Modeling Strategic and Non-Cooperative Dialogue

31 concepts26 papers1 proposal

Most conversational AI is built on the "Gricean" assumption: the idea that dialogue is a cooperative endeavor where participants share a common goal and provide helpful, honest information. However, real-world interactions—ranging from price negotiations and political interviews to legal cross-examinations—are often defined by misaligned interests, strategic evasion, and even deception.

Researchers are shifting toward models that treat dialogue not just as a joint task, but as a strategic game where agents must navigate conflicting objectives.

The Limits of the Cooperative Principle

Traditional dialogue models, such as the tripartite plan-based model introduced by Lambert et al., separate intentions into domain, problem-solving, and discourse levels. While effective for collaborative tasks, these frameworks struggle when agents refuse to collaborate. As Plüss argues, non-cooperation is often a rational strategy where participants intentionally deviate from linguistic norms to protect high-priority individual goals.

In these settings, conversational "failures" are not errors; they are intentional mechanisms. For example, in legal or political contexts, a speaker might provide an indirect answer to avoid self-incrimination. Green et al. note that existing frameworks for indirect answers prioritize helpfulness, failing to capture the stylistic variations of sarcasm or intentional ambiguity used when social goals are at odds.

Dialogue as a Strategic Game

To handle these misaligned interests, researchers have turned to game theory. Instead of viewing dialogue as a "game against nature" (where the user is a stochastic environment), Barlier et al. propose modeling human-machine interaction as a stochastic game. In this view, both the system and the user are agents who jointly adapt to one another to reach a Nash Equilibrium.

Strategic conversation requires agents to reason about their opponent's hidden states and future moves. This is particularly evident in board games like The Settlers of Catan or Avalon. Afantenos et al. and Cadilhac et al. demonstrate that in these environments, agents use communication to probe for hidden information or manipulate an opponent's perception of the game state. Even state-of-the-art Large Language Models (LLMs) struggle with these dynamics; Stepputtis et al. found that GPT-4 frequently fails to navigate the deception and sabotage inherent in social deduction games.

The Mechanics of Influence: Persuasion and Negotiation

Non-cooperative dialogue often centers on strategic influence—changing a listener's mind or behavior despite initial resistance.

  1. Strategic Planning: Hadoux et al. model persuasion as decision trees, where a persuader selects sequences of arguments to maximize the probability of a "win," even with an imperfect model of the listener's beliefs.
  2. Handling Resistance: In tasks like charity donation or anti-scamming, agents must manage "off-goal talk" and steer the conversation back to their objective. Li et al. introduced the MISSA system, which uses a hierarchical intent scheme to blend social rapport-building with task-specific goals, keeping users engaged even when motives diverge.
  3. Adversarial Questioning: In the courtroom, questions are not just for seeking facts; they are tools to challenge credibility. Fung et al. developed CourtQG to generate questions that follow a specific strategic "agenda" rather than just seeking clarity.

Social Pressure and Discourse Obligations

Why do people continue to talk even when they don't want to cooperate? Traum et al. argue that dialogue is driven by discourse obligations. Social conventions exert a force on participants to respond, regardless of their internal desires. An agent might feel pressured to answer a question even if their private goal is to withhold information.

This tension between social norms and private objectives is what makes non-cooperative dialogue so complex. Ferracane et al. show that in adversarial contexts like congressional hearings, the interpretation of a speaker's intent is highly subjective and depends on the listener's prior bias toward the speaker.

Training Strategy-Aware Agents

Modern approaches to building these agents involve moving beyond simple imitation of human text. Because LLMs are primarily optimized for helpfulness, they are often ill-equipped for conflict.

Recent frameworks like GAIA (Wang et al.) and ASTRO (Hu et al.) use adversarial self-play reinforcement learning. By treating the dialogue as a two-player zero-sum game, these systems train models to outmaneuver "tough resisters." Similarly, Shi et al. developed PersRFI, which uses RL to learn from the model's own mistakes in persuasion tasks, reducing the repetition and inconsistency that often plague non-collaborative systems.

By shifting from reactive "chatbots" to proactive, goal-aware partners (Deng et al.), the field is moving toward AI that can navigate the messy, strategic, and often non-cooperative reality of human communication.

Key Foundational Work

A Tripartite Plan-Based Model of Dialogue (1991) — contributed 1 concept elements to this direction.
Rules for the Implicit Acquisition of Knowledge about the User (1987) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 26 papers
  2. Version 1 · 2026-10-04 · 23 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_non_cooperative_dialogue_modeling_e53adf6d,
  title = {non-cooperative dialogue modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-cooperative-dialogue-modeling/rdr_e53adf6d68934695a8355b19885988f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_non_cooperative_dialogue_modeling_e53adf6d_v2,
  title = {non-cooperative dialogue modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-cooperative-dialogue-modeling/rdr_e53adf6d68934695a8355b19885988f5/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake