In human conversation, what is said is rarely all that is meant. If someone asks, "Can you pass the salt?" they are not inquiring about your physical capabilities; they are making a request. This gap between literal meaning and intended message is the domain of pragmatics.
For decades, artificial intelligence struggled with this gap. Traditional systems treated language as a static mapping of symbols to meanings. However, a growing body of research suggests that communication is better modeled as a cooperative game of social reasoning. By using the Rational Speech Acts (RSA) framework, researchers are enabling machines to "read between the lines" by recursively modeling the mental states of their interlocutors.
The Hierarchy of Reasoning
The core intuition behind recursive pragmatic reasoning is that speakers and listeners are rational agents who simulate each other to resolve ambiguity. This is typically structured as a hierarchy of increasingly sophisticated agents:
- The Literal Listener ($L_0$): This agent interprets a message based solely on its conventional, dictionary definition. It does not consider context or the speaker's goals.
- The Pragmatic Speaker ($S_1$): This agent wants to convey a specific state of the world. To choose the best word, it simulates a literal listener and selects the utterance that is most likely to lead that listener to the correct conclusion while minimizing the "cost" (effort) of speaking.
- The Pragmatic Listener ($L_1$): This agent receives a message and asks: "Why would a rational speaker have chosen this specific word instead of another?" By simulating the pragmatic speaker, the listener can infer hidden intents.
As Frances Yung and Kevin Duh demonstrate in Modelling the Interpretation of Discourse Connectives by Bayesian Pragmatics, this recursive structure allows models to outperform literal statistical approaches. For example, a pragmatic listener can infer that if a speaker says "some," they likely mean "not all," because if they meant "all," a rational speaker would have used the more informative word "all."
Communication as Utility Maximization
Why does a speaker choose one word over another? In the RSA framework, this is treated as a probabilistic trade-off between informativeness and cost.
A speaker aims to maximize the "utility" of an utterance. Philippe Schlenker and Emmanuel Chemla describe this in Anti‐Babel: Three degrees of interspecies comprehension as a coordination problem where the sender selects signals that maximize information transfer efficiency.
This utility function can be tuned for different goals. In How do Large Language Models Navigate Conflicts between Honesty and Helpfulness?, Ryan Liu explores how models weight literal truth against the benefit to the listener. Sometimes, a "helpful lie" (like rounding 4:58 to "5:00") has higher communicative utility than a pedantically precise truth.
Grounding Language in Social Context
Pragmatic reasoning is essential for "grounded" tasks where language refers to the physical world. In Colors in Context: A Pragmatic Neural Model for Grounded Language Understanding, Will Monroe, Noah Goodman, and Christopher Potts show that a neural model using RSA can identify a specific shade of "blue" among several similar distractors. It does this by simulating how a speaker would describe that specific blue to distinguish it from the others.
This ability to resolve ambiguity through social context is a major frontier for robotics. Yanming Wan and Joshua B. Tenenbaum introduced the HandMeThat benchmark to highlight that while AI can label objects, it often fails to infer that "that one" refers to the knife a human is currently reaching for.
Scaling Pragmatics to Large Language Models
While Large Language Models (LLMs) like GPT-4 show some emergent pragmatic sensitivity, they often lack the explicit recursive depth of Bayesian models. Linlu Qiu and Yoon Kim found in On the Same Wavelength? Evaluating Pragmatic Reasoning in Language Models across Broad Concepts that LLMs struggle to generate effective pragmatic clues unless they are augmented with RSA-style reranking.
This has led to the development of "listener-adaptive" frameworks. Jacopo Teneggi and Zhenzhen Wang have applied these concepts to AI explainability in Adaptive Explanations via Direct Preference Optimization. Instead of a one-size-fits-all explanation, the system treats the explanation as a communication game, selecting concepts that maximize the specific listener's ability to infer the model's internal state.
The Information-Theoretic Frontier
The future of this field lies in grounding these social heuristics in formal mathematics. Noga Zaslavsky and Jennifer Hu have proposed a Rate-Distortion view of human pragmatic reasoning, suggesting that RSA is not just a clever trick, but a fundamental optimization of the trade-off between communicative benefit and cognitive effort.
As we move toward Strategic Foundation Models, as envisioned by Amy Greenwald and Roma Patel, the integration of game theory and recursive pragmatics will be the key to moving AI from simple text generators to sophisticated, collaborative agents.