Planning Ahead with RSA: Efficient Signalling in Dynamic Environments by Projecting User Awareness across Future Timesteps

TL;DR

This work extends Rational Speech Act (RSA) from single-turn communication to dynamic, multi-step signaling in time-evolving environments. By introducing a finite-horizon planning framework, user-specific priors, and explicit belief tracking, the authors design an adaptive assistant that optimizes what to say, how to say it, and when to say it to preserve operator situational awareness. The full model—d-RSA with Priors and Planning—outperforms baselines across diverse, high-difficulty scenarios, reducing timely-delivery of critical alerts and exploiting anticipatory communications when user knowledge permits. The study provides theoretical foundations for pragmatic, human-aware AI in safety-critical domains and points to future hybrid approaches that combine RSA with neural methods and planning under uncertainty. Overall, this framework advances principled, cognitively grounded human-AI collaboration by explicitly modeling attention, belief updates, and temporal planning in dynamic environments.

Abstract

Adaptive agent design offers a way to improve human-AI collaboration on time-sensitive tasks in rapidly changing environments. In such cases, to ensure the human maintains an accurate understanding of critical task elements, an assistive agent must not only identify the highest priority information but also estimate how and when this information can be communicated most effectively, given that human attention represents a zero-sum cognitive resource where focus on one message diminishes awareness of other or upcoming information. We introduce a theoretical framework for adaptive signalling which meets these challenges by using principles of rational communication, formalised as Bayesian reference resolution using the Rational Speech Act (RSA) modelling framework, to plan a sequence of messages which optimise timely alignment between user belief and a dynamic environment. The agent adapts message specificity and timing to the particulars of a user and scenario based on projections of how prior-guided interpretation of messages will influence attention to the interface and subsequent belief update, across several timesteps out to a fixed horizon. In a comparison to baseline methods, we show that this effectiveness depends crucially on combining multi-step planning with a realistic model of user awareness. As the first application of RSA for communication in a dynamic environment, and for human-AI interaction in general, we establish theoretical foundations for pragmatic communication in human-agent teams, highlighting how insights from cognitive science can be capitalised to inform the design of assistive agents.

Figures (3)

• Figure 1: Impact of individualised alert sequences (x-axis) for distinct users in the same environment on user awareness (y-axis). Timesteps (x-axis) are annotated with the onset of critical events. Our model uses recursive Bayesian reasoning to construct user-specific models of how each user updates their world beliefs in response to alert sequences, optimising for global situational awareness over time.
• Figure 2: Total reward (y-axis) for each model variant across scenario factors. (a) Reward as a function of the number of critical properties (higher = more challenging). (b) Reward vs. temporal dispersion of critical onsets (lower = more tightly clustered, i.e., challenging). The full model achieves the greatest relative gains as scenarios become more difficult.
• Figure 3: (a) Message Specificity Ratio (y-axis) vs. Prior User Awareness of critical properties (x-axis). Specificity decreases with higher prior awareness only in the full model, showing its adaptive use of less specific alerts; baseline variants remain less flexible. (b) Median Delay (in timesteps, y-axis) to the First Alert about Low-Awareness Critical Properties vs. Number of Critical Properties (x-axis). The full model is consistently faster as scenario difficulty grows, outperforming baseline variants.