Abstract
Asking for clarification is fundamental to effective collaboration. An interactive artificial agent must know when to ask a human instructor for more information in order to ascertain their goals. Previous work bases the timing of questions on supervised models learned from interactions between humans. Instead of a supervised classification task, we wish to ground the need for questions in the acting agent’s predictive uncertainty. In this work, we investigate if ambiguous linguistic instructions can be aligned with uncertainty in neural models. We train an agent using the T5 encoder-decoder architecture to solve the Minecraft Collaborative Building Task and identify uncertainty metrics that achieve better distributional separation between clear and ambiguous instructions. We further show that well-calibrated prediction probabilities benefit the detection of ambiguous instructions. Lastly, we provide a novel empirical analysis on the relationship between uncertainty and dialog history length and highlight an important property that poses a difficulty for detection.
Introduction: The Challenge of Grounded Clarification
In collaborative environments, such as a human "architect" guiding an AI "builder" in a 3D Minecraft world, effective communication is paramount. However, human instructions are frequently underspecified or vague (e.g., "Put the red one there"). Traditional approaches to this problem involve training a separate "meta-model"—a binary classifier—to predict whether an agent should ask a question based on the current dialog state.
While functional, this approach has several drawbacks: it requires large amounts of labeled interaction data, lacks interpretability, and decouples the decision to ask for help from the agent's actual internal state and capabilities. This research proposes an alternative: grounding the decision to clarify in the agent's own predictive uncertainty. If an agent is unsure of the next action required to fulfill an instruction, that internal signal should trigger a clarification request.
Methodology: Measuring Model Confusion
The researchers employed a T5 encoder-decoder architecture adapted for the Minecraft Collaborative Building Task. The core idea is to treat ambiguous instructions as "out-of-distribution" (OOD) events that should naturally trigger higher uncertainty in a well-calibrated model.
To quantify this uncertainty, several metrics were evaluated:
- Negative Log-Likelihood (NLL): Measures the model's "surprise" at the most probable action.
- Entropy (ENT): Captures the spread of the probability distribution; high entropy indicates the model is distributing its confidence across multiple potential actions.
- Deep Ensembles: By training ten models with different initializations, the researchers could better calibrate the agent. Ensembles help prevent the "confidently wrong" behavior common in single neural networks.
- Beam Search Statistics: Uncertainty data was also extracted from the top-N hypotheses generated during decoding.
Experimental Setup and Datasets
The study utilized the IGLU (Interactive Grounded Language Understanding) framework, focusing on three datasets:
- Minecraft Dialogue Corpus (MDC): Real human-human collaborative building logs.
- IGLU-NLP / IGLU-MULTI: Standardized multi-turn dialogs paired with 3D grid-world states.
- IGLU-MINI: A novel evaluation set created by the authors. It consists of "minimal pairs" where clear instructions are edited to remove specific information (like color or quantity), allowing for precise testing of whether uncertainty spikes exactly when ambiguity is introduced.
Key Results: The Power of Calibration
The experiments demonstrated that predictive uncertainty is a viable signal for ambiguity detection, particularly when using ensembles.
Key findings include:
- Superior Detection with Ensembles: On the IGLU-MINI dataset, the ensemble model correctly identified ambiguity (via higher uncertainty) in 92-95% of color-ambiguous cases.
- Distributional Separation: Using Area Under the ROC Curve (AUC), metrics like KL-divergence and Entropy achieved scores between 0.56 and 0.65 in distinguishing clear from ambiguous turns.
- Interpretability: Token-level analysis revealed that entropy spikes occurred specifically on the tokens related to the missing information (e.g., the "color" slot in an action sequence), providing a clear link between linguistic ambiguity and model confusion.
Discussion: The History Confounder and Limitations
A critical insight from this work is the impact of dialog history length. The researchers found that as a conversation progresses, the model's internal uncertainty naturally increases, regardless of the current instruction's clarity. This "history confounder" makes it difficult to set a static uncertainty threshold for asking questions; a clear instruction late in a session might appear more "uncertain" than a vague one at the start.
Furthermore, while the model can identify when it is confused, the next step—determining what to ask—remains an open challenge. Future work will focus on normalizing uncertainty against history length and leveraging specific entropy spikes to generate targeted clarification questions (e.g., "Which color did you mean?").
Conclusion
This research demonstrates that an agent's internal predictive uncertainty, when properly calibrated through ensembles, can serve as a grounded signal for initiating clarification dialog. By moving away from separate "black-box" classifiers, we move toward agents that truly "know what they don't know," leading to more natural and efficient human-AI collaboration.
Summary updated