When humans interact, we rarely communicate our intents perfectly on the first try. If a request is vague, the listener does not simply guess blindly; they ask a clarifying question, receive feedback, and iteratively narrow down the possibilities.
Yet, traditional AI systems have long operated under a strict single-turn interaction constraint. A user provides a prompt, and the model must deliver a flawless final output in a single pass. This one-shot paradigm creates a severe bottleneck when dealing with complex, real-world tasks where human intent is inherently under-specified, ambiguous, or evolving. To overcome this limitation, a major research direction focuses on transitioning models from static, single-turn execution to multi-turn iterative clarification.
The Single-Turn Interaction Bottleneck
Why does the single-turn paradigm fail? At its core, it forces a model to resolve a massive, high-entropy search space without a feedback loop.
Consider code generation or data visualization. As Li et al. (2023) observe, user prompts for text-to-code tasks are frequently too vague or under-specified to yield correct programs on the first attempt. Similarly, in text-to-visualization tasks, Luo et al. (2025) point out that a single imprecise query can map to dozens of equally valid charts. When forced to respond immediately, models either make presumptuous guesses or fail entirely, leaving the user stuck without a clear path forward (Elgohary et al., 2020).
Even in information retrieval and web search, web queries are often ambiguous. Traditional search engines that rely on a single-turn paradigm fail to capture the nuanced, exploratory nature of complex information seeking (Zamani et al., 2020). Human understanding of dialogue naturally involves a process of re-evaluating previous statements as new information emerges; without a recurrent mechanism to refine initial estimates, models struggle to resolve complex dependencies between intertwined conversational tasks (Xing, 2022).
By breaking this single-turn constraint, we can treat interaction not as an isolated transaction (Rago et al., 2024), but as a dynamic control problem where intent is progressively uncovered. As mapped out by Qu et al. (2018) in technical support dialogues, successful problem-solving naturally progresses through predictable, multi-turn patterns of feedback and clarification.
Quantifying Ambiguity: When and What to Ask
For an AI agent to engage in multi-turn clarification, it must possess two critical capabilities: it must know when it is confused, and it must know how to formulate a question that maximally reduces that confusion.
If a model asks questions constantly, it becomes tedious and over-communicative; if it never asks, it blunders (Kahlon et al., 2025). To mathematically determine the optimal moment to pause and clarify, Choi et al. (2025) introduced a framework called INTENT-SIM. Instead of relying on standard token-level confidence, INTENT-SIM simulates potential user intents to calculate "intent entropy." If the entropy across possible interpretations crosses a threshold, the model proactively pauses to ask for clarification.
[User Prompt: "Plot the revenue data"]
│
▼
┌──────────────────────────────┐
│ Simulate Possible Intents │ ──► Intent A: Bar chart by year?
└──────────────────────────────┘ ──► Intent B: Line graph over time?
│
▼
┌──────────────────────────────┐
│ Calculate Intent Entropy │
└──────────────────────────────┘
│
┌───────┴───────┐
▼ ▼
[Low Entropy] [High Entropy]
│ │
▼ ▼
Execute Task Proactively Pause & Ask Clarifying Question
Once a model decides to clarify, it must generate a question that targets the missing factual information efficiently. Toles et al. (2023) address this by training models to identify specific missing facts using fact-level masking, optimizing the question to fill the exact information gap.
To teach models the long-term value of clarification, Zhang et al. (2025) developed a double-turn preference labeling method. Instead of evaluating a clarifying question in isolation, this method trains the model to simulate the successful outcomes of future conversation turns. By looking ahead, the model learns that investing an immediate turn into asking a question yields a significantly higher probability of a correct final answer. This same logic applies to Retrieval-Augmented Generation (RAG) systems: rather than providing a binary rejection when a query is unanswerable, a robust system engages in iterative dialogue to resolve the underspecified request (Choubey et al., 2025).
Incremental Grounding Across Diverse Domains
Multi-turn clarification works because it allows the model and the user to build "common ground" incrementally. This mechanism is being deployed across a wide variety of domains to bridge the semantic gap between human intent and machine execution.
- Content Creation: Instead of requiring non-expert users to write highly complex, detailed instructions from scratch, strategies like "HELP ME THINK" shift the burden of detail-gathering to the AI (Mishra et al., 2023). The model initiates a multi-turn dialogue, asking a series of targeted questions to iteratively refine the user's customized goals.
- Multimodal Grounding and UI Automation: In phone UI automation, an agent might encounter multiple identical icons or ambiguous choices. Frameworks like
Moraeproactively pause at these critical decision points, using multi-turn negotiations to resolve nested user preferences (Peng et al., 2025). By allowing the agent to refine its actions based on iterative user feedback, systems achieve massive performance boosts on complex, ambiguous UI tasks (Li et al., 2024). - Embodied Navigation: In physical spaces, static, one-way instructions often fail due to environmental uncertainties. Grounded language models, such as those utilizing the Rational Speech Acts (RSA) framework, simulate how a speaker chooses words to distinguish targets from neighbors, but require multi-turn feedback loops to handle natural communication (Monroe et al., 2017). Embodied agents that can proactively ask questions (
ELBA) utilize multi-turn dialogue to resolve deeply nested ambiguities during household tasks (Shen et al., 2025), leveraging a full history of past questions and answers to navigate effectively (Cakmak, 2019). - Niche and Scientific Reasoning: In specialized domains like virtual physiotherapy coaching, one-off feedback is insufficient; users need to ask follow-up questions about specific postural corrections over a continuous dialogue (Wang, 2023). Similarly, scientific question answering across complex tables and text requires multi-turn dialogues where users can refine queries based on intermediate findings (Wang, 2025), just as session-based recommendation systems require iterative dialogue to synthesize deep user intent (Wang et al., 2025).
Internalizing the Dialogue: Multi-Turn Self-Correction
The intuition behind multi-turn clarification extends beyond human-to-computer interactions; it is also being used to unlock advanced reasoning within the models themselves.
When a model attempts to solve a highly complex logical puzzle or mathematical problem in a single-turn inference pass, its ability to explore the state space is severely limited (Chen et al., 2025). Traditional single-turn reasoning exploration leaves the dynamics of iterative self-correction completely untouched (Hu, 2026).
To solve this, researchers are structuring the reasoning process as an internal multi-turn dialogue. By pairing a primary agent with an environmental verifier (Chen et al., 2025) or an internal "reflector" model (Dou et al., 2024), the system can treat its own initial outputs as ambiguous or potentially flawed hypotheses. The agent and reflector engage in a multi-turn exchange, using environmental feedback to iteratively isolate errors, refine hypotheses, and self-correct.
Ultimately, whether interacting with a human user or probing a logical state space, moving beyond the single-turn constraint allows AI systems to treat communication and reasoning as they truly are: iterative, adaptive, and collaborative processes.
Key Foundational Work
Spoken Interactive ODQA System: SPIQA (2003) — contributed 1 concept elements to this direction.
Colors in Context: A Pragmatic Neural Model for Grounded Language Understanding (2017) — contributed 1 concept elements to this direction.