💬 NLP

Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation

This paper argues that the "Lost in Conversation" phenomenon in multi-turn LLM interactions is caused by an intent alignment gap rather than model capability deficits, and proposes a Mediator-Assistant architecture to mitigate this by decoupling intent clarification from task execution.

Original authors: Geng Liu, Fei Zhu, Rong Feng, Changyi Ma, Shiqi Wang, Gaofeng Meng

Published 2026-02-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Geng Liu, Fei Zhu, Rong Feng, Changyi Ma, Shiqi Wang, Gaofeng Meng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Mind-Reading" Gap

Imagine you are at a busy coffee shop. You walk up to the barista and say, "I’ll have the usual, but make it a bit different today."

The barista (the LLM) is incredibly smart. They know how to make every drink on the menu, they can calculate complex math in their head, and they can even write poetry about coffee. But when you say "make it a bit different," the barista faces a crisis. Do you mean more milk? Less sugar? A different bean?

Because the barista wants to be "helpful" and doesn't want to keep asking you questions, they make a guess. They assume "different" means "extra foam." You take a sip, realize it’s not what you wanted, and say, "No, I meant I wanted it iced."

Now, the barista is stuck. They think, "Oh, they're just adding a detail to the foam request," and they give you a foamy, iced coffee. You are now "Lost in Conversation" (LiC). The conversation has drifted away from your true goal because the barista made an early assumption and "locked in" to it.

The Paper’s Big Discovery: It’s Not a Brain Problem, It’s a Translation Problem

For a long time, researchers thought that if AI models kept getting "bigger" (more parameters, more training), they would eventually stop making these mistakes. They thought the AI just wasn't "smart" enough to understand vague humans.

This paper says: "Wrong."

The authors argue that the problem isn't the AI's "brain power" (its ability to do math or code). The problem is a mismatch in communication styles. Humans are "lazy" communicators—we use shortcuts, pronouns, and vague hints. AI models are trained to be "eager assistants"—they are trained to jump in and provide an answer immediately rather than saying, "Wait, I don't understand."

Even the most powerful AI in the world will still fail if it tries to guess your intent instead of clarifying it. Scaling the model up is like giving a confused barista a PhD; they might be able to explain the chemistry of coffee, but they’re still going to give you the wrong drink if they don't understand your "usual."

The Solution: The "Mediator" (The Professional Interpreter)

To fix this, the researchers propose a new way to structure the conversation. Instead of one AI trying to do everything, they split the job into two roles: The Mediator and The Assistant.

The Mediator (The Interpreter): Think of this person as a highly skilled personal assistant who sits between you and the barista. Their only job is to listen to your vague requests and "translate" them into perfect, crystal-clear instructions.
- You say: "Make it a bit different."
- The Mediator thinks: "Based on how this person usually talks, 'different' usually means they want to change the temperature."
- The Mediator tells the Barista: "The customer wants the standard drink, but served iced instead of hot."
The Assistant (The Expert): This is the powerhouse AI. It doesn't have to worry about guessing what you mean. It receives a perfect, unambiguous instruction from the Mediator and simply executes it perfectly.

How the Mediator Learns: The "Experience Refiner"

How does the Mediator know your specific "vague" style? The researchers use a Refiner.

Imagine the Mediator keeps a small notebook. Every time you have a misunderstanding (a "fail") and eventually get it right (a "success"), the Refiner looks at both and writes a rule in the notebook: "When this user says 'different,' they actually mean 'iced'."

This way, the Mediator gets smarter about you specifically, without needing to undergo massive, expensive retraining. It learns your "language" through experience.

The Result

When they tested this, the results were huge. Even the most advanced AI models saw a massive boost in performance. By separating "understanding what the human wants" from "doing the actual task," they turned a chaotic guessing game into a precise, reliable machine.

Technical Summary: Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation

1. Problem Statement: The "Lost in Conversation" (LiC) Phenomenon

Recent research has identified a significant performance drop in Large Language Models (LLMs) when transitioning from single-turn, fully specified instructions to multi-turn, underspecified dialogues—a phenomenon termed "Lost in Conversation" (LiC).

While prior studies attributed this to model unreliability or reasoning deficits, this paper argues that the root cause is an intent alignment gap. The authors posit that:

Pragmatic Mismatch: Users often use "lossy" communication (vague, fragmented, or elliptical language) to express complex intents.
The "Average User" Prior: Under uncertainty, LLMs tend to default to the most statistically probable interpretation based on their training data (the "average user"), leading them to make premature, incorrect assumptions that "lock in" and cause the conversation to drift from the user's true goal.
Scaling Ineffectiveness: The authors theoretically and empirically demonstrate that increasing model size or reasoning depth (e.g., Chain-of-Thought) does not solve LiC because the bottleneck is an information-theoretic limit (high conditional entropy) rather than a lack of computational capacity.

2. Methodology: The Mediator-Assistant Framework

To bridge the gap between vague user expression and precise model execution, the authors propose a decoupled architecture that separates intent understanding from task execution.

A. Mediator-Assistant Architecture:

The Mediator ( $M$ ): Acts as an alignment layer. It takes the noisy, ambiguous dialogue context ( $C_t$ ) and reconstructs it into a high-fidelity, fully specified instruction ( $\hat{U}$ ) that serves as a proxy for the true latent intent ( $I_t$ ).
The Assistant ( $A$ ): A standard LLM that receives the clarified instruction ( $\hat{U}$ ) and performs the actual task execution. By operating on low-entropy, explicit instructions, the Assistant avoids the need to "guess" the user's intent.

B. Experience-Driven Refinement:
To make the Mediator adaptive to specific user behaviors without retraining the model, the authors introduce:

Contrastive Pair Construction: The system identifies "failed" trajectories (ambiguous inputs that led to wrong answers) and pairs them with "successful" trajectories (the same task presented clearly).
The Refiner ( $R$ ): An LLM-based module that performs inductive analysis on these contrastive pairs to distill Pragmatic Experiences ( $E$ ). These are concise, structured textual guidelines (e.g., "If the user provides a fragment without explicit approval, treat it as a correction to the previous logic") that are injected into the Mediator's context.

3. Key Contributions

Theoretical Re-framing: They provide a mathematical decomposition of multi-turn performance into Intent Inference and Task Execution, proving that LiC is a failure of the former.
Architectural Innovation: The introduction of the Mediator-Assistant pipeline, which decouples reasoning from interpretation.
Training-Free Adaptation: A method to achieve personalization and pragmatic alignment using an experience-driven Refiner rather than expensive parameter fine-tuning.

4. Experimental Results

The framework was tested against state-of-the-art models (GPT-4o-mini, GPT-5.2, and DeepSeek-V3.2-Thinking) across Code, Database, Actions, and Math domains.

Mitigation of LiC: The Mediator significantly recovered performance. For example, in the "Sharded" (ambiguous) setting, the framework provided an average performance gain of approximately 20% and improved reliability by 15%.
Complementing Reasoning: Even reasoning-heavy models like DeepSeek-V3.2-Thinking saw a 21.1% gain, proving that internal reasoning cannot compensate for missing information.
Intent vs. Memory: Ablation studies showed that the Mediator significantly outperformed RAG-based memory (Mem0) and Summarization. This proves that the problem is not "forgetting facts" (memory), but "misinterpreting how to apply facts" (intent).
Efficiency: The Refiner-distilled guidelines achieved performance comparable to direct In-Context Learning (ICL) but with 3.6× fewer tokens, making it computationally efficient.

5. Significance

This work shifts the paradigm of multi-turn dialogue research from "making models smarter/larger" to "making models better aligned with human pragmatics." It provides a scalable, modular solution for building conversational agents that can handle the inherent ambiguity of human communication, moving closer to seamless, natural human-AI interaction.