When designing conversational agents, a natural instinct is to maximize helpfulness by providing as much relevant information as possible. However, human communication operates under a strict constraint: cognitive load. If an agent provides too much information at once, it overwhelms the user's working memory, leading to a paradoxical drop in comprehension. Conversely, if it is too brief, the user must endure tedious clarification loops.
This tension defines the helpfulness-brevity tradeoff. To build effective conversational systems, researchers have long studied how to dynamically manage information density—ensuring that responses are sufficiently supportive without becoming mentally taxing.
The Cognitive Bottleneck: Why Less is More
Why does providing more information often result in a worse user experience? The answer lies in how humans process language sequentially.
In early spoken dialogue systems (SDS), this bottleneck was highly visible because speech is a serial and transient medium; unlike text, a listener cannot easily skim an audio stream. As Brian Langner et al. demonstrated in Generating time-constrained audio presentations of structured information, simply reading out structured data linearly causes user retention to plummet. When presenting information over audio, a concise summary yields significantly higher comprehension than an exhaustive list.
[Too Brief] --------> High Temporal Cost (Endless clarification loops)
[Optimal] --------> Low Cognitive Load + High Task Success
[Verbose] --------> High Mental Tax (Working memory saturation)
This constraint becomes even more acute in multitasking environments, such as in-car navigation. Work by Heather Pon-Barry et al. in Evaluation of content presentation strategies for an in-car spoken dialogue system and Andi Winterboer et al. in The influence of user tailoring and cognitive load on user performance in spoken dialogue systems highlighted that controlling turn length and information density is vital for preventing driver distraction. If a system delivers an overly dense response while a user is driving, the user's performance on the primary task degrades. Similarly, for indoor navigation, Perttu Prusi et al. developed specialized grammars in Towards generic spatial object model and route guidance grammar for speech-based systems to carefully meter out spatial details, preventing the system from overloading the user's mental map.
Dialogue Generation as an Optimization Problem
To resolve this tension, researchers moved away from static, rule-based responses toward dynamic optimization. If a system treats information delivery as a sequential decision-making problem under uncertainty, it can mathematically compute the sweet spot between brevity and informativeness.
A foundational series of papers by Verena Rieser et al. pioneered this approach using Reinforcement Learning (RL). In Learning dialogue strategies for interactive database search, the authors used RL to decide exactly how many search results to present at once. The model learned to balance the cognitive load of a large result set against the temporal cost of asking additional filtering questions.
By framing natural language generation as statistical planning under uncertainty across several works, the authors showed that an adaptive policy could maximize task success by adjusting information density based on the user's immediate context and the volume of data available:
- In Natural Language Generation as Planning Under Uncertainty for Spoken Dialogue Systems and Natural Language Generation as Incremental Planning Under Uncertainty: Adaptive Information Presentation for Statistical Dialogue Systems, they demonstrated that incremental planning allows a system to adapt its generation strategy mid-dialogue to avoid exceeding human working memory limits.
- In Optimising Information Presentation for Spoken Dialogue Systems and Adaptive Information Presentation for Spoken Dialogue Systems: Evaluation with real users, they confirmed through user studies that joint optimization—penalizing both excessive length and insufficient detail—significantly outperforms human-crafted heuristics.
This optimization perspective dates back to early computational linguistics. For instance, the structural need to balance helpfulness with conciseness when providing indirect or non-explicit answers was recognized by Nancy L. Green et al. in A Hybrid Reasoning Model for Indirect Answers, establishing that dialogue systems must explicitly reason about when to add supplementary details versus when to remain concise.
The Modern Parallel: Verbosity Bias in LLMs
While early research was driven by the physical constraints of speech and audio, the helpfulness-brevity tradeoff has re-emerged as a central challenge in modern Large Language Models (LLMs).
Today's LLMs are heavily optimized via reinforcement learning from human feedback (RLHF) to be maximally helpful. However, this optimization introduces a severe side effect: conversational verbosity bias. Because models are rewarded for completeness, they default to exhaustive, over-explained, and robotic responses that conflict with human preferences for casual engagement.
This phenomenon creates what Md Awsafur Rahman et al. term a "memory-likability gap" in LikeBench: Evaluating Subjective Likability in LLMs for Personalization. Their evaluation framework reveals that while a model might excel at retaining and reciting facts, its subjective likability drops when it fails to meet user expectations for brevity.
This is echoed by the findings of Lisa Dunlap et al. in VibeCheck: Discover and Quantify Qualitative Differences in Large Language Models. By building an automated framework to quantify the stylistic "vibes" of different models, the authors discovered that excessive verbosity in technical or reasoning tasks frequently correlates with lower user satisfaction. In practice, users often prefer concise, formal narratives over long-winded, over-explained text.
Adapting to User Registers
To resolve the verbosity bias, modern systems must dynamically align their information density with the user's implicit preferences and emotional state.
In sensitive contexts like emotional support, over-explaining can alienate users. To address this, Junlin Li et al. introduced a framework in Be Helpful but Don't Talk too Much - Enhancing Helpfulness in Conversations through Relevance in Multi-Turn Emotional Support based on the "Cognitive Relevance Principle." They demonstrate that excessively wordy responses feel repetitive and counterproductive to a user's emotional well-being, requiring a strict calibration between supportive content and brevity.
| Conversational Context | Risk of High Density | Mitigation Strategy |
|---|---|---|
| Task-Oriented / Navigation | Cognitive overload, safety risks | Hierarchical summaries, strict turn-length limits |
| Emotional Support | Perceived as robotic, repetitive, or apathetic | Cognitive Relevance Principle, empathetic brevity |
| Open-Domain / Chat | Lower subjective likability, frustration | Conversational style mirroring |
One powerful mechanism for achieving this calibration automatically is conversational mirroring. As Katherine Metcalf et al. demonstrated in Mirroring to Build Trust in Digital Assistants, an assistant can reliably predict whether a user prefers a concise or a talkative response by analyzing the acoustic and stylistic features of their initial query. By matching the user's preferred level of "chattiness," the system avoids the friction of a mismatched conversational register—proving that managing information density is not just about data transmission, but about building human trust.
Key Foundational Work
Towards generic spatial object model and route guidance grammar for speech-based systems (2005) — contributed 2 concept elements to this direction.
A Hybrid Reasoning Model for Indirect Answers (1994) — contributed 1 concept elements to this direction.