Adapting Generative Recommenders for Multi-Turn Interaction

1 Introduction

• Fixed-history generative recommenders offer no direct way to correct recommendations that miss a user’s current intent.Preference changes become observable only through later item interactions, making correction indirect and potentially slow.Source: §1 · p. 2
• Because conversation and item generation share model parameters, conversational training may overwrite the history-to-item mapping.Items and words share an output space, allowing conversation without an architectural change but exposing the learned mapping to forgetting.Source: §1 · p. 2
• INTEGER treats conversation as a control layer that determines when the recommender’s history-to-item mapping is invoked and how recommendations are presented or revised.Its design exposes recommendation as a learned decoding action and combines behavioral replay with instruction-data rehearsal to preserve accuracy during adaptation.Source: §1 · p. 2
• On Amazon Beauty and Toys, INTEGER is competitive with end-to-end and modular systems and significantly outperforms its generative-recommender starting point.The paper also analyzes how training erodes conversational and behavioral capabilities, how re-anchoring and rehearsal counteract this, and how refinement learns a single intent-agnostic replacement.Source: §1 · p. 2

2.1 Multi-Turn Interaction

• Multi-turn recommendation must decide when to recommend and generate an item consistent with the intent expressed so far, while still producing fluent responses.The dialogue context D_t ends with a pending user message, and the response is generated autoregressively from that context.Source: §2.1 · p. 3
• Prior approaches learn recommendation from dialogue, convert conversation into feedback for an existing recommender, or delegate recommendation to an external model or tool.These approaches respectively include end-to-end conversational recommenders, pseudo-interaction or preference-signal methods, and modular pipelines.Source: §2.1 · p. 3

2.2 Generative Recommendation

• Generative recommendation decodes an item from a fixed interaction history, leaving users unable to revise the input except through further interactions.Items are represented as sequences of discrete tokens and generated autoregressively; each ItemID uses codewords derived from item content embeddings, with similar items sharing longer prefixes.Source: §2.2 · p. 3
• At inference, a prefix-trie-constrained beam search produces a ranked top-N list whose generated sequences correspond to valid catalog items.The trie restricts decoding to prefixes of valid ItemIDs.Source: §2.2 · p. 3
• INTEGER keeps recommendation grounded in the history-to-item mapping learned from interaction sequences, with conversation acting as a control layer rather than requiring a separate recommender.Its overview depicts a learned routing token choosing between chat and recommendation, history placed before ItemID generation, and feedback triggering a revised recommendation.Source: §2.2 · p. 4, Figure 1 · p. 4

3 INTEGER: Interactive Generative Recommendation

• INTEGER starts from a generative recommender and supplies the interaction history in the system prompt alongside each turn’s dialogue context.The model decides whether to chat or recommend, and places the history next to the generated ItemID when recommending.Source: §3 · p. 4

3.1 Recommendation as a Decoding Action

• A learned <hist> routing token initiates recommendation; otherwise, the model continues with a conversational response.Recommendation turns follow <hist> H <rec> z′, after which the ItemID is decoded under a valid-item trie constraint and response generation resumes.Source: §3.1 · pp. 4–5
• History re-anchoring inserts H immediately before ItemID generation, restoring the local history-to-item format used to train the original recommender.The history remains available in the system prompt during dialogue, while the re-anchored copy lets item generation draw on behavior and expressed preferences together.Source: §3.1 · p. 5

3.2 Conversational Adaptation with Behavioral Replay

• Conversational adaptation mixes multi-turn recommendation dialogues, sequential recommendation examples, and general instruction-following examples under next-token prediction.The dialogues teach routing and response behavior, while the other sources retain recommendation behavior and strengthen response generation.Source: §3.2 · pp. 5–6
• Behavioral replay continues history-to-ItemID training during conversational adaptation to counter weakening of the mapping learned in generative recommendation.For recommendation turns, the re-anchored history and <rec> are excluded from the loss; supervision covers <hist>, the ItemID, and the subsequent response.Source: §3.2 · pp. 5–6

3.3 Item Refinement

• INTEGER uses preference pairs to train revised recommendations after rejection or requests for variety within the same dialogue.For rejection, the preferred item is the ground-truth item; for variety, it is the current item’s nearest neighbor by content-embedding cosine similarity, and DPO applies the preference update.Source: §3.3 · p. 6

4.1 Experimental Settings

• The evaluation uses Beauty and Toys benchmarks with synthesized category-specific conversations, and compares sequential, end-to-end conversational, and modular baselines.For conversational systems, accuracy is measured at the ground-truth recommendation turn; conversation quality is the mean LLM-as-judge rating across four criteria.Source: §4.1 · pp. 6–7

4.2 Main Results

• GCRS has near-zero Hits@10 (0.004 and 0.002) despite competitive conversation quality, while PECRS has the strongest baseline Toys NDCG@10 but the lowest conversation quality (3.92 and 4.05).The modular pipelines fall between these extremes, with conversation quality of 4.40–4.58 at moderate accuracy.Source: §4.2 · pp. 7–8
• On Beauty, INTEGER scores 0.051 Hits@10, 0.027 NDCG@10, and 4.77 conversation quality; on Toys, it ties PECRS at 0.057 Hits@10 but trails its 0.032 NDCG@10 with 0.028.On both datasets, INTEGER significantly outperforms TIGER in Hits@10 and NDCG@10 (paired t-test, p < 0.001).Source: §4.2 · p. 8, Table 3 · p. 8

4.3 Ablation Studies

• Removing history re-anchoring or behavioral replay costs roughly 33% of recommendation accuracy, while instruction-data rehearsal raises conversation quality by 0.63 with accuracy essentially unchanged.Dialogue-only training collapses to Hits@10=0.029 and Conv. Q=2.39; removing replay raises conversation quality from 4.14 to 4.37.Source: §4.3 · p. 9, Table 4 · p. 9

5.1 Routing Behavior

• INTEGER assigns routing-token probability 0.88 to recommendation turns versus 0.09 before and 0.00 after them, with AUC 0.94; bare replay has AUC 0.51.Bare replay's median probability is 0.99 before the target and 0.86 on recommendation turns, so it tends to recommend prematurely rather than distinguish turn types.Source: §5.1 · pp. 9–10, Figure 2 · p. 10

5.2 Two Directions of Forgetting

• At recommendation time, INTEGER attends to prompted history 2.11× more than while chatting, while dialogue-only finetuning attends 0.86× as much.Attention to far prompt history decays to 0.81× for INTEGER and 0.63× for dialogue-only; adjacent recap attention stays flat at 1.00×.Source: §5.2 · p. 10, Figure 3 · p. 11
• Conversation quality rises sharply through 10% instruction data, then plateaus by +0.5 to +1.2, while Hits@10 is non-monotone and peaks at different fractions by dataset.The analysis reports no systematic accuracy trend as the instruction-following-data portion changes.Source: §5.2 · pp. 10–11, Figure 4 · p. 11

5.3 Item Refinement

• DPO lowers Repeat@10 to 0.284 and 0.269 and raises gold-item recovery from 0.228 to 0.367 and 0.182 to 0.364 across variety and rejection intents.The table covers recoverable Amazon Beauty cases where the gold item originally ranked 2–50.Source: §5.3 · p. 11, Table 5 · p. 12
• Because both preference-pair types reward replacement without an intent label, DPO learns one intent-agnostic replacement operation, yielding similar outcomes for variety and rejection.The measured similarity between re-offered and rejected items is 0.885 for variety and 0.886 for rejection.Source: §5.3 · pp. 11–12

5.4 Case Study

• In the case study, bare replay recommends at every assistant turn, while INTEGER asks clarifying questions and waits until the user's goal is clear.INTEGER recommends at the third turn, after the user mentions patchy foundation.Source: §5.4 · p. 12

• INTEGER exposes recommendation as a decoding action and is competitive with end-to-end conversational recommenders and modular pipelines in accuracy and conversation quality on Amazon Beauty and Toys.It significantly outperforms the generative recommender from which it is adapted.Source: §6 · p. 12
• The findings are limited by synthesized dialogues and refinement that recovers near-misses rather than satisfying stated preferences.Attribute-level feedback would require conditioning refinement on item attributes; the learned operation is intent-agnostic and moves within the embedding neighborhood.Source: §6 · p. 12
• In the illustrated dialogue, bare replay recommends prematurely at every turn, while INTEGER asks clarifying questions and recommends only at the gold turn.The example is from Amazon Beauty.Source: Figure 5 · p. 13