Timing the Message: Language-Based Notifications for Time-Critical Assistive Settings

Ya-Chuan Hsu, Jonathan A. DeCastro, Andrew Silva, Guy RosmanarXiv 2025 · DOI · arXiv

Abstract

In high-stakes environments such as autonomous driving or piloting, AI assistants must deliver notifications that are both timely and informative. However, traditional AI communication models often treat language as an instantaneous event, neglecting the critical time required for speech conveyance and human cognitive processing. This paper introduces a reinforcement learning framework that explicitly models these temporal delays. By incorporating a reactive human model and an LLM-based surrogate for comprehension mapping, the proposed "Convey & React" policy optimizes the trade-off between speed and detail. Experimental results across three time-critical domains demonstrate a significant improvement in task success rates compared to delay-agnostic baselines.

Introduction: The Timeliness-Informativeness Trade-off

Effective AI assistance in dynamic environments requires navigating a fundamental tension: providing enough detail for the human to act correctly while ensuring the message is completed before the window of opportunity closes. Most current systems either provide rapid but context-poor alerts (e.g., "Stop!") or detailed instructions that arrive too late.

The core of this challenge lies in the temporal nature of language. As shown in Figure 1, a notification involves several distinct phases: the conveyance time ($l$) required to speak the words, the human reaction delay ($d_r$), and the subsequent follow-through time ($d_f$). Crucially, humans often reach a point of "maximum understanding" ($k$) before a sentence is finished, allowing for incremental action.

Figure 1: The temporal components of a language-based notification, including conveyance time, reaction delay, and the point of maximum understanding.

Methodology: Modeling the Human in the Loop

To address these delays, the researchers formulated the problem as an augmented-state Markov Decision Process (MDP). This allows the AI agent to reason over a history of states, accounting for the time elapsed during communication.

Reactive Human and LLM-Surrogate Models

The framework utilizes a mathematical model to simulate human responses, characterized by a reaction delay and a follow-through duration. To scale the training process without exhaustive human labeling, the authors employed Large Language Models (LLMs) as surrogate annotators. These LLMs predict word-by-word comprehension levels, identifying the "actionable index"—the specific word where a human is likely to grasp the command's intent.

Policy Architecture

The system employs a "Timing-and-Intent" network trained via Proximal Policy Optimization (PPO). This network decides precisely when to initiate a notification and selects the semantic intent from an offline taxonomy. This taxonomy is a pre-generated library of utterances mapped to their specific temporal properties, such as length and comprehension points.

Experimental Evaluation and Results

The framework was evaluated in three simulated domains: Lunar Lander (piloting), Highway Merging (driving), and a collaborative Steakhouse (cooking) environment.

As summarized in Figure 6, the "Convey & React" model significantly outperformed traditional "Delay-Free" and "Heuristic" baselines. In the piloting and driving tasks, the model achieved success rates of 97% and 93% respectively—representing a performance jump of over 40% compared to models that ignore communication delays.

Figure 6: Success rates and secondary metrics across different notifier policies, highlighting the superior performance of the Convey & React model.

Key observations from the results include:

  • Adaptive Strategy: The AI learns to favor shorter, urgent alerts in highly volatile situations while switching to longer, context-rich instructions when the environment is stable.
  • Robustness: The policy remains effective even when the human's actual reaction time deviates slightly from the training parameters, provided the delay remains within a manageable threshold.

Conclusion

This research highlights that language in assistive AI is not a point-in-time event but a temporal process. By explicitly modeling conveyance and comprehension delays, AI assistants can better support humans in time-critical settings. Future work will focus on personalizing these models to individual human reaction profiles and addressing potential linguistic biases in the LLM-surrogate models.

Summary updated

Abstract

In time-critical settings such as assistive driving, assistants often rely on alerts or haptic signals to prompt rapid human attention, but these cues usually leave humans to interpret situations and decide responses independently, introducing potential delays or ambiguity in meaning. Language-based assistive systems can instead provide instructions backed by context, offering more informative guidance. However, current approaches (e.g., social assistive robots) largely prioritize content generation while overlooking critical timing factors such as verbal conveyance duration, human comprehension delays, and subsequent follow-through duration. These timing considerations are crucial in time-critical settings, where even minor delays can substantially affect outcomes. We aim to study this inherent trade-off between timeliness and informativeness by framing the challenge as a sequential decision-making problem using an augmented-state Markov Decision Process. We design a framework combining reinforcement learning and a generated offline taxonomy dataset, where we balance the trade-off while enabling a scalable taxonomy dataset generation pipeline. Empirical evaluation with synthetic humans shows our framework improves success rates by over 40% compared to methods that ignore time delays, while effectively balancing timeliness and informativeness. It also exposes an often-overlooked trade-off between these two factors, opening new directions for optimizing communication in time-critical human-AI assistance.
BibTeX
@misc{hsu2025timing,
  author = {Ya-Chuan Hsu and Jonathan A. DeCastro and Andrew Silva and Guy Rosman},
  title = {{Timing the Message: Language-Based Notifications for Time-Critical Assistive Settings}},
  year = {2025},
  eprint = {2509.07438},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.07438},
}
Figure