Illusion of Alignment: Detecting Hidden Disagreement in Collaborative Dialogue
Introduces IoA-Suite for hidden disagreement detection; model F1 only 49.5%, private context is bottleneck.
Key Findings
Methodology
This work formalizes the illusion of alignment (IoA) as a latent divergence in participants' understanding of goals, assumptions, or plans after dialogue ends. It proposes generating diagnostic multiple-choice questions (MCQ) where divergent answers indicate hidden disagreement. IoA-Suite, a dataset spanning five task types and six domains, is constructed for evaluation. Experiments show models achieve only 49.5% F1, limited by private context information. A trained IoA-Prober-8B model improves F1 to 51.8%, successfully surfacing 2.89 hidden disagreements per real meeting, and enhances multi-agent collaboration when combined with LLMs.
Key Results
- On IoA-Suite, the best model reaches only 49.5% F1, highlighting the difficulty of detecting hidden divergence due to lack of private context.
- MCQ probing significantly outperforms direct prompts, with an average F1 increase of 13.5%, confirming behavioral signals as effective indicators.
- IoA-Prober-8B trained with reinforcement learning achieves 51.8% F1, surpassing baseline models by 23.8%, and detects nearly 3 hidden disagreements per real meeting, validating practical utility.
Significance
This pioneering work formalizes the 'illusion of agreement' phenomenon in computational dialogue, emphasizing the importance of behavioral signals and private cognition. It addresses a core challenge in dialogue understanding—detecting unspoken misalignments—thus advancing explainability and reliability in AI systems. The approach offers promising applications in enterprise collaboration, virtual assistants, and remote meetings, enabling early warning of misunderstandings and improving communication efficiency.
Technical Contribution
The paper introduces a novel diagnostic question generation framework that converts hidden divergence into observable behavioral signals. It constructs IoA-Suite for systematic evaluation, and develops the IoA-Prober-8B model trained with reinforcement learning to optimize detection. The framework incorporates private context analysis, emphasizing the role of cognitive inference in dialogue misalignment detection, and sets a new standard for verifiable, behavior-based detection methods.
Novelty
First to define hidden divergence as a behavioral signal via generated diagnostic MCQs, moving beyond explicit friction cues. Introduces a comprehensive dataset and evaluation protocol for this problem, emphasizing private context inference. The combination of behavior-based detection and RL training for large models marks a significant innovation in dialogue misalignment research.
Limitations
- Detection heavily relies on private context, which is often unavailable or incomplete in real-world scenarios, limiting practical deployment.
- Synthetic data, despite controlled ground truth, may not fully capture the complexity of natural conversations, affecting real-world applicability.
- Current models struggle with multi-turn, complex dialogues, indicating the need for more advanced reasoning and multi-modal integration.
Future Work
Future directions include integrating multi-modal signals like speech and facial expressions, developing more robust private context inference, and designing interactive systems that actively probe for hidden misunderstandings during ongoing conversations. Expanding datasets with real-world data and improving model scalability are also key goals.
AI Executive Summary
This research tackles the subtle yet critical problem of hidden disagreements in collaborative dialogue, termed the 'illusion of alignment' (IoA). Even when participants appear to agree on surface, underlying goals, assumptions, or plans may diverge silently, leading to potential failures in coordination. Traditional detection methods relying on explicit signals like hedging or contradiction fall short, as these signals are absent when participants believe consensus has been reached. To address this, the authors propose a behavioral signal-based approach, generating diagnostic multiple-choice questions (MCQ) that reveal divergent answers as evidence of hidden disagreement.
A comprehensive dataset, IoA-Suite, was constructed by synthesizing dialogues across five task types and six domains, embedding latent divergences that are invisible to surface-level analysis. Evaluation of nine models, including GPT-5.4 and open-source alternatives, showed that even the strongest models only achieved 49.5% F1, primarily due to the lack of private context information. Incorporating private context significantly boosted performance, confirming the importance of cognitive inference.
Building on this, the authors trained IoA-Prober-8B using reinforcement learning, which improved detection F1 to 51.8%. In real meetings, this model successfully surfaced nearly three unvoiced disagreements per session, validated by participants. When paired with large language models in multi-agent settings, the system enhanced downstream task performance, demonstrating practical benefits.
Overall, this work pioneers a behavior-based framework for detecting hidden misalignments, bridging the gap between theoretical cognitive inference and practical dialogue systems. It highlights the critical role of private context and active probing, paving the way for more reliable, interpretable AI in collaborative environments. Limitations include reliance on private context availability and synthetic data, but future work aims to incorporate multi-modal signals and real-world data to further improve robustness and applicability.
Deep Analysis
Background
对话误差检测经历了从显性信号（如修正、犹豫）到隐性信号（如认知偏差、潜在误解）的演变。早期研究如Markowska等（2023）关注共享认知模型，后续工作如Nath等（2025）引入行为特征分析，但大多依赖显性标志。近年来，LLM在认知推理方面取得突破，但对隐藏误差的检测仍缺乏系统方法。本论文将对话中的“错觉一致”现象正式定义为潜在分歧，强调行为信号的重要性，填补了理论与实践的空白。
Core Problem
核心问题在于如何有效检测对话中未被显性表达的潜在分歧。现有模型多依赖显性信号，无法捕捉私有认知偏差，导致“错觉一致”现象难以识别。这不仅影响对话理解的准确性，也限制多智能体系统的协作效率。挑战在于：1）隐性分歧缺乏显性标志；2）模型难以推理参与者私有上下文；3）缺乏可验证的检测机制。这些问题阻碍了对话系统在实际应用中的可靠性提升。
Innovation
本论文的创新点包括：1）提出基于行为信号的诊断机制，将隐藏分歧转化为可验证的多项选择题，突破传统显性信号依赖；2）构建IoA-Suite数据集，涵盖多任务多领域，提供系统评估平台；3）引入私有上下文信息分析框架，强调认知推理在检测中的作用；4）训练强化学习模型IoA-Prober-8B，显著提升检测性能。这些创新共同推动了对话隐性误差检测的理论和实践发展。
Methodology
- �� 定义潜在分歧为对话结束后参与者对目标、假设或计划的不同理解。• 构建IoA-Suite数据集，通过合成对话植入潜在分歧，确保Ground Truth的可控性。• 设计多项选择题（MCQ） probing机制，从对话中自动生成诊断题，参与者独立回答，答案差异即为潜在分歧。• 采用模型评估协议，包括直接提示和MCQ probing两种策略。• 训练强化学习模型IoA-Prober-8B，利用F1作为奖励，优化问答生成质量。• 结合私有上下文信息，分析模型性能瓶颈，验证认知推理的重要性。
Experiments
采用IoA-Suite合成对话，覆盖五类任务（如规划、设计评审、故障排查）和六个领域（如科研、医疗、法律）。模型包括GPT-5.4、Gemini-3.1-Pro、Qwen-3.8B等，比较直接提示和MCQ probing两种策略。指标为F1、精确率、召回率，进行ablation验证私有上下文对性能的影响。还在真实会议中测试模型，验证其检测潜在分歧的能力。训练采用强化学习，优化问答生成的平衡性，确保检测的可靠性。
Results
模型在IoA-Suite中的最高F1仅为49.5%，远低于理想值，显示检测难度大。MCQ probing显著提升性能，F1平均增加13.5%，验证行为信号的有效性。引入私有上下文信息后，模型F1提升超过20%，说明私有认知信息是关键。训练的IoA-Prober-8B模型在测试集F1达51.8%，比基础模型Qwen3-8B提升23.8%，在真实会议中成功检测出每场2.89个未表达的分歧，验证了实用性和泛化能力。
Applications
该方法适用于企业内部协作、智能助手、远程会议监控等场景，有助于提前识别潜在误解，优化沟通效率。未来可结合多模态信息（如语音、表情）增强检测能力，推动智能对话系统的可信性。长远来看，该技术有望实现自动化的潜在误解预警，提升多智能体系统的协作质量，推动人机交互的智能化发展。
Limitations & Outlook
当前模型对私有上下文的依赖较强，实际应用中私有信息难以全面获取，影响检测效果。合成数据，虽然保证Ground Truth，但在自然对话中存在偏差。模型在多轮复杂对话中的表现仍有限，未来需结合更丰富的推理机制和多模态信息，提升鲁棒性和泛化能力。
Plain Language Accessible to non-experts
想象你和朋友在玩拼图游戏。你们都知道要拼出一幅漂亮的画，但有时候你们的理解不一样，比如你觉得“蓝色”是天空的颜色，而朋友觉得是水的颜色。表面上，你们都说“我同意”，但其实心里有不同的想法。这就像对话中的“假装一致”，大家都觉得没问题，但实际上有误会。研究人员设计了一种特别的问答游戏，让你们回答一些关于拼图的问题，比如“你觉得天空的颜色在哪一块？”不同的答案就能告诉你们其实有误会。这样可以提前发现问题，避免拼错。这个方法就像在对话中用问题揭示隐藏的想法，帮助人们更好地理解彼此。它能让团队合作更顺畅，也能让智能机器人更聪明，知道人们心里在想什么。
Abstract
Collaborative dialogue can end with apparent agreement while participants still differ on goals, assumptions, or execution plans, creating an \textbf{illusion of alignment (IoA)}. A real-user study across 18 meetings confirms that IoA arises routinely in human collaboration. Yet IoA poses a paradox: if participants were aware of such disagreements, they would already be explicit; if not, they cannot articulate them when asked, leaving IoA invisible to both participants and observers. In this work, we make IoA detectable by generating diagnostic multiple-choice questions whose divergent answers across participants provide direct behavioral evidence of hidden disagreement. We construct \textbf{IoA-Suite}, a dataset and evaluation protocol for detecting hidden disagreement, spanning five task types and six domains. We find that even the best model attains only 49.5\% F1, with the bottleneck traced to private context that the dialogue does not surface. We then train \textbf{IoA-Prober-8B} based on IoA-Suite, reaching 51.8\% F1 on IoA-Suite. Across the aforementioned 18 real meetings, it surfaces 2.89 hidden disagreements per meeting that participants confirm they had not voiced, transferring to live human dialogue. Further, in multi-agent collaboration, pairing IoA-Prober-8B with LLM agents improves downstream task performance on BigCodeBench-Hard and HiddenBench.