--- title: Learning as Conversation url: https://www.emergentmind.com/topics/learning-as-conversation type: topic --- # Learning as Conversation Searching arXiv for the provided IDs and topic to ground the article in the cited literature. Learning as conversation is a family of theories, system designs, and learning algorithms that treats dialogue not as a superficial interface over an already solved task, but as the medium through which understanding, supervision, adaptation, and knowledge construction occur. In one influential formulation, conversational search is a user-system cooperation “with the objective to solve an information-seeking task,” in which the user learns content, search strategies, and self-regulation while the system learns from actions, needs, context, knowledge state, and satisfaction signals [2001.02912]. Across later work on teachable agents, dialog-based language learning, conversation-based tutoring, conversational emotion recognition, and dialogic pedagogy for LLMs, the same principle recurs: learning is produced by iterative exchange, and the quality of that exchange constrains what can be learned [1909.13443][1604.06045][2403.14071][2506.19484][2602.14910]. ## 1. Conceptual foundations A central conceptual move in this literature is to broaden what counts as learning. Conversational-search work explicitly rejects reducing success to “did the user find the right document?” and instead includes procedural learning, knowledge-based learning, task-level success, metacognitive self-regulation, and later transfer or generalization across contexts [2001.02912]. On this view, dialogue matters because information needs are often not fully formed at the outset, especially in exploratory or educational settings; conversation incrementally surfaces and refines them. The user’s utterances, clarifications, corrections, and modality choices also provide evidence about expertise, uncertainty, cognitive load, and satisfaction, allowing the system to guide, scaffold, simplify, elaborate, or redirect [2001.02912]. Educational theory gives this conversational view a longer genealogy. A recent review of dialogic pedagogy for LLMs maps conversational AI onto Vygotsky’s sociocultural learning, scaffolding and the Zone of Proximal Development, the Socratic method, and Laurillard’s conversational framework, while also stressing that LLMs do not automatically satisfy those theories merely by generating fluent text [2506.19484]. Dialogue in this tradition is valuable because it externalizes reasoning, makes misunderstandings visible, and supports co-construction rather than one-way transmission. Position work in 2026 radicalizes the same idea by arguing that robust reasoning emerges from linguistic self-reflection internalized from high-quality social interaction, so that private reasoning is the afterlife of public dialogue rather than an independent capacity [2602.14910]. Empirical human-LLM work reaches a compatible conclusion in more restrained terms: in political learning dialogues, “learning from LLMs is an interactional achievement, not a uniform outcome of better explanations,” with confidence gains and knowledge gains mediated by different engagement processes [2601.07796]. A plausible implication is that “learning as conversation” now denotes both a pedagogical principle and a modeling hypothesis: cognition improves when explanation, critique, and reflection are organized dialogically. ## 2. Pedagogical roles and instructional dialogue One major line of work studies how conversational roles reshape pedagogy. Chhibber and Law’s teachable-agent paper contrasts teaching agents, peer agents, and teachable agents, and aligns itself with the teachable category: the agent is deliberately less knowledgeable, elicits the student’s explanations and uncertainties, and uses “cognitive, affective, and meta-cognitive interventions” such as asking for examples, rephrasings, repeated lessons, and rationales [1909.13443]. Their system, Curiosity Notebook, has a simple but revealing interaction pattern: a student highlights a sentence in a reading, the agent asks a question about it, and the student responds. The broader environment displays the agent’s learning progress, applies sanity checks to student input, and maintains a hidden ground-truth model of what the student is supposed to learn [1909.13443]. Conversation here functions as self-explanation, monitoring, and repair rather than content delivery. A complementary configuration reverses the direction of expertise in a different way. In “Learning as Conversation: Dialogue Systems Reinforced for Information Acquisition,” a teacher bot has access to a source passage, a student bot does not, and the objective is for the teacher to convey the passage through multi-turn dialogue so that the user can learn without reading the source directly [2205.14748]. The dialogue must be simultaneously informative and attentive: it should cover as much of the passage as possible while still responding coherently to the learner’s utterances. This is not ordinary question answering, because the system is optimized for conversational knowledge transfer rather than isolated answer extraction [2205.14748]. A more explicitly personalized tutoring architecture combines diagnostic modeling with prompt-based dialogue. The proof-of-concept SAT Writing tutor of Kim and colleagues uses an onboarding survey, pre-test, tutoring session, and post-test in a cycle over the concepts Pronouns, Punctuation, and Transitions [2403.14071]. Its student model tracks cognitive state, affective state, and learning style; the conversation-based tutor then uses GPT-4 with prompt engineering that incorporates student assessment outcomes, IRT-selected exercises, and session summaries. The tutoring policy is deliberately dialogic: one question per turn, no immediate revelation of correct answers, deeper questioning after correct responses, and summary-and-check transitions before moving on [2403.14071]. In this configuration, tutoring is not a static script but a repeated loop of diagnosis, conversation, and model update. ## 3. Multimodality and embodied conversational learning A distinctive claim in the conversational-search literature is that speech-only interaction is insufficient for serious learning support because human thought is “spatialized,” and spoken conversation is not itself a spatial modality [2001.02912]. The proposed remedy is multimodality: speech, text, digital pen or handwriting, touch, tangible manipulation, gaze, gesturing, and multimedia output. Each modality contributes differently. Pen input preserves spatial relations and leaves a durable trace; touch and manipulation support embodied interaction with 3D spatial information and are described as especially beneficial for procedural learning, such as learning to drive in a simulator; multimodal fusion also improves robustness because users can switch modalities when one is error-prone, and systems can reduce recognition errors by combining multiple information sources [2001.02912]. The same work also frames multimodality as an accessibility issue, noting potential support for blind and hearing-impaired students, lower-performing students, and non-native speakers [2001.02912]. Embodied group dialogue in augmented reality extends this claim from multimodal input to situated, social practice. ConversAR places two embodied humanoid LLM agents in a learner’s physical environment through a Unity application on Meta Quest 3, with GPT-4o used both for agent conversation and for a separate Moderator LLM that selects which agent should respond after the user speaks [2505.24000]. The system includes live captions, scene-context conditioning, and agent-to-agent as well as agent-to-learner exchange. In a study with 10 adult university undergraduates practicing Spanish, perceived speaking anxiety was lower for ConversAR than for in-person group methods, with means of 3.70 versus 5.10 on a 7-point scale, while engagement means were 5.40 and 5.20 respectively [2505.24000]. At the same time, the paper reports no coded instances of social engagement in the form of backchannels, indicating that embodied multi-party interaction can lower anxiety without yet reproducing the full reciprocity of human group conversation [2505.24000]. This suggests that embodiment and multimodality can change the conditions of conversational learning even when they do not fully solve its social realism. ## 4. Conversation as supervision, representation, and optimization Several technical papers turn conversation itself into the supervisory signal. Weston’s dialog-based language learning formulates interaction as tuples \((x,c,a,\bar{x})\): the last teacher utterance, dialogue context or memory, the learner’s action, and the teacher’s subsequent response [1604.06045]. Instead of relying only on reward, the model can learn by predicting the teacher’s next utterance. In the bAbI setting, the forward-prediction model reached 100%, 98%, and 99% on Task 7 at \(\pi_{acc}=0.5, 0.1, 0.01\), showing that a learner can answer correctly without reward-based supervision when the teacher’s responses remain informative [1604.06045]. The core idea is that conversational feedback—correction, hint, explanation, or approval—contains enough structure to supervise the latent task. Grounded interactive learning pushes this further by combining imitation and reinforcement inside dialogue. “Listen, Interact and Talk” places an agent in a visual environment, lets a teacher ask grounded questions or provide statements, and then trains the learner from both the teacher’s next sentence and scalar reward [1705.09906]. The joint model decisively outperformed reward-only and imitation-only baselines: in the mixed configuration it achieved 98.9% on compositional generalization and 97.5% on knowledge transfer, whereas Reinforce remained at 0.0% in both settings [1705.09906]. Here, conversation is not only a source of labels; it is the sequential process through which grounded language competence is acquired. Document-grounded teaching dialogue has also been optimized directly. The teacher bot in “Learning as Conversation: Dialogue Systems Reinforced for Information Acquisition” is fine-tuned by self-play with a frozen student bot using a mixed reward \(R=\beta R_{cov} + (1-\beta)R_{coh}\), with \(\beta=0.7\) in the full model, where coverage measures incremental passage information introduced by a turn and coherence estimates response appropriateness [2205.14748]. On CNN/DailyMail, this moved human QA from 28.12 to 52.5 and overall human-rated quality from 1.87 to 2.53 relative to the pretrained DoHA baseline [2205.14748]. The system therefore treats conversation as an optimization target for information acquisition rather than merely an output format. Other work formalizes conversation as a structure over which representations should be learned. Conversation Kernels models online discussions as trees and learns which parts of the tree constitute task-relevant context, achieving Acc 0.7896 and macro-F1 0.7804 for “informative” comments with its ancestor-sibling-children kernel, well above the best RoBERTa baseline at Acc 0.7059 and macro-F1 0.7058 [2505.20482]. Theta, the three-stream hierarchical conversation model, induces DA-like turn labels and HMM sub-dialogue states without manual annotation, and raises CraigslistBargain test accuracy from \(54.1 \pm 2.4\) to \(66.1 \pm 1.0\) [2205.12244]. In emotion recognition in conversation, hybrid curriculum learning orders training from easy to hard at both conversation and utterance levels and improves TODKAT from 43.12 to 46.11 on EmoryNLP, while emotion-anchored contrastive learning reaches weighted F1 scores of 70.41 on IEMOCAP, 67.12 on MELD, and 40.24 on EmoryNLP by explicitly separating confusable emotion anchors [2112.11718][2403.20289]. Related work learns conversation-level characteristics for ASR through role and topic latent variables, reporting a maximum 12% relative CER reduction, and learns dialogue concepts by masking context words and focusing on those with high mutual information for response generation [2202.07855][2003.04976]. Across these models, a shared premise is visible: utterances are not independent objects but nodes, stages, or latent states inside an evolving exchange. ## 5. Personalization, student modeling, and continual adaptation A second major axis of the literature concerns who the conversation is with and how the system should adapt to them. “Towards Teachable Conversational Agents” treats text classification as teacher-learner dialogue: a web-based bot asks about which words are internally relevant, internally irrelevant, or externally relevant to a news category, and updates an interactive Naive Bayes classifier from those conversational cues [2102.10387]. Across 60 crowdworkers, more dialogues were associated with better agent F1 \((\beta = 0.001,\ t(56)=3.68,\ p<0.001)\), and more frequent testing by participants also improved F1 \((\beta = 0.005,\ t(56)=4.69,\ p<0.001)\) [2102.10387]. Yet the same study also shows the fragility of conversational supervision: the best teacher modestly improved over the supervised baseline, while the worst teacher degraded performance. A concrete implication is that teachability requires mechanisms for filtering, grounding, or weighting teacher input. Personalization can also be learned relationally rather than only from the target learner’s own history. In few-shot personalized conversation with an auxiliary social network, each speaker is treated as a task with \(K=10\) support examples, and a Task Aggregator uses mutual-follow relations to provide a prior for new speakers [2105.10323]. The motivating empirical observation is that average response similarity between neighboring speakers is 0.47, compared with 0.38 for random speaker pairs and 0.50 for the same speaker [2105.10323]. On the Weibo benchmark, the full model reaches BLEU-4 \(=4.145\), NIST \(=0.940\), CIDEr \(=0.238\), Dist-1 \(=0.169\), Dist-2 \(=0.530\), Grd-F1 \(=0.024\), and TokSim \(=0.062\), outperforming the meta-learning baseline PAML on appropriateness and speaker consistency [2105.10323]. This suggests that conversational behavior can be personalized not only from what a learner or speaker has already said, but also from the structure of their conversational neighborhood. The lifelong-learning agenda generalizes this adaptivity from individual sessions to deployment as a whole. LINC—Lifelong INteractive learning in Conversation—argues that a new learning task \(T_{N+1}\) arises when the system encounters something it cannot currently handle, and that it must then formulate a dynamic interaction strategy \(\mathcal{I}\), acquire ground-truth data through conversation or demonstration, and learn incrementally during use [2009.10750]. The proposed targets are new world knowledge, new language expressions grounded to actions, and new conversational skills. To protect correctness, newly acquired items are placed in an unverified knowledge buffer and are considered trustworthy only after verification from \(K\) different random users [2009.10750]. This is a strong version of the “learning as conversation” thesis: dialogue is not only a site of tutoring or retrieval, but the mechanism through which the system discovers its own missing tasks and teaches itself on the job. ## 6. Evaluation, limitations, and unresolved questions Because conversational learning is richer than one-shot prediction, evaluation has repeatedly had to expand. Conversational-search work explicitly states that standard IR metrics are too narrow for settings concerned with learning, self-regulation, accessibility, and reduced cognitive load [2001.02912]. Document-grounded teacher bots therefore combine coverage metrics such as ROUGE and SummaQA with coherence classifiers and human evaluations of readability, coherence, and whether a hidden passage can be reconstructed from the dialogue [2205.14748]. In personalized tutoring, however, the gap between personalization and measurable learning remains visible: the GPT-4-based SAT tutor produced a corpus of 74 dialogues and 1470 turns with an average of 19.87 turns per dialogue, but average student utterances were only 3.9 words long, and the reported IRT-based learning gains were \(-0.0753\) for Pronouns, \(0.0159\) for Punctuation, and \(-0.0102\) for Transitions [2403.14071]. The system personalized its prompts, but did not guarantee notable learning gains [2403.14071]. Human-LLM learning studies further show that outcomes are highly conditional. In 397 conversations from 152 participants discussing political issues, partial mediation fit better for confidence change than full mediation, with \(\chi^2_{diff}=18.35,\ df=1,\ p<.05\), whereas for knowledge gain full mediation was sufficient, with \(\chi^2_{diff}=0.41,\ df=1,\ p=.52\) [2601.07796]. Richer explanations supported confidence partly through reflective insight, but their effect on knowledge depended entirely on cognitive engagement; longer conversations benefited primarily reflective or high-efficacy users [2601.07796]. This supports a general conclusion already anticipated in conceptual work: conversation is a condition for learning, not a guarantee of it. The literature also converges on several risks. Deeply human-centered multimodal modeling raises privacy concerns because it requires sensitive behavioral data [2001.02912]. Lifelong self-learning systems face the possibility of wrong or malicious user input, difficult knowledge revision, and the need to revise consequences of false beliefs already stored in a KB [2009.10750]. Dialogic pedagogy reviews warn that LLMs tend to provide direct answers instead of fostering co-construction, may hallucinate, and may reduce learner agency if they remove productive struggle too quickly [2506.19484]. Embodied systems can lower anxiety yet still fail to reproduce the social reciprocity of human conversation, as shown by the absence of coded social engagement in ConversAR [2505.24000]. Position work pushes the normative point further by claiming that “Dialogue Quality is the New Data Quality,” implying that future progress depends less on conversation per se than on whether conversational environments contain rigor, diversity, repair, and productive friction [2602.14910]. Taken together, these studies define learning as conversation neither as a metaphor nor as a single method, but as a broad research program. Its strong form holds that dialogue is a developmental substrate for reasoning, teaching, retrieval, personalization, and continual adaptation; its weaker form holds that many learning systems improve when they treat utterances as contextually embedded, pedagogically structured, and interactionally diagnostic. The unresolved question is no longer whether conversation matters, but how to engineer, model, and evaluate conversations whose structure genuinely produces understanding rather than only the appearance of it. Source: https://www.emergentmind.com/topics/learning-as-conversation