type
Post
status
Published
date
Jul 4, 2026 07:02
slug
rec-weekly-en-2026-W27
summary
24 papers this week, 4 from industrial online deployments (Meta, Netflix, Alibaba, Kuaishou), covering retrieval, ranking, re-ranking, and full-page generation. The underlying logic of core technical density is shifting—generative recommendation moves from "being able to generate" to "being able to reason," retrieval shifts from embedding matching to navigational exploration, and the ranking stage seeks balance between constraints and interpretability. Generative recommendation enters the "reasoning + RL" era: GR2, ShopX, and GenPage all showcased different architectural directions for generative systems in the same week. GR2 introduces reasoning chains (CoT) and RL post-training to the re-ranking stage for the first time, achieving +18.7% in R@1 on live traffic. ShopX pushes generative recommendation from candidate generation to end-to-end "intent-to-item" execution, boosting complex request satisfaction by 55–75% in Taobao's agent scenario. GenPage goes furthest—replacing Netflix's entire multi-stage homepage pipeline with a single Transformer, delivering +0.24% on the core metric while cutting latency by 20%. The common thread across all three: the core barrier for generative recommendation has shifted from "can it generate?" to "can it find an industrially feasible solution that balances reasoning quality and deployment efficiency?" Retrieval moves from static matching to dynamic graph exploration: Meta's hard negative sampling uses LLM clustering to generate real-time same-cluster negatives, lifting online recall by +8.5% and reducing popularity bias by -12.3%. Kuaishou's IID-Nav models retrieval as autonomous graph exploration, supporting unlimited indirect depth traversal. Kuaishou's POEM uses multi-task ranking scores to construct partial order sequences, enabling real-time per-request interest updates. All three technical paths share a trend: retrieval is moving from static embedding lookup to dynamic, context-aware behavior modeling. Constrained optimizati
tags
Recommendation Systems
Weekly
Papers
category
Rec Tech Report
icon
📚
password
priority
1
Weekly Overview
24 papers this week, 4 from industrial online deployments (Meta, Netflix, Alibaba, Kuaishou), covering retrieval, ranking, re-ranking, and full-page generation. The underlying logic of core technical density is shifting—generative recommendation moves from "being able to generate" to "being able to reason," retrieval shifts from embedding matching to navigational exploration, and the ranking stage seeks balance between constraints and interpretability.
Generative recommendation enters the "reasoning + RL" era: GR2, ShopX, and GenPage all showcased different architectural directions for generative systems in the same week. GR2 introduces reasoning chains (CoT) and RL post-training to the re-ranking stage for the first time, achieving +18.7% in R@1 on live traffic. ShopX pushes generative recommendation from candidate generation to end-to-end "intent-to-item" execution, boosting complex request satisfaction by 55–75% in Taobao's agent scenario. GenPage goes furthest—replacing Netflix's entire multi-stage homepage pipeline with a single Transformer, delivering +0.24% on the core metric while cutting latency by 20%. The common thread across all three: the core barrier for generative recommendation has shifted from "can it generate?" to "can it find an industrially feasible solution that balances reasoning quality and deployment efficiency?"
Retrieval moves from static matching to dynamic graph exploration: Meta's hard negative sampling uses LLM clustering to generate real-time same-cluster negatives, lifting online recall by +8.5% and reducing popularity bias by -12.3%. Kuaishou's IID-Nav models retrieval as autonomous graph exploration, supporting unlimited indirect depth traversal. Kuaishou's POEM uses multi-task ranking scores to construct partial order sequences, enabling real-time per-request interest updates. All three technical paths share a trend: retrieval is moving from static embedding lookup to dynamic, context-aware behavior modeling.
Constrained optimization and interpretability make a return: Avito's PermR demonstrates on 56 million search queries that adjacent swaps can approximate integer programming gains within production latency. KakaoBank's ChunkGroupSHAP bridges the granularity gap between word-level explanations and dense rankers using grouped Shapley values. Both works serve as a reminder: amidst the LLM hype, incremental engineering optimizations and interpretability tools can still deliver concrete gains.
Generative Recommendation: Reasoning Chains, RL Post-training, Model-native Execution
Five generative recommendation papers this week span the full spectrum from candidate generation to full-page construction. The shared signal: RL post-training is moving from "optional" to "standard."
GR2 (Yongkang/Yang/Yongkang team) — This is the only generative reasoner deployed in an industrial re-ranking stage so far. GR2's core design combines three components: a semantic ID tokenizer (≥99% uniqueness), reasoning trajectory distillation from a strong teacher, and RL post-training based on verifiable rewards. The technical report provides several key details: SFT actually degrades at industrial scale, making On-Policy Distillation (OPD) a viable alternative. On live traffic, GR2 improves over traditional re-ranking baselines by +18.7% R@1, +7.1% R@3, and +9.6% N@3. The reward design section is worth a close read for practitioners: LLMs can "hack rewards" by preserving input order or exploiting position biases—an issue reminiscent of the semantic ID sequence selection problem noted in GLASS.
ShopX (Alibaba) — Unlike GR2's focus on re-ranking, ShopX targets end-to-end intent-to-item execution in agent shopping scenarios. It uses a single foundation model to perform intent understanding, execution planning, and native SID operations (SID beam-search retrieval, list-level ranking, product bundling), eliminating information loss in traditional tool-calling pipelines. Evaluations on Taobao production logs show that model-native execution improves intent satisfaction by 55–75% over tool-mediated systems like ChatRec and InteRecAgent on complex/vague requests. ShopX's "semantically recoverable SID" design follows the generative lineage pioneered by TIGER, but generalizes the output space from single candidates to flexible item-space operations.
GenPage (Netflix) — The most aggressive generative recommendation practice to date: a single Transformer generates the entire Netflix homepage (multiple rows, structured layout) end-to-end, replacing the full multi-stage pipeline. GenPage uses a standard LLM training recipe: pre-training → WBC or RL post-training. In online A/B tests, the WBC variant yields +0.24% on core engagement metrics (p<0.001) with a 20% end-to-end latency reduction. Two offline findings stand out: at current scale, prompt optimization yields more gains than scaling model capacity; RL post-training increases homepage diversity—even though diversity was not an optimization target.
IntuRec (Academic) — Introduces "recommendation intuition" as a prior for latent reasoning. IntuRec first has an LLM generate a top-K candidate set from user history (intuition extraction), then encodes it into a preference-aligned intuition embedding via self-attention and cross-attention, initializing the starting point of latent reasoning. On Amazon Beauty, Sports, and Yelp, IntuRec outperforms baselines like P5, TIGER, and SASRec, demonstrating that the choice of starting point for latent reasoning is more critical than the reasoning itself.
Diffusion-GR2 (Academic) — A solution to GR2's inference speed problem: convert the autoregressive generative re-ranker into a block diffusion model, bridging the structural and distributional gap through a three-stage training pipeline (CFT + OPD + RL). On Amazon Beauty, Diffusion-GR2 achieves 2.4–3.5x inference throughput while maintaining accuracy close to GR2. Ablations show CFT recovers most of the conversion gap. This is a natural extension of the dense supervised re-ranking idea from DeGRe into the diffusion paradigm.
Retrieval and Candidate Generation: Hard Negative Sampling, Graph Exploration, and Multi-Sequence Decoupling
Eight retrieval papers this week take divergent paths but converge logically: single-vector matching is insufficient—context-aware negatives, disentangled multi-intent, or dynamic graph traversal are needed.
Meta Hard Negative Sampling (Meta) — In production two-tower retrieval, uses LLM clustering to generate same-cluster hard negatives, replacing in-batch or out-of-batch negative sampling. The key design is self-supervised: the LLM does not directly annotate, but clusters by learning item representation distributions; items in the same cluster serve as hard negatives for each other. Online deployment yields +8.5% recall and -12.3% popularity bias. The latter means hard negative sampling breaks the feedback loop where "popular items are always treated as positives, cold items are always ignored"—a debiasing effect consistent with the motivation of Sampling-Bias-Corrected Neural Modeling but achieved via a different path.
CMSL (Meta) — Challenges the single-sequence modeling assumption from a structural perspective. CMSL argues that the "multi-faceted" nature of user behavior is fundamentally inconsistent with the unidirectional linear nature of natural language sequences, leading to unrelated interests diluting each other in a single sequence ("context contamination"). The solution decomposes user history into multiple thematic sequences via learnable sequence construction modules, then models them in parallel with linear attention. CMSL is deployed across four major Meta surfaces (covering ranking and retrieval), but the abstract does not disclose specific online metrics—the most notable missing data point among this week's industrial papers.
IID-Nav (Kuaishou) — Redefines retrieval as stateful autonomous graph exploration, rather than static neighbor expansion. Three core components: a goal-aware navigation policy (steering expansion toward user intent), recursive state evolution (reusing state across requests to support logically infinite depth traversal without linear inference latency growth), and trajectory alignment training (stabilizing optimization with graph hard negatives). On a billion-scale industrial dataset, IID-Nav outperforms i2i and index retrieval baselines under strict latency budgets. This work extends the dynamic graph idea from Dynamic Graph with Similarity-Aware Attention, but makes search drift an explicit optimization target.
MDP Planning Retrieval (VK / Academic) — A narrow but deep question: when is it worth planning for the dynamics of user state caused by fold-in? The authors model top-K retrieval as an MDP over ALS posteriors, with items as actions, closed-form rank-one fold-in as transitions, and a reward combining relevance similarity and posterior alignment. On 5 datasets, a single-step lookahead captures most of the gain—meaning dynamic awareness is better than static top-K, but full MCTS is not needed. A practical takeaway: use cosine similarity instead of inner product for relevance, since inner product is entangled with popularity.
PPRO (Baidu) — User-aware optimization for retrieval from long-term dialogue memory. PPRO builds two memory banks (episodic, semantic) and derives a user profile from cumulative memory as an explicit personalization prior. It then trains a query reformulator using GRPO with a joint reward based on evidence retrieval quality and downstream answer quality. On LoCoMo and LongMemEval-S, PPRO improves F1 by 12.3% and 9.8% respectively over training-free baselines. Ablations show that both profile-guided ranking and retrieval-oriented reformulation contribute—contrasting with the retrieval-feedback reformulation approach in ReFeed.
SchemaRAG (Microsoft) — Addresses RAG efficiency for large-schema information extraction: full schemas lead to high cost, slow latency, and loss of mid-position items. SchemaRAG dynamically reduces the output schema space via retrieval, achieving +8.8% F1, 47% latency reduction, and 48% token cost reduction on medical and e-commerce datasets.
Vocabulary Transfer (Amazon) — Reveals a paradox: advanced encoders like ModernBERT outperform BERT-base on dense retrieval but lag behind on learned sparse retrieval (LSR). The root cause is a "vocabulary gap": the original vocabularies of modern tokenizers, designed for lossless reconstruction, map the same semantic unit to multiple redundant surface forms. The VT framework transfers ModernBERT to a sparse-friendly normalized vocabulary via semantically initialized space-topology-preserving and activation potential calibration, achieving 52.4 nDCG (+4.7) on BEIR. Practitioners on the SPLADE path should take note.
GeoRAG (Academic) — Recasts RAG context selection as information-need-coverage optimization: builds a multi-dimensional demand distribution using diverse sub-queries and inverse verification weighting, then selects contexts by minimizing the Sinkhorn-Wasserstein distance. The setup is monotone submodular with a (1-1/e) greedy guarantee. On 6 QA benchmarks, GeoRAG's exact match surpasses top-k truncation by +6.5 to +9.7 points, without training and independent of the retriever.
Ranking and Re-ranking Optimization: Real-time Sequence Modeling, Constrained Re-ranking, Interpretability
Three industrial papers this week in the ranking stage focus on real-time modeling, constrained optimization, and interpretability—all non-LLM incremental improvements, but validated online.
POEM (Kuaishou) — The core tension in real-time sequence modeling: user interests drift between requests, but traditional models only see static history. POEM uses upstream multi-task ranking scores (estimated CTR, estimated watch time) as supervision signals to construct dynamic partial order sequences. It also introduces a hierarchical sample learning strategy, treating system-preferred high-ranking items as positives, combined with graph-mined hard negatives and a margin-based pairwise loss. Full deployment on two Kuaishou pages yields average watch time lifts of +0.249% (KS Single Page) and +0.213% (KS Lite Page). The partial order sequence idea extends the session partitioning from DSIN, but uses ranking scores rather than time intervals to drive sequence reconstruction.
PermR (Avito) — A textbook engineering optimization case: after ranking, use adjacent swapping to maximize revenue while satisfying relevance constraints. In an online A/B test on 56 million search queries, PermR increases revenue by +2%, achieving 63% of the gain from an exact ILP solution, while meeting all constraints and keeping latency production-acceptable. +2% may seem modest, but given PermR is a lightweight permutation search (non-learned model), the cost-benefit ratio is noteworthy.
ChunkGroupSHAP (KakaoBank) — A granularity gap exists between word-level explanations and dense rankers. ChunkGroupSHAP clusters semantically related text blocks into cross-document features and assigns importance via Shapley values. Different rankers require different explanation units: the E5 dense ranker prefers corpus-level grouping, while BM25 prefers word-level features. This finding has direct implications for building industrial interpretability toolchains—not all explanation methods suit dense ranking.
Bi-NAS (Academic) — Uses bi-level neural architecture search to simultaneously optimize cross-attention mechanisms and feature interaction functions, combined with LLM zero-shot generation of personalized explanations. Improves recommendation accuracy by 5–10% on Yelp and Amazon datasets. However, the feasibility of NAS training costs at industrial scale is questionable—experimental scale is limited (max params not provided).
Contextual Slate GLM Bandits (Microsoft Research India) — Addresses the slate bandit problem in the limited-adaptivity setting. Two algorithms (B-SlateGLinCB and RS-SlateGLinCB) achieve regret bounds of O(Nd^{3/2}√T) and O(Nd√T) respectively, independent of the nonlinear parameter κ. Simulations show B-SlateGLinCB matches the fully adaptive Slate-GLM-OFU baseline under limited adaptivity. A theoretical direction, but offers a practical limited-adaptivity slate decision module—potentially applicable to ad slot selection in industrial settings.
IntentTune (eBay) — Disambiguates vague query intent (gender, age, category, size) using user-level behavioral signals (mainly historical search queries), outperforming demographic profiling. A concrete optimization for e-commerce search, but note the dataset is internal eBay data and lacks online validation.
Directions to Watch
RL post-training is becoming the "third stage" of generative recommendation
GR2's RL post-training, GenPage's RL post-training, Diffusion-GR2's RL stage—three independent works this week adopt RL post-training as a standard component for generative re-ranking. The traditional "pre-training → SFT" two-stage recipe shows clear shortcomings at industrial scale: GR2 explicitly reports that SFT collapses at scale, making OPD and RL necessary. This signal is critical for teams building generative recommendation pipelines: if you plan to use LLMs for re-ranking or generative retrieval, you need to start thinking about reward design (verifiable vs. model rewards, and defenses against reward hacking).
The "granularity alignment" problem between retrieval and ranking appears across multiple works
ChunkGroupSHAP finds a mismatch between word-level explanations and dense ranker granularity; IID-Nav finds a mismatch between static entry nodes and real-time user intent; Meta's hard negative sampling reveals bias at the "item granularity" level in traditional negative sampling; POEM finds that static historical sequences cannot capture request-level interest granularity. These four works can be unified under one question: different stages of a recommendation system have inconsistent understandings of "granularity," and the resulting information loss is comparable to model capacity bottlenecks. A likely trend for H2 2026: more work will focus on cross-stage granularity alignment mechanisms.
Agent recommendation moves from proof-of-concept to evaluation-driven system design
ShopX's model-native framework and EvoRec's multi-agent self-evolving framework (+1.85% revenue, +1.02% CTR) represent two different paths for agent recommendation, but both rely on structured evaluation protocols. ShopX uses Taobao production logs to build an 8-dimensional evaluation metric set (intent satisfaction, constraint anchoring, catalog coverage, etc.). EvoRec enables agents to accumulate methodologies autonomously through a dual-loop cycle (model iteration + skill distillation). Both cases show that the core bottleneck for agent recommendation has shifted from "can we build it?" to "can we systematically evaluate and improve it?"—consistent with the simulation-based evaluation approach in ContextSim, but closer to production environments.
Paper Roundup
Generative Recommendation
GR2 — Yongkang/Yang team's first industrial generative reasoning re-ranker, combining semantic IDs, reasoning trajectory distillation, and RL verifiable rewards; online R@1 +18.7%, R@3 +7.1%.
ShopX — Alibaba's unified foundation model for intent understanding, execution planning, and native SID operations; +55–75% complex request satisfaction in Taobao agent scenarios.
GenPage — Netflix replaces its entire multi-stage homepage pipeline with a single Transformer, end-to-end generation of structured multi-row homepage; core metric +0.24%, latency -20%.
IntuRec — Academic: "recommendation intuition"-guided latent reasoning framework; outperforms P5, TIGER, and other baselines on Amazon Beauty/Sports/Yelp.
Diffusion-GR2 — Academic: converts GR2 into a block diffusion reasoning re-ranker; 2.4–3.5x inference throughput, accuracy close to AR baseline.
ReasonRec — Meta's three-stage explicit reasoning multimodal recommendation agent; ranking metrics improve >30%, dynamically delegates 35% of queries to efficient sub-models.
Retrieval and Candidate Generation
Meta Hard Negative Sampling — Meta's LLM clustering-based real-time hard negative sampling; online recall +8.5%, popularity bias -12.3%.
CMSL — Meta's multi-sequence learning paradigm: decomposes user history into thematic sequences then models in parallel; deployed on four surfaces.
IID-Nav — Kuaishou models retrieval as autonomous graph exploration, supporting unlimited depth navigation; outperforms baselines on billion-scale industrial data.
MDP Planning Retrieval — VK/Academic: top-K retrieval as ALS posterior MDP; single-step lookahead captures most gains.
PPRO — Baidu's user profile-guided personalized memory retrieval; +12.3% F1 on LoCoMo.
SchemaRAG — Microsoft's dynamic schema reduction RAG framework; +8.8% F1, -47% latency.
Vocabulary Transfer — Amazon's vocabulary transfer framework; ModernBERT achieves 52.4 nDCG (+4.7) on BEIR.
GeoRAG — Academic: information-need-coverage optimized context selection; unsupervised, +6.5–9.7 EM on 6 datasets.
CoPersona — Academic: multi-faceted persona graph collaborative completion for sparse user profiles; +5–10% across domains and model sizes.
IntentTune — eBay uses behavioral signals to disambiguate vague query intent; outperforms demographic profiling.
POEM retrieval linkage via — Kuaishou partial order sequence modeling (see Ranking section).
Ranking and Re-ranking Optimization
POEM — Kuaishou uses multi-task ranking scores to construct partial order real-time sequences; online watch time +0.249% (KS Single Page) / +0.213% (KS Lite Page).
PermR — Avito's adjacent swap constrained re-ranking algorithm; +2% revenue in 14-day online A/B test.
ChunkGroupSHAP — KakaoBank's grouped Shapley values for explaining dense rankers, bridging word-level and dense representation granularity.
Bi-NAS — Academic: bi-level NAS for interpretable recommendation; +5–10% accuracy on 4 datasets.
Contextual Slate GLM Bandits — Microsoft Research India's limited-adaptivity slate bandit algorithm; regret bounds independent of nonlinear parameter κ.
LinkedIn Policy Personalization — LinkedIn deploys hybrid ranking + treatment effect extrapolation framework on job platform; guardrail risk reduced by >10%.
FedCGNM — Intel's grouped normalization momentum federated learning optimizer; consistently outperforms FedAvg on long-tail benchmarks.
EvoRec — Industry's multi-agent self-evolving recommendation framework; online revenue +1.85%, CTR +1.02%.
Other
PixelRAG — Academic: end-to-end visual RAG based on webpage screenshots; 30 million screenshot library; EM improvement up to 18.1% on text tasks.
STRUCTSURVEY — Bloomberg's hierarchical multi-agent structured survey generation framework; +2.9 ROUGE-1.
Gaming Consensus — X reveals vulnerability in matrix factorization bridging mechanism: 10.7% low-quality notes can be manipulated by fewer than 10 ratings; mitigation deployed.
Mandol — Academic: coalesced unified memory system; 5.4x retrieval speedup, 4.8x insertion speedup in long conversations.