Conversational Recommendation

• Conversational recommendation is a dialogue-based system that uses multi-turn interactions to clarify user preferences and provide personalized suggestions.
• It combines dialogue management, natural language processing, and reinforcement learning to optimize both immediate engagement and long-term satisfaction.
• Recent architectures range from modular KG-enhanced models to unified encoder-decoder systems, showcasing diverse strategies for integrating conversation with recommendations.

Searching arXiv for recent and foundational work on conversational recommendation to ground the article in the provided literature. Conversational recommendation is a recommendation paradigm in which suggestions are produced through a multi-turn dialogue rather than a one-shot ranked list. In this setting, a conversational recommender system supports users in achieving recommendation-related goals through a multi-turn dialogue, combining recommendation, dialogue management, and natural language technologies. Across the literature, the object of optimization ranges from item choice to the full conversational strategy: what to ask, when to recommend, how to explain, and how to adapt recommendations to user feedback, constraints, and context (Jannach et al., 2022, Liu, 30 Jun 2025, Zhang et al., 21 May 2026).

1. Definition, scope, and interaction structure

A conversational recommender system is commonly defined as a goal-oriented, multi-turn dialogue system that recommends items by proactively interacting with users in natural language. This distinguishes it from conventional recommenders that are typically one-shot, reactive, and based primarily on historical behavior such as clicks, ratings, or purchases. In conversational recommendation, the system continuously interacts over multiple turns, clarifies and refines user preference through dialogue, proactively guides indecisive users, and must both recommend appropriate items and talk about them in fluent, coherent language (Jannach et al., 2022, Wang et al., 2022).

The literature describes a broad range of interaction forms. Early systems emphasized forms, critiquing, and knowledge-based advisors; more recent systems emphasize natural-language chatbots, end-to-end neural models, and large-language-model interfaces. The “grand challenge” view organizes this space around input/output processing, user modeling and recommendation reasoning, background knowledge, and dialogue management, and also enumerates rich user intents such as initial query, reformulate, provide preference, accept, reject, inquire, critique-feature, critique-add, and critique-similar, together with recommender actions such as request information, clarify question, recommend-show, recommend-explore, explain-introduction, explain-preference, and explain-suggestion (Jannach et al., 2022).

The same term also covers substantially different operational settings. Some systems are explicitly multi-turn and task-oriented, such as knowledge-graph and reinforcement-learning CRSs; some are single-turn but still conversational in the sense that they consume free-form spoken utterances and sentiment to drive category recommendations. The prototype in “Conversational Recommendation System using NLP and Sentiment Analysis” is a voice-driven, sentiment-aware recommender that maps spoken utterances to top-3 product categories, whereas “Optimizing Conversational Product Recommendation via Reinforcement Learning” treats conversational recommendation as an interactive, multi-turn dialogue in which the sequence and style of utterances matter for conversion, engagement, satisfaction, and compliance (Talegaonkar et al., 17 May 2025, Liu, 30 Jun 2025).

2. Sequential decision-making formulations

A central formalization models conversational recommendation as sequential decision-making. In the reinforcement-learning view, the interaction is an episodic Markov Decision Process with states $s_t$, actions $a_t$, transitions, and rewards. In the enterprise product-recommendation formulation, the state encodes the dynamic dialogue context using language-model embeddings over conversation history and contextual features; the action is an utterance choice or talk track; rewards combine short-term cues such as engagement and positive sentiment with long-term outcomes such as conversion, satisfaction, or survey scores; and the objective is to maximize expected cumulative reward,

$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$

This formulation makes the trade-off between asking more questions and recommending earlier explicit (Liu, 30 Jun 2025).

The same general idea appears in multi-round conversational recommendation over items and attributes. UNICORN formulates three coupled decisions—what attributes to ask, which items to recommend, and when to ask versus recommend—as a single unified policy-learning problem. Its state combines conversation history and a dynamic user-specific graph; its unified action space contains both attribute-ask and item-recommend actions; and its objective is

$$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$

The dynamic weighted graph updates user–item scores as accepted and rejected attributes accumulate, enabling the policy to compare attribute actions and item actions in the same action space (Deng et al., 2021).

Later work extends the single-conversation assumption. “Towards Multi-Subsession Conversational Recommendation” defines a session $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$ composed of multiple subsessions, where the current subsession may begin with vague interests and requires “activation attributes” $\mathcal{A}_{v^n}^*$ to elicit explicit preferences. MSCAA therefore combines a context-aware recommendation module, an attribute selection policy, and a conversation policy to manage asking and recommending across previous subsessions and current feedback (Ji et al., 2023).

Not all sequential formulations rely on explicit reinforcement learning. “Conversational Recommendation System with Unsupervised Learning” instead treats final orders as delayed rewards and learns user intention prediction and dialog state tracking from unlabeled conversations plus transactional outcomes, avoiding hand-labeled dialog acts and hand-written rules (Sun et al., 2016). Conversely, some systems remain deliberately simpler: the 2025 voice-based prototype has no multi-turn dialogue manager, no reinforcement learning policy, and no follow-up clarification strategy, but it still demonstrates sentiment-driven conversational adaptation through ranking inversion and local keyword-frequency updates (Talegaonkar et al., 17 May 2025).

3. Architectural paradigms

The field contains several recurring architectural paradigms, differing mainly in how tightly recommendation and dialogue are coupled, how items are represented, and how control is exercised.

Paradigm	Representative systems	Core idea
Modular KG-enhanced CRS	KBRD, KGSF, CR-Walker, EGCR	Separate recommender and generator, often linked through KG-derived representations
Unified encoder–decoder or PLM CRS	RecInDial, BARCOR, PECRS	One model handles recommendation and generation with shared contextual representations
Fully generative CRS	GCRS	A single autoregressive model generates intent, item identifiers, and response text
Retrieval-style CRS	BM25-based baseline	Conversations are queries and items are documents to be retrieved
LLM-orchestrated CRS	LLMCRS	An LLM manages sub-tasks and expert models, then generates the final response

Modular systems historically dominated the literature surveyed in the 2022 “grand challenge” review: separate recommenders and response generators, often with knowledge graphs and hand-designed interfaces between components (Jannach et al., 2022). BARCOR characterizes this modular tradition explicitly through systems such as KBRD and KGSF, arguing that their connections between recommendation and generation are often complicated and unintuitive. BARCOR therefore uses a single BART-based sequence-to-sequence model, with the encoder acting as recommender and the decoder acting as response generator, plus a lightweight movie knowledge graph called CORG (Wang et al., 2022).

RecInDial pursues a different unification strategy. It expands the vocabulary with item tokens, introduces a binary vocabulary pointer $I_{vp}$, and integrates recommendation into dialogue generation itself. Recommendation becomes part of the decoding process rather than a separate module plus copy mechanism, and the model is evaluated in an end-to-end manner by whether the final generated responses actually contain the correct items (Wang et al., 2021). PECRS pushes this unification further by representing items as natural-language descriptions, using a single frozen GPT-2 backbone with LoRA adapters to encode items, understand dialogue, retrieve and rerank candidates, and generate responses. Its training objective is a single-stage multi-task loss,

$$\mathcal{L} = \alpha \mathcal{L}_\text{recall} + \beta \mathcal{L}_\text{rerank} + \gamma \mathcal{L}_\text{gen},$$

with $\alpha = 0.15$, $\beta = 0.85$, and $a_t$0 (Ravaut et al., 2024).

Fully generative architectures attempt to eliminate the retrieval/generation split entirely. GCRS represents items as semantic IDs produced by residual quantization and inserts them directly into the token stream as

$a_t$1

It factorizes the system response as

$a_t$2

where $a_t$3 is response intent, $a_t$4 is the target item, and $a_t$5 is the natural-language response. This explicit factorization separates high-level planning from surface realization and supports constrained decoding for faithful item generation (Zhang et al., 21 May 2026).

LLMCRS adopts yet another structure. Rather than making the LLM the recommender, it uses the LLM as a controller over four stages—sub-task detection, model matching, sub-task execution, and response generation—and delegates specialized computation to expert models such as KBRD, KGSF, TG-ReDial, or database lookup (Feng et al., 2023). This orchestration perspective contrasts with both retrieval baselines and end-to-end generative models.

4. Knowledge sources, supervision, and item representations

A major dividing line in conversational recommendation concerns the knowledge source used to represent items and conversations. Knowledge graphs are pervasive in modular neural CRSs. BARCOR builds CORG from movies, genres, cast members, directors, production companies, and genre–genre relations, then encodes it with R-GCN; EGCR uses DBpedia and ConceptNet with R-GCN to ground recommendations and explanations; UNICORN and DICR reason directly over user–item–attribute or entity–relation graphs (Wang et al., 2022, Wen et al., 2022, Deng et al., 2021, Zhou et al., 2022).

Review text is another important signal. EGCR augments a CR-Walker-style backbone with BERT-encoded IMDb reviews, averaging up to 30 reviews per movie and fusing them with KG-based movie embeddings. The stated purpose is to enhance item representation and increase the informativeness of the whole conversation, while also supporting explanations that surface reasons users commonly give for liking items (Wen et al., 2022). Self-supervised bot play with justifications similarly grounds conversational recommendation in subjective aspects extracted from reviews, using aspect vectors both for recommendation and for critique-driven update of user state (Li et al., 2021).

Several papers explicitly reject non-textual metadata or external KGs as a requirement. PECRS formulates conversational recommendation as a language processing task in which each item is represented by a textual template such as “Movie title [SEP] Actors [SEP] Director(s) [SEP] Genre(s) [SEP] Plot,” then encoded by GPT-2 into a shared semantic space with dialogue (Ravaut et al., 2024). The BM25 retrieval baseline goes further: it treats conversations as queries and items as documents, with each item document built as

$a_t$6

where $a_t$7 is the set of training conversational contexts that led to recommending item $a_t$8. Even the conversation-only variant, without metadata, reaches $a_t$9, $$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$0, and $$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$1, outperforming several knowledge-graph-heavy baselines on ReDial (Gupta et al., 2023).

The supervision signal also varies widely. Some systems depend on fully supervised conversational data; others operate with weaker or more indirect supervision. The 2016 unsupervised system learns from raw chat logs and final orders as delayed rewards, with no hand-labeled dialog acts or per-turn slot annotations (Sun et al., 2016). The 2025 RL product-recommendation framework proposes aggregate reward modeling from historical patterns, such as average conversion rate for a $$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$2 pattern or distributions of likely user responses, explicitly to protect privacy and smooth noise (Liu, 30 Jun 2025). The single-turn voice prototype relies entirely on pre-trained tools—Google speech recognition, NLTK, GloVe, TextBlob—and does not train any CNN, RNN, or LSTM despite referring to them in the surrounding literature (Talegaonkar et al., 17 May 2025).

5. Personalization, explanation, and alignment

Personalization in conversational recommendation may refer to item choice, language style, fine-grained intention modeling, or the internal alignment between recommendation and conversation. CCRS treats user experience as a first-class objective and customizes the system from three perspectives: user-specific fine-grained intentions over knowledge-graph relations, context-aware speaking style, and meta-learned user adaptation. Its relation-aware, user-conditioned graph encoder models different user preferences over relations such as actor, director, and producer, while a style-conditioned vocabulary bias modulates response generation through latent style embeddings selected by the current intention vector (Li et al., 2022).

Explanation is another axis of personalization and transparency. EGCR explicitly generates an explanation for each agent action, grounding it in the same knowledge sources—KG and reviews—used by the CRS backbone. The motivation is that existing CRS work remains a black box to users, whereas explicit explanation can improve transparency, informativeness, and trust (Wen et al., 2022). DICR makes explanation part of the internal training signal rather than a post-hoc add-on. It models recommendation actions as recommendation paths on a KG, extracts user interest shift paths from dialogue, and uses dual imitation so that recommendation paths and conversation paths align. The result is a bidirectional link: recommendation paths inform coherent explanations, while conversation-side knowledge and semantics reward better recommendation paths (Zhou et al., 2022).

The justifications framework of “Self-Supervised Bot Play for Conversational Recommendation with Justifications” is more structured than free-form explanation. It predicts subjective aspects as justifications, allows users to critique those aspects, and updates user representations through an aspect encoder

$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$3

Its second-stage bot-play fine-tuning uses a discounted cross-entropy objective over multiple turns to optimize the system for iterative conversational success without requiring dialog transcripts (Li et al., 2021).

Alignment with human preferences also appears in recent RL and LLM-based systems. The 2025 product-recommendation paper explicitly includes RLHF for nuanced goals such as helpfulness, politeness, persuasiveness without pressure, safety, and tone (Liu, 30 Jun 2025). LLMCRS uses schema-based instruction, demonstration-based instruction, dynamic sub-task and model matching, and summary-based generation, then fine-tunes the LLM with reinforcement learning from CRS performance feedback, using

$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$4

as the reward for policy-gradient updates (Feng et al., 2023).

6. Evaluation, empirical findings, and open problems

Evaluation in conversational recommendation is intrinsically multi-dimensional. The surveyed work uses recommendation metrics such as Recall@$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$5, NDCG@$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$6, MRR@$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$7, SR@$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$8, AT, hDCG@(15,10), and hN@(10,10), and dialogue metrics such as BLEU, ROUGE-L, Distinct-$$\max_\theta \; \mathbb{E}_{\pi_\theta} \Big[ \sum_{t=0}^{T} \gamma^t r_t \Big].$$9, Item-F1, AIN, Perplexity, human fluency, relevance, informativeness, coherence, and explainability (Deng et al., 2021, Wang et al., 2022, Wang et al., 2021, Ji et al., 2023, Zhou et al., 2022). The “grand challenge” review argues that offline recommendation accuracy is insufficient on its own and calls for user-centric evaluation that also addresses efficiency, dialogue quality, and usability (Jannach et al., 2022).

Representative empirical results span very different architectural regimes. On ReDial, BARCOR reports $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$0, $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$1, $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$2, and $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$3, while also achieving Dist-2 $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$4, Dist-3 $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$5, Dist-4 $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$6, and Item-F1 $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$7 (Wang et al., 2022). RecInDial emphasizes end-to-end evaluation and reports end-to-end $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$8, $$\pi^*=\argmax_{\pi\in\Pi}\mathbb{E}\left[\sum_{t=0}^T r(s_t,a_t)\right].$$9, $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$0, with Item Ratio $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$1, BLEU-2 $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$2, and ROUGE-L $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$3 (Wang et al., 2021). PECRS-medium reaches $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$4, $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$5, $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$6 on ReDial while remaining parameter-efficient, and its authors argue that Distinct metrics can be misleading because models with poor response quality can still score high on diversity (Ravaut et al., 2024). GCRS reports gains of up to $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$7 on Recall@1 over strong baselines, including $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$8 on Inspired versus a best baseline of $S_u = [s_u^1, s_u^2, \ldots, s_u^{n-1}, s_u^n]$9, while also obtaining the best Perplexity on both ReDial and Inspired (Zhang et al., 21 May 2026). The BM25 retrieval baseline, despite its simplicity, achieves $\mathcal{A}_{v^n}^*$0, $\mathcal{A}_{v^n}^*$1, and $\mathcal{A}_{v^n}^*$2 with user selection, and $\mathcal{A}_{v^n}^*$3, $\mathcal{A}_{v^n}^*$4, and $\mathcal{A}_{v^n}^*$5 after adding data augmentation for low-frequency items (Gupta et al., 2023).

At the same time, several papers are explicitly conceptual or prototype-oriented rather than benchmark-driven. The 2016 unsupervised conversational product-recommendation demo reports no detailed numerical metrics, emphasizing end-to-end feasibility instead (Sun et al., 2016). The 2025 voice-and-sentiment prototype also provides no precision, recall, F1, NDCG, MAP, or user-satisfaction scores, instead illustrating feasibility through example interactions (Talegaonkar et al., 17 May 2025). The 2025 RL product-recommendation paper is more conceptual than empirical and does not provide benchmark tables or specific numerical results, though it specifies reward design, algorithmic options, and deployment patterns (Liu, 30 Jun 2025).

Open problems recur across the literature. The “grand challenge” perspective highlights the difficulty of integrating recommendation, dialogue management, NLU, NLG, and user modeling at a level comparable to human advisers (Jannach et al., 2022). Many papers note sparse or delayed rewards, data hunger, long-tail items, metric mismatch, interpretability, safety, and trust. UNICORN and MSCAA show that explicit policy learning becomes difficult as action spaces grow, motivating pruning and hierarchical control (Deng et al., 2021, Ji et al., 2023). KG-heavy methods depend on graph quality and domain coverage (Wang et al., 2022, Zhou et al., 2022). Retrieval-style methods remain vulnerable to lexical mismatch and cold start (Gupta et al., 2023). LLM-based orchestration and fully generative models raise additional concerns about faithfulness, scalability, bias, persuasive power, and the need for transparency and user control (Feng et al., 2023, Zhang et al., 21 May 2026).

Taken together, these strands suggest that conversational recommendation is not a single architecture but a family of approaches for coupling recommendation with dialogue. Some emphasize policy learning, some unify generation and recommendation, some retrieve over conversational evidence, and some treat the LLM as controller rather than recommender. The common research goal is consistent: to use interaction itself—questions, critiques, explanations, and multi-turn context—not merely as an interface layer, but as a primary source of preference signals and a primary object of optimization (Jannach et al., 2022, Liu, 30 Jun 2025).