CS Paper Radar

Better Rankers, Worse Retrievers: The Hidden Cost of RL Post-Training in Generative Recommendation

Yuheng Chen, Fan Zhang, Yixuan Xu, Chao Yang, Yang Xiang, Can Yang

cs.IR cs.AI ·submitted 2026-11-17 14:19 UTC · v1

Abstract

Generative recommenders decode the semantic ID of a user's next item with a language model, and post-training them with reinforcement learning (RL) after supervised fine-tuning (SFT) has become standard practice. The gains are mostly reported on the first ten results, yet in deployment such models hand hundreds of candidates to a downstream ranker or reward model. We show that RL post-training has a hidden cost: it improves accuracy at the top of the list (hit@K for small K) but, on all five datasets we study, lowers the recall of the candidate list (hit@K for K in the hundreds). The model becomes a better ranker and a worse retriever. Controlled experiments trace the loss to the negative advantages that RL assigns to the candidates it samples, the rollout. The root cause is that only one next item is logged per user, so RL treats every other candidate as wrong. Yet these candidates are the model's most plausible alternatives, and often the next items of other users with similar histories. Penalising them pushes down exactly what similar users choose next, and these items drop out of those users' candidate lists. This diagnosis points directly to a remedy: out-of-rollout anchoring, a single additional loss term that preserves the SFT model's ordering of the items outside the rollout. It keeps the gain at the top and restores recall to the SFT level on all five datasets, and it is effective under other RL configurations as well. To our knowledge, this is the first systematic study of this recall loss: we measure it, identify its cause, and show that, once the cause is known, a simple remedy suffices.

Closest papers

Nearest neighbours by meaning across the whole index.

Many Right Answers, One Training Example: Removing Responder Noise from GRPO for LLM Recommenders

cs.CL ·2026-10-13 ·Xin Li, Hao Jiang, Jinghao Guo, Weisi Lin, Chau Yuen
positioncode/data similarity 0.74
Abstract · Project Page: https://lixin.ai/ReMaRk/

A recommendation request on Reddit draws several replies, each naming different titles. When an LLM recommender is trained on such threads with reinforcement learning (RL), which replies should a generated list be rewarded against? The state-of-the-art recipe, Rank-GRPO, picks one reply per training example, so the advantage mostly records which reply was drawn (85% of its variance) rather than how good the list is. We argue that RL should train on the request, not on each answer, and propose ReMaRk: it scores every list once against the union of all replies' titles and samples each request equally. This makes RL cheaper. On seven backbones and three datasets, ReMaRk reaches the validation recall of 2,000 per-reply RL steps in 25 to 53% of the steps and leads at that budget; trained until neither improves, it is never significantly behind Rank-GRPO and in four of seven pairs reaches Rank-GRPO's best with fewer GPU-hours. The same reward helps on ProtoQA, where about 100 people answer each question, which points to a simple rule for RL with several valid human answers: reward them jointly.

0.47

The Recall Ceiling of LLM Recommendation Reranking

cs.IR cs.LG ·2026-08-27 ·Zhaohui Wang
top venuetheory similarity 0.72
Abstract · 11 pages, 1 figure, 8 tables. Accepted for oral presentation at CIKM 2026

Some LLM-based recommendation rerankers are evaluated under an oracle protocol that guarantees the ground-truth item is present in the scored set, either by injecting it into the candidate list or by scoring it against sampled negatives. Across three primary Amazon datasets, we show that this protocol overestimates realistic NDCG@10 by 92--95%. The cause is a recall ceiling: realistic retrieval covers only 2--19% of relevant items at $K=100$ across eight datasets in three domains, imposing a deterministic upper bound on any closed-candidate reranker's top-$k$ NDCG. Under leave-one-out evaluation, $\mathbb{E}[\mathrm{NDCG}@k] \leq \mathrm{Recall}@|W_π|$, where $W_π$ is the reranker's candidate window. Under realistic retrieval, none of the tested optimisation strategies significantly improves over the collaborative-filtering baseline on our primary Amazon datasets. These strategies include prompt engineering, model scaling over a 168$\times$ parameter range, sequential models, supervised neural rerankers, LoRA fine-tuning, hybrid retrieval, score-aware prompting, and LLM+CF fusion. Text-aware retrieval increases recall on one dataset but does not improve end-to-end NDCG, while providing upstream CF scores mainly makes the LLM reproduce the CF order. We therefore propose the Recall-Aware Evaluation Protocol (RAEP): first classify the retrieval-recall regime, then evaluate reranking where the ceiling permits meaningful differentiation. In the low-recall regimes measured here, improving retrieval is more consequential than increasing reranker sophistication; this ordering need not hold in production systems with higher recall, richer features, or online feedback.

0.43

Training with Missed Targets in Generative Recommendation: Separating Supervision from Probability Competition

cs.IR cs.LG ·2026-10-07 ·Xuesi Wang, Yangbin Shi, Xiaolin Zheng
similarity 0.71
Abstract · 12 pages, 4 figures, 8 tables

Generative recommenders return a limited candidate set and may omit observed targets before reranking. A training strategy appends these missed targets to reranker training lists, although inference still ranks only original candidates. This operation simultaneously changes retrieved-target weight, adds supervision over appended targets, and makes the two groups compete for probability. An append/no-append comparison therefore cannot explain changes in returned-item rankings. We construct three matched losses that hold retrieved-target weight fixed while introducing appended-target supervision and group competition separately. The intermediate loss trains within both groups but normalizes them separately, preventing training-only targets from competing with inference candidates. Experiments with a released OneRec model and locally trained Amazon generators show that this competition can harm returned-item ranking. In four prespecified Amazon Video Games comparisons, removing it improved full-target normalized discounted cumulative gain (FT-NDCG) by 7.8--22.2\%; 95\% intervals over users and three of four intervals over training runs excluded zero. A conservative development-set rule selected appended-target training for two of three generators in one held-out category and rejected it for all three in another, avoiding a 1.7\% loss. Candidate completion should therefore be evaluated for each generator rather than applied automatically.

0.17

R2QRec: Learning Recommendation Queries from Retrieval Feedback

cs.IR ·2026-10-29 ·Weixin Chen, Shuchang Liu, Hailan Yang, Yuhan Zhao, Li Kang +11 more
hot direction similarity 0.69
Abstract

Query-based recommendation separates expressing user needs from locating matching items. Its central learning problem is how to turn item interactions into effective retrieval requests when interaction logs do not annotate queries. We present R2QRec, which uses a fixed retrieval system to construct query supervision and evaluate recommendation outcomes. Retrieval verification first selects generated expressions that recover observed items as supervised fine tuning targets, allowing retrieval behavior to determine the target wording. Policy refinement then evaluates generated query sets by the observed item's rank in their fused recommendation list. Log-rank feedback distinguishes outcomes beyond the recommendation cutoff, while a KL penalty constrains policy change relative to the supervised reference. The stages have distinct roles: verification identifies expressions to imitate, and policy refinement improves the combined ranking produced by newly generated queries. Only the query generator is trained; the retriever and item index remain fixed, and no query annotations or learned reward model are required. Across three public datasets, R2QRec achieves state-of-the-art overall recommendation performance among the evaluated methods. Extensive experiments demonstrate the effectiveness of the proposed design and its generalization across generators and domains.

0.25

What Gets Hacked in a Collaborative Reward for LLM Recommenders: The Part That Does Not Depend on the User

cs.IR ·2026-11-05 ·Zhuangsheng Li, Liang Dong, Yu Wu, Dailin Li, Yu Xia +4 more
code/data similarity 0.69
Abstract

Collaborative-filtering (CF) scores provide graded, personalized rewards for training large language model (LLM) recommenders through reinforcement learning. However, these rewards can cause exposure collapse: even though the CF model itself produces diverse top recommendations across users, an LLM optimized with its rewards concentrates on fewer items. We decompose each score into an item-mean component, the item's average CF score across user histories, and a residual component, the current user's score deviation from the average. Through controlled interventions, we identify the item-mean component as a contributor to exposure collapse. Accordingly, we propose two reward corrections: per-item centering removes each item's estimated mean score, while per-item standardization further rescales the residual using its estimated standard deviation. Experiments on two datasets with two identifier schemes show that both corrections substantially improve catalog coverage and exposure evenness. Compared with training on the uncorrected CF reward, our corrections increase the number of distinct items ranked first across users by up to $9.8\times$, while maintaining comparable ranking accuracy. Code is available at https://anonymous.4open.science/r/LLM4Rec-CF-Bias-27EF/.

0.38

Recommendation Retrievers Need Verifiers: Universal Generative Reranking for Sequential Recommendations

cs.IR cs.AI ·2026-09-10 ·Benyu Zhang, Qiang Zhang, Rui Li, Qunshu Zhang, Devansh Tandon +1 more
similarity 0.69
Abstract

First-stage recommenders in multi-stage systems produce a ranked candidate list from which a limited prefix is forwarded to downstream rankers. Because each forwarded item must be processed by more expensive ranking stages, this shortlist cannot be arbitrarily large. The first-stage objective is therefore high coverage of relevant items within the forwarded prefix, commonly measured by Recall@$k$. A relevant item may be available deeper in the retrieved list but absent from the shorter prefix that is actually consumed. This paper studies post-hoc verification for promoting such candidates into the consumed shortlist without retraining or replacing the retriever. We introduce a lightweight generative verifier for retrieval models. Given a retriever state and a candidate item, the verifier scores the item through the likelihood of its identifier tokens. It is trained post hoc with next-token cross entropy, requires no sampled negatives or candidate pool during training, and scores only the retriever's top-$K$ candidates at inference. The interface is minimal: the retriever supplies a query state and candidate items, and the item representation can use any fixed tokenization. Across Amazon product recommendation and YaMBDa music recommendation, the same verifier training recipe improves Recall@10 for SASRec, GRU4Rec, NextItNet, and MiniOneRec. Ablations show that the improvements are not explained solely by injecting item-content features into the retriever, supporting verification as a post-hoc output-side adaptation mechanism.

0.10

Revisiting Generative vs. Dense Retrieval: A Static Value Prior Reverses the Gap

cs.IR cs.AIcs.LG ·2026-10-13 ·Renbo Tu, Yang Liu, Shuying Sun
published similarity 0.68
Abstract · ACM CIKM EARL 2026

Generative retrieval (GR) has emerged as a strong candidate generator for sequential recommendation as language models and semantic-item identifiers have improved. However, GR is not a uniformly stronger replacement of dense retrieval (DR) in online retail candidate generation. We examine their differences on two Amazon Toys datasets under different metrics. GR leads on best relevant item value among users both methods cover, while DR leads on exact-target recall and relevant value across the ranked list. With deep analysis, we find that popularity contributes the most to GR's measured value lead. Adding the same static item value to DR in one reranking step reverses this lead and keeps recall above GR. Finally, on an industrial-scale internal e-commerce dataset with LLM-judged relevance, we show that a 14B SID-LLM still underperforms a $47\times$ smaller dense encoder.

0.42

More Than a Score: Reinforcement Learning with Verbalized Feedback for Generative Recommendation

cs.IR ·2026-11-17 ·Zipeng Chen, Xinyu Lin, Yuliang Sun, Sheng Zhang, Zeyu Wang +3 more
similarity 0.68
Abstract · 10 pages, 11 figures and tables

Large language models (LLMs) are increasingly used for generative recommendation, where recommended items are generated directly from users' interaction histories. To further improve recommendation quality, reinforcement learning (RL) has been applied to align these models with recommendation objectives. However, existing RL methods rely on scalar rewards to implicitly guide exploration. These rewards do not specify which aspects of a candidate conflict with the user's preferences or how the next candidate should differ. To address this gap, we propose VerRec, a verbalized feedback-guided RL framework for generative recommendation. Specifically, a user simulator generates personalized verbalized feedback based on the user's historical preferences and the candidate item profile. The policy uses this feedback to produce refined recommendations. Original and refined candidates form an expanded group and are jointly optimized with group relative policy optimization. The user simulator is only required during training and is removed at inference time. Experiments on three Amazon review datasets with two generative recommendation backbones demonstrate that VerRec improves recommendation accuracy and ranking quality. Further analyses show that verbalized guidance encourages exploration of more diverse candidates and helps the policy discover more rewarding recommendations during training.

0.13