Referential ambiguity and clarification requests: comparing human and LLM behaviour

In this work we examine LLMs’ ability to ask clarification questions in task-oriented dialogues that follow the asynchronous instructiongiver/instruction-follower format. We present a new corpus that combines two existing annotations of the Minecraft Dialogue Corpus — one for reference and ambiguity in reference, and one for SDRT including clarifications — into a single common format providing the necessary information to experiment with clarifications and their relation to ambiguity. With this corpus we compare LLM actions with original human-generated clarification questions, examining how both humans and LLMs act in the case of ambiguity. We find that there is only a weak link between ambiguity and humans producing clarification questions in these dialogues, and low correlation between humans and LLMs. Humans hardly ever produce clarification questions for referential ambiguity, but often do so for task-based uncertainty. Conversely, LLMs produce more clarification questions for referential ambiguity, but less so for task uncertainty. We question if LLMs’ ability to ask clarification questions is predicated on their recent ability to simulate reasoning, and test this with different reasoning approaches, finding that reasoning does appear to increase question frequency and relevancy.

Introduction. Large Language Models (LLM) are much maligned for their tendency to act presumptively, “hallucinating” in the absence of knowledge. Until the recent advent of reasoning orientated LLMs (i.e. models deliberately fine tuned with reasoning as an objective such as DeepSeek-R1 (Guo et al., 2025)), models struggled asking clarification questions and rarely proactively sought missing information (Deng et al., 2023; Li et al., 2022, 2024). Prior works have tested how LLMs respond to uncertainty, and proposed benchmarks (Zhang et al., 2024). However, this remains a challenge.

This is perhaps further complicated as clarification is a conversational strategy applied sparingly by humans (Purver et al., 2003; Rodríguez and Schlangen, 2004; Rieser and Moore, 2005). Certain situations promote greater clarification question usage; for example, situations in which information is asymmetric, and which concerns a task requiring information seeking. This happens to be a popular paradigm for tasks created with the objective of soliciting dialogue (sometimes referred to as instruction giver/instruction follower) and also, in recent years, for studying clarification questions (Chi et al., 2020; Madureira and Schlangen, 2023a; Testoni and Fernández, 2024; Shen and Lourentzou, 2023). In this work we look at the Minecraft Dialogue Corpus (Narayan-Chen et al., 2019), a task orientated, grounded corpus that follows this paradigm. We select this corpus as it has benefited from multiple separate annotation efforts (Thompson et al., 2024; Madge et al., 2025) that extend its already richly structured offering, with useful supplemental information that can inform the experiments undertaken in this work. We focus on one particular area of linguistic uncertainty, referential ambiguity. There has been a long standing interest in reference, with ambiguity featuring as an interest in the first popular corpora (Pradhan et al., 2012). Our first contribution of this work is combining prior annotation efforts providing annotations for clarification questions (Thompson et al., 2024) and reference (Madge et al., 2025) into a single aligned corpus in the MMAX format (described in Section 3). This annotation supports our next contribution, a comparison of how LLMs and humans resolve uncertainty. We ask, “does referential ambiguity really trigger clarification requests from humans, and is this different for LLMs?”. We look at both the annotated instances of linguistic ambiguity, and the original clarification questions, as posed by human interlocutors for correlation. We test both against different LLM based approaches. One proposed approach to improving clarification questions with LLMs is through adding further reasoning capabilities with variations on the Chain of Thought approach (Deng et al., 2023). Despite extensive testing in epistemic, aleatoric, linguistic uncertainty (Ortega-Martín et al., 2023) and proposed benchmarks (Zhang et al., 2024), it remains somewhat unclear how effective LLMs are in identifying uncertainty and even more so how LLMs may consistently generate the relevant questions to address it. For our second research question, we ask if the ability to ask a useful clarification question, or indeed judge when to ask a question, is based on a model’s ability to simulate reasoning. We test this hypothesis with an experiment, comparing models that were trained to include reasoning and prompt engineering strategies for inducing reasoning at test time against ordinary models/methods. For our final contribution, we look further into human reasoning and its constituent parts, with a discussion on how these may affect LLMs seemingly emergent ability to ask some clarification questions, and liken this to human reasoning.

Related work. There has been a great interest in clarification questions in the literature on dialogue systems going back at least twenty years (Purver et al., 2003; Schlangen, 2004; Gabsdil, 2003). More recently, there has been extensive work in modern Natural Language Processing modelling clarification question generation or indeed when to ask them (Majumder et al., 2021; Aliannejadi et al., 2019; Kiseleva et al., 2022). This section will primarily focus on prior works that target clarification question with overlap to our specific goals (i.e. task orientated dialogue with situated and/or embodied agents). Previous works have gathered or annotated datasets in situated dialogue with clarification questions. For example, (Gervits et al., 2021) gather a corpus (HuRDL - Human-Robot Dialogue Learning) and annotate clarification questions in a dialogue gathered from human participants in a robot situated tool gathering task. (Gella et al., 2022) annotate dialogue acts in the TEACh (Task-driven Embodied Agents that Chat) dataset (Padmakumar et al., 2022); the product of a task that has human participants collaborating to perform household tasks in a virtual house environment. A particularly popular task/dataset for this is Co- Draw (Kim et al., 2017). The CoDraw task (Kim et al., 2017) is similar to the previously discussed Minecraft task, in that an instruction giver communicates with an instruction follower to collaboratively reach a goal. As opposed to constructing a 3D voxel based structure, they recreate a scene formed of clipart images. (Madureira and Schlangen, 2023b) annotate this dataset with clarification questions. Previous works have also compared when humans and models would ask clarification questions (Testoni and Fernández, 2024) use the aforementioned CoDraw dataset to investigate this the relationship between model uncertainty, and human clarification questions based on task properties (e.g. size, orientation, position etc.). The presence of a clarification question is used as the measure of measure uncertainty, and they use logistic regression to see if they can predict this. The Minecraft Dialogue Corpus (Narayan-Chen et al., 2019) used in this work is different, in that rather than referencing direct objects, continuously changing abstract shapes are created and manipulated during the dialogue. We expand on this further in the following section. There has been other work using Minecraft-like environments as a test-bed for the study of clarification questions in dialogue.

Method. In this section we will motivate and describe our effort to add MSDC information to MDC-R to produce a new version of the corpus combining both types of annotation. We identify two types of utterances or phrases may provoke clarification requests. Firstly, utterances that have been annotated as the subject of a clarification or confirmation request in the dialogue (typically related to task orientated uncertainty , and secondly instances of referential ambiguity. To support our experimentation, and investigation of any possible relationship between the two, we present a corpus that merges two existing corpora that identifies these. This is a combination of the previously discussed MDC-R (Madge et al., 2025) (providing reference annotations for MDC), and MSDC corpora (Thompson et al., 2024) (providing clarification questions for MDC), permitting convenient examination of reference and more specifically types of referential ambiguity aligned with clarification questions. We automated the merge of these two corpora, through use of a script that operated at token level to produce a common method of addressing and aligning the respective segments in each. We add a new MMAX1 layer, referred to as SDRT. Each MMAX markable in this layer represents an Elementary Discourse Unit and the relations between those markables are represented by a to attribute on each markable, with the related markables unique ID. Crucially, these relationships describe links between utterances and their clarification questions. This combined format allows parsing and examination of adjacent reference in MMAX (shown in Figure 1) and clarification (shown in Figure 2). Reference annotation has many parameters, with each relationship holding many attributes. Whilst the SDRT annotations exhaustively cover all of the original MDC dataset (Narayan-Chen et al., 2019), MDC-R covers a subset of 100 dialogues (some more detailed descriptive statistics taken from MDC-R (Madge et al., 2025) are given in Table 1). As such, our corpus will be limited to the same 100 dialogues.

Using the combined corpus, we have counted instances of utterances exhibiting certain attributes that may motivate a question. Firstly, using annotations originating from MSDC, we look at the counts of confirmation questions and clarification questions occurring in the original human dialogue. Secondly, we count specific instances of referential ambiguity, using annotations originating from MDC-R. These are instances of discourse deixis - relating to parts of the discourse (e.g. “as I said earlier”), and spatial deixis in a real world space - typically our voxel world environment (e.g. “next to that block”). Table 2 shows the frequency of these different instances of utterances that may motivate a question as a percentage of all utterances in the selected subset of the corpus.

To test how LLMs perform clarification questions, we first discover points in the conversation that may require clarification, then we sample from various LLMs with different approaches, feeding the context of the conversation up to the appropriate point. Examples of our prompts are given in the Appendix (see Section A.1). A system prompt describes the nature of environment and it’s constraints. This differs slightly between architects and builders, in that architects can see the target structure, and the builder’s system prompt specifies the required JSON response format necessary to encode their resultant actions or expect the world state (this experiment is text only, with the world state/actions encoded in JSON - no images are used). For the chain-of-thought treatments, we supplement these system prompts as follows. The zero-shot chain-of-thought approach simply adds, “Think step by step” to the system prompt. The one-shot approach follows the system prompt with an example exchange that incorporates thinking.These approaches are deliberately selected to compare reasoning based models (e.g. llama3.2, DeepSeek-R1 (Guo et al., 2025)) and sampling methods (llama2:13b-COTZERO, llama2:13b-COTONE) vs. non-reasoning (llama2:13b). We compare the approach taken by LLMs to the approach originally taken by humans.

Discussion. We show here in this section that whilst linguistic ambiguities rarely trigger direct human clarification, despite the fact they may translate to uneconomical task action, LLMs can and do sometimes respond with targeted clarification questions.

Our experiments would appear to show that the recent advent of reasoning in models has the emergent benefit of allowing models to ask clarification questions. In this section, following our hypothesis that clarification is dependent on an ability to reason, we look at clarification questions in relation to reasoning through the lens of human psychology, and where available, assessments of LLM abilities to perform the required components of related human reasoning abilities Examining the role of reasoning in clarification, knowing when to ask a question, requires reflection on the gaps in ones own knowledge, or a higher order of thought, referred to in human psychology as Metacognition (Flavell, 1979). LLMs originally lacked any awareness of gaps in their knowledge, acting presumptively, leading to “hallucination” where, albeit often grammatically valid, a model’s output would be factually incorrect or possibly nonsensical. This is most commonly due to a lack of knowledge (Zhang et al., 2023), or perhaps the inability to reason when knowledge is absent. Metacognition is the ability to reflect on held knowledge by self questioning. The Metacognitive capabilities of LLMs have been explored in previous works in relation to reasoning (Didolkar et al., 2024). Metacognitive prompting that, self questions to enhance reasoning (Wang and Zhao, 2023) has been explored in LLMs, as has self-questioning with the goal of reducing hallucination (Dhuliawala et al., 2023). A model knowing whether it has the applicable knowledge or skill to proceed, or whether to direct its process to clarification, could be seen as metacognitive regulation. When an agent is working in collaboration with other interlocutors, clarification may be dependent on discovering other parties knowledge, abilities or attitudes and approach to a task. This crucial component, in psychology, is an aspect of reasoning known as Theory of Mind (Premack and Woodruff, 1978), which relates to reasoning about other participants belief states. There is some evidence to suggest, as a consequence of simulated reasoning ability, LLMs may now be able to simulate this also (Kosinski, 2023). This has been explored with LLMs in Mindcraft, which is a collaborative task in which the players have separate skills and must negotiate to reach a common goal (Bara et al., 2021). This is particularly important for LLMs in referential communication (Sidera et al., 2018)

Conclusion. To conclude, we find that in our conversations human participants do not commonly ask a clarification question when language is ambiguous. In the majority of these cases of linguistic ambiguity, the ambiguity appear deliberate in the interest of conversational efficiency. Consequently, there may be little utility to asking a question in many of those situations. (The resultant utility of a clarification question is not examined in this work, and may be the subject of a future work. ) We did however identify one situation in which that uncertainty did propagate to create future issues. Regardless of conversational efficiency, we do find that LLMs, particularly reasoning orientated ones, are capable of asking relevant clarification questions under those circumstances. On the topic of the originally posed human clarification questions, we found they largely followed a specific pattern and strategy not adopted by the LLMs. That was, to perform actions then use a clarification question to verify they were correct. These instances of human clarification questions largely relate to task based ambiguity rather than linguistic ambiguity. Across all instances, we find a greater tendency of reasoning orientated approaches to pose clarification questions and find that this can be somewhat matched at test time, with methods such as COT.

Limitations. Another limitation to our experiment. It’s challenging to communicate a sense of perspective to the LLMs. As a consequence, some instructions do not make sense, e.g.: “A: can you come to the side of the structure so you have a side view”; “B: left or right”; “B: forward to my right or in front of me?”; “B: this perspective ?”.