2: Closed-loop style adaptation system. Each turn, the user’s style vector $\mathbf{u}_t$ is extracted, the policy combines it with prior state $\mathbf{b}_{t-1}$ (and persona centroid $\mathbf{b}_c$) to choose a target $\tilde{\mathbf{b}}_t$, which is translated into a delta of style instructions and composed with a base prompt. The LLM’s reply realizes $\mathbf{b}_t$, which is logged and fed back into session state for the next turn.
• Figure 3: The Synchrony--Stability Frontier on Human-Log Data. Each point is a policy mean. Uncapped anchors the high-synchrony/low-stability extreme. Bounded Hybrid variants trace the efficient frontier. With $\epsilon{=}0.1$, Dead-Band sits near Uncapped---yielding only a negligible stability gain---and is not strictly dominated.
• Figure 4: Effect of the Hybrid Policy versus Uncapped (LLM-in-the-loop). Mean per-participant changes with 95% bootstrap CIs. For both GPT-4.1 nano and Claude Sonnet 4, Hybrid significantly increases Stability and decreases Synchrony; confidence intervals do not include zero.
• Figure 5: Impact of User-History Window on Predictive Synchrony. A one-turn window is most predictive of the user's next-turn style (mean=0.645), with performance declining and then plateauing as the window grows (mean $\approx$ 0.597 at 8 turns).