type
Post
status
Published
date
Jul 18, 2026 05:33
slug
rec-weekly-en-2026-W29
summary
This week's recommendation system research clusters around four technical themes: generative recommendation entering industrial deep waters, ranking models evolving toward long sequences and fine-grained semantics, retrieval systems breaking through on heterogeneous indexing and causal optimization, and LLM-enhanced recommendation moving from experiments to engineering deployment. Of the 34 papers, 23 come from industry (18 deployed), and 13 report online A/B results. Theme 1 "Generative Recommendation: From DocID Design to Fine-Tuning Alignment": Alibaba's CRID encodes business value ranking directly into DocIDs, achieving +1.06% GMV on a 300M item catalog at full traffic. GFlowGR fine-tunes generative recommendation with GFlowNet, delivering +0.4% annual revenue in Taobao search ads. Meituan's NONTP extends NTP training signals via temporal contrastive learning and cross-domain learning, lifting online CTR by +1.8% and GMV by +2.1%. Common thread: generative recommendation is shifting from "being able to generate" to "optimizing better." Theme 2 "Ranking Models Pursue Deep Decoupling and Long-Term Modeling": Meta's SlimPer formulates personalized ranking as iterative refinement of a <user, item> knowledge base, supporting 10k+ historical events with O(N) complexity, deployed on Instagram. Yandex's Long-History User Transformers decouple long-history inference via offline encoding + caching + a lightweight online model, achieving +2.77% in search ads. Alibaba's SAM uses satiety-gated explicit modeling of interest lifecycles, reducing post-purchase repetition rate by 60%. Theme 3 "Engineering and Causal Paradigms in Retrieval": Pinterest's causal retrieval framework reduces shopping triggers by 85% without harming key sessions. MESH uses modular architecture and gated bias correction to boost the scaling exponent for fresh items by 14x, with user retention +0.46%. Microsoft's FlashTrie fully migrates constrained decoding for generative retrieval to GPU, handling an
tags
Recommendation Systems
Weekly
Papers
category
Rec Tech Report
icon
📚
password
priority
1
Weekly Overview
This week's recommendation system research clusters around four technical themes: generative recommendation entering industrial deep waters, ranking models evolving toward long sequences and fine-grained semantics, retrieval systems breaking through on heterogeneous indexing and causal optimization, and LLM-enhanced recommendation moving from experiments to engineering deployment. Of the 34 papers, 23 come from industry (18 deployed), and 13 report online A/B results.
Theme 1 "Generative Recommendation: From DocID Design to Fine-Tuning Alignment": Alibaba's CRID encodes business value ranking directly into DocIDs, achieving +1.06% GMV on a 300M item catalog at full traffic. GFlowGR fine-tunes generative recommendation with GFlowNet, delivering +0.4% annual revenue in Taobao search ads. Meituan's NONTP extends NTP training signals via temporal contrastive learning and cross-domain learning, lifting online CTR by +1.8% and GMV by +2.1%. Common thread: generative recommendation is shifting from "being able to generate" to "optimizing better."
Theme 2 "Ranking Models Pursue Deep Decoupling and Long-Term Modeling": Meta's SlimPer formulates personalized ranking as iterative refinement of a <user, item> knowledge base, supporting 10k+ historical events with O(N) complexity, deployed on Instagram. Yandex's Long-History User Transformers decouple long-history inference via offline encoding + caching + a lightweight online model, achieving +2.77% in search ads. Alibaba's SAM uses satiety-gated explicit modeling of interest lifecycles, reducing post-purchase repetition rate by 60%.
Theme 3 "Engineering and Causal Paradigms in Retrieval": Pinterest's causal retrieval framework reduces shopping triggers by 85% without harming key sessions. MESH uses modular architecture and gated bias correction to boost the scaling exponent for fresh items by 14x, with user retention +0.46%. Microsoft's FlashTrie fully migrates constrained decoding for generative retrieval to GPU, handling an 800M keyword catalog in <3ms, driving online revenue +0.71%.
Theme 4 "Lightweight and Agentic LLM Recommendation": Vrbo uses training-free LLM-synthesized queries to cover long-tail items, closing the recall gap between a 3B model and an API model to <1%. QuintoAndar's LLM re-ranking fuses conversational context in real estate search, yielding CTR +5.3%. Kuaishou's RashomonLLM couples explanation generation with prediction, improving live-stream CTR AUC +2.3% and explanation quality +8.7%.
Generative Recommendation: DocID Design, Fine-Tuning Alignment, and Engineering Frameworks
This week saw 7 papers on generative recommendation, 6 from industry deployment. Three sub-problems are in focus: how DocIDs encode business value, how fine-tuning aligns with set-level objectives, and how engineering frameworks decouple feature logic from model architecture.
CRID (Alibaba) — Cluster-Ranked Identifier. Existing semantic ID-based DocIDs (e.g., TIGER, RQ-VAE) have two problems: collisions hurt retrieval accuracy, and the encoding objective (reconstructing semantic similarity) mismatches the business objective (GMV). CRID decouples DocIDs into two levels: first cluster items into semantic cluster IDs, then rank items within each cluster by business value (e.g., expected revenue) to produce ranking IDs. This supports incremental updates — new items only need re-ranking within their cluster without affecting the global structure. On a 300M Taobao item catalog, CRID's Top-K Hitrate exceeds the strongest embedding retrieval baseline (Hitrate relative improvement ~6-8%), and full-traffic deployment yields +1.06% GMV. The paper also provides an analytical framework that decomposes retrieval gains into personalized preference and statistical prior generalization, finding that cluster size balances the two. Related work traces back to DSI and TIGER; CRID injects ranking intent directly at the DocID level rather than relying solely on implicit learning via attention.
GFlowGR (Alibaba) — Generative recommendation faces two train-deployment mismatches: NTP optimizes only single-step prediction but must generate a set during inference, and all user interactions are weighted equally. GFlowGR introduces a GFlowNet fine-tuning framework, modeling generative recommendation as sequential set generation. It has three components: a trajectory sampler constructs training trajectories from candidate sets for set-level learning, a behavior-aware reward model quantifies item utility, and the GFlowNet objective provides token-level supervision. The key insight is that GFlowNet's flow-matching property makes probability proportional to reward, without needing policy gradients from RL. Deployed in Taobao search ads, it yields +0.4% annual revenue (billion-level impact). GFlowGR differs from alignment methods like RLHF and DPO: those optimize pairwise preferences, while GFlowNet natively aligns with set-level diversity-value trade-offs.
NONTP (Meituan) — NTP has two structural limitations: temporal locality (optimizes only single-step predictions) and spatial locality (gradients in cross-domain sequences propagate only through the previous hidden state). NONTP extends signal coverage with two auxiliary objectives: temporal contrastive learning (TCL) uses a BYOL-style EMA teacher to align hidden states with K-step future trajectories, and cross-domain learning (TDL) average-pools cross-domain hidden states and predicts via a shared head. Both are discarded at inference, adding zero overhead. On Meituan's four-domain ranking dataset, HR@10 improves 34.3% over NTP, with online CTR +1.8% and GMV +2.1% (p<0.01). This work continues the multi-domain generation paradigm of MBGR but is the first to systematically analyze NTP's locality bottlenecks.
Prompt Generation (Taobao) — A config-driven feature decoupling framework. In generative recommendation, tight coupling between feature processing logic and model architecture leads to slow iteration and heavy deployment. PG uses two declarative JSON files to uniformly describe feature types (raw values/sequences/text/indexes) and processing components (mappers/poolers/aggregators), enabling feature experiments via config changes alone. The framework also includes token compression for very long sequences. On Taobao search, it drives +0.47% transactions and +0.51% GMV, and has been rolled out to multiple teams. This complements xGR's service-side optimization — PG tackles model-side engineering efficiency.
TmallGS (Alibaba) — A large-scale ranking architecture for Tmall search, with five components: hierarchical distribution-aware tokenization (FSR+DCP), field-adaptive gated Transformer, decoupled FiLM late fusion, context-aware bias network, and error-aware progressive training. This work demonstrates feature heterogeneity handling during the transition from DLRM to Transformer — instead of tokenizing and concatenating all features like OneTrans, it assigns projection subspaces per field. Online UCTCVR and GMV both improve substantially.
FlashTrie (Microsoft) — The constrained decoding bottleneck in generative retrieval. The standard practice uses a CPU-based Trie for beam-search constraint, but this becomes a latency bottleneck as beam width increases. FlashTrie migrates constraint decoding entirely to GPU: an integer-aware compact trie layout (bit-compressed to reduce memory footprint) and collaborative CUDA kernels that handle beam expansion, validation, and pruning on-device. With 800M keywords and beam width 1000, trie search latency is <3ms, delivering a 24x speedup over an optimized multi-threaded CPU baseline. An A/B experiment on a commercial search engine shows +0.71% revenue lift. This provides a viable path for generative retrieval to use larger beam widths in latency-sensitive scenarios like sponsored search; the Generative Conversational Recommender System faced a similar decoding constraint.
GUIDE (Alibaba) — Generative auto-bidding. A Decision Transformer models historical bids and environmental states, a Q-value module guides exploration, and an inverse dynamics module provides safe fallback. Online, it lifts ad GMV +4.10%, ad clicks +1.40%, ad cost +1.66%, and ad ROI +3.52%. This work follows the generative bidding line of Constraint-Aware Generative Auto-bidding but is the first to explicitly unify exploration, protection, and selection.
Ranking Models: Long-Sequence Decoupling, Fine-Grained Semantics, and Debiasing
Highlights in ranking this week center on decoupled architectures enabling real-time long-history inference, fine-grained semantic modeling (satiety, discount rates), and debiasing (habit decoupling, position bias).
SlimPer (Meta / Instagram) — Redefining the role of Transformers in recommendation. The large intermediate tensors of generative models are a poor fit for recommendation, which only needs a single relevance score per <user, item> pair. SlimPer models personalized ranking as iterative refinement of a compact <user, item> knowledge base (KBase). Each layer selectively queries user-side multimodal tokens at O(N) cost, computes explicit relevance matching scores, and refines the KBase. Model depth is decoupled from sequence length, supporting 10k+ fine-grained historical events. At inference, request-level optimization shares user-side tokens across candidate items, further reducing memory. The architecture unifies sparse/dense/sequence features, and the attention mechanism provides intrinsic interpretability. Deployed on Instagram Reels and Feed, it improves user engagement. This design complements HiGR's hierarchical generation approach — one compresses history at the ranking stage, the other generates sets at the retrieval stage.
Long-History User Transformers (Yandex) — A classic decoupling with offline encoding + caching + lightweight online model. An offline Transformer asynchronously encodes the user's full cross-surface interaction history into a compact representation (6000 dimensions), stored in a feature store. The runtime model uses only the cached representation + recent events + request context for CTR prediction. The offline encoder is pre-trained with dual objectives (feedback prediction + next-item prediction), then fine-tuned on ad surfaces. This split architecture recovers 72-80% of the quality attained by a full-history runtime Transformer (which is too slow for deployment), and the cache is robust to staleness. Search ads improve +2.77%, ad network +2.1%, with revenue lifts of +2.26% and +0.43% respectively.
SAM (Alibaba) — In e-commerce, a purchase often signals the end of an intent rather than its continuation, but sequential models treat all user interactions as positive signals, leading to post-purchase redundancy (PPRR). SAM explicitly models interest lifecycles via a satiety-aware mechanism: dual cross-attention suppresses historical clicks on already-satisfied interests, an adaptive satiety gating unit outputs time-sensitive soft masks, and a self-supervised TTNP task learns product repurchase cycles. Online PPRR is reduced by 60%.
DANet (Alibaba) — Discount rates affect CVR but are often overlooked. DANet uses Fourier transforms to capture long-term discount trends, a distribution debiasing module to mitigate user discount rate bias, and a regression auxiliary task to provide explicit discount labels. Offline AUC improves +1.61%, online pCVR +3.63%, GMV +2.23%.
OrDA (Ant Group) — Clicks on homepage marketing blocks are driven by both content interest and access habits; habitual clicks create false positive samples. OrDA decouples interest and habit: a gated allocation layer adaptively routes features, orthogonal regularization constrains the two latent spaces to be geometrically perpendicular, and at inference do-calculus ranks only by purified interest. Online UCTR improves +5.64%.
RecRec (Academic) — Decouples reasoning from prediction by distilling hidden states into multi-interest representations via a context compressor, with a recursive reasoner that progressively refines interests in an intermediate latent space. Outperforms SASRec, BERT4Rec, etc. on four datasets, and reasoning depth can be freely adjusted at inference.
Another non-deployed paper with the same name: RecRec (Academic) — A lightweight recursive refinement model (3.9M-14M parameters) that maintains a compact hidden state and updates via evidence-anchored gated recursion. Matches or exceeds larger models. Both RecRec approaches share the same trend: recommendation models need multiple reasoning/refinement steps, not a single forward pass.
MMRM (JD) — Multi-task multi-modal representation. MMRM uses a shared backbone with task-specific tokens and projection layers to simultaneously align multiple collaborative signals, introducing multi-faceted user representations (different user representations for different tasks retrieved from item embeddings). Deployed on JD search.
RashomonLLM (Kuaishou) — Couples explanation generation with prediction. Introduces the Rashomon explanation set concept, proving that explanation fidelity constraints provide an upper bound on model performance. An agentic workflow iteratively aligns explanations with predictions. On Kuaishou live-stream CTR, AUC improves +2.3% and explanation quality +8.7%. This work breaks the long-standing accuracy-interpretability trade-off narrative in XAI.
Long-term User Engagement Optimization (Pinterest) — A model-agnostic downstream reward framework. Offline, it screens session-level behaviors (e.g., saves, shares) for early-observable, retention-predictive signals, constructing downstream rewards that are added to the ranking model. Deployed across Homefeed, Related Pins, Search, and Notifications, with consistent retention improvements.
Retrieval Systems: Heterogeneous Indexing, Causal Triggering, and Multi-Modal Retrieval
At the retrieval level, the week revolves around unified heterogeneous content scaling, causal trigger optimization, and multi-modal/multi-language retrieval, with most papers from industry deployment.
MESH (Pinterest) — Scaling bias in heterogeneous retrieval systems: model capability improvements disproportionately benefit high-frequency content. MESH uses a modular architecture to partition feature spaces, each domain scaled independently, with gated bias correction reducing interference between sparse content and high-frequency features. Experiments show a 14x improvement in the scaling exponent for fresh items (power-law scaling exponent). On Pinterest's Related Pins (billion-scale recommender system), fresh items see +5.5% repins, user retention +0.46%, and funnel efficiency +55%. An asynchronous serving strategy boosts throughput 2.87x. This work provides a scalable paradigm for converging a fragmented "retrieval zoo."
Causal Retrieval Optimization at Pinterest (Pinterest) — Treats shopping trigger candidate generators as causal decisions: trigger only when helpful, not distracting. A multi-task deep model jointly predicts baselines and uplifts for multiple events, trained with doubly robust pseudo-outcomes. A linear-time offline replay strategy accurately predicts online effects. Online shopping triggers are reduced by 85%, total sessions +0.26%, Pin saves +1.10%, with no latency regression.
FlashTrie was analyzed in the generative recommendation section, but it is fundamentally an engineering breakthrough on the retrieval side.
SilverTorch (Meta) — Migrates ANN indexing and filtering services from CPU into a unified model layer. A model-based GPU Bloom index + fused Int8 ANN kernel, with the OverArch scoring layer supporting multi-task retrieval. On industrial datasets, throughput is 23.7x and cost efficiency 13.35x. Deployed with hundreds of models online, supporting multiple applications.
Apple Music Multi-Language Semantic Retrieval (Apple) — A 305M-parameter Siamese bi-encoder (based on GTE-multilingual-base) with curriculum-scheduled multi-objective training (Hit@10 +69% relative improvement). Deployed via a hybrid retrieval architecture (dense + sparse) using quantile distribution matching, requiring no retraining of downstream rerankers. Across 150+ global storefronts, CR improves +2.28%, no-result rate drops -86%, and tail query CR improves +7.93%.
Apple TV Incremental Search Personalization (Apple) — A two-tower system mixing text embeddings (TextEmb, fine-tuned with contrastive learning) and ID embeddings (IdEmb), with XGBoost reranking. Short prefix queries (1-3 characters) see NDCG@10 improve +8.63%, online tap-through rate +1.14%, conversion rate +1.23%.
Walmart EBR Upgrade (Walmart) — Mixed hard negative sampling (online cross-batch sampling + offline cross-encoder metadata mining) + old-model warm-start distillation from DistilBERT to GTE-base. NDCG@5 improves +7.34%, revenue +0.50%.
Vrbo LLM Candidate Generation (Vrbo / Expedia) — A training-free LLM candidate generation pipeline. LLMs are used to synthesize diverse semantic queries for each property, pre-trained text encoders embed them, and ANN indexes retrieve from an 11.7M property catalog. A union fusion strategy merges this with item-level KNN without degrading popular properties. The system covers tens of thousands of long-tail properties unreachable by IBKNN, and the recall gap of a 3B open-source model (27-46% vs. API models) is reduced to <1% via fusion.
CwA (Meta FAIR) — Jointly learns database partitioning and probe functions. Uses an auction algorithm to balance partitions, achieving 4.7x throughput improvement in OOD scenarios. For in-distribution scenarios, simple linear probe functions outperform deep neural network methods.
Proximity Features (Airbnb) — Privacy-compliant cold start. Groups ~1000 nearby users via geographic IP adaptive clustering, generating aggregated signals without persistent identifiers. Deployed on marketing landing pages and destination recommendations, driving substantial booking increases.
ZoRRO (JP/Politikens Hus) — Zero-weight, training-free news recommendation. Online CTR approaches SOTA deep learning models, with inference speeds over 600x faster. Reveals offline-online gaps and cases where CTR is the same but recommendation distributions differ.
AGREE (Alibaba) — Uses MLLM cross-attention maps as local relevance supervision, jointly trained with global labels for retrieval. On ViDoRe V2, nDCG@1 improves +12.82%.
Mutable Sketches (Academic) — Real-time recommendation without retraining. A KP-tree stores user preferences, low-rank projections are fixed, and embeddings are recomputed in real time as new ratings arrive. On KuaiRec, 1.8% of the data achieves 0.810 RMSE (ALS: 0.822), and new users get recommendations in <1ms. Proves that each new rating monotonically tightens the error bound.
LLM-Enhanced Recommendation: Reranking, Agent Architectures, and Feature Fusion
The role of LLMs in recommendation systems is shifting from fully replacing traditional models to hybrid architectures: LLMs handle unstructured/heterogeneous context, while traditional ML handles structured signals at scale.
LLM Re-Ranking for Real Estate (QuintoAndar) — Introduces LLM re-ranking for conversational property search. Constructs a dataset of 960k query-item pairs (synthetic + production queries, LLM-as-a-Judge + human verification). Online CTR improves +5.3%, scheduled visits +4.8%. This work shows LLM reranking contributes substantially to ambiguous-intent scenarios (like real estate), but only if the input context is sufficiently rich (multi-turn dialogue).
Agentic Recommendation for CTV (Meta/Amazon) — An LLM orchestrates specialized components to handle heterogeneous context (trending news, cross-surface activity). The key challenge is LLM inference latency — solved by an agent architecture where each subtask is handled by the most appropriate method (sub-modules can be fully LLM or traditional ML). The engineering contribution lies in how to plug LLMs into an existing pipeline without slowing overall inference.
Tokenizing Numerical and Embedding for LLM RecSys (Meta/Amazon) — Maps continuous numerical values and dense embeddings into soft tokens fused into the LLM input space. In a shared-parameter two-tower LLM recommendation setup, interactive fusion outperforms direct concatenation. Improves three Amazon benchmarks.
AWA-RL (Alibaba) — Hallucination mitigation for search agents. Traditional RL rewards correct answers but does not penalize fabrication when retrieval fails. AWA-RL dynamically adjusts a refusal reward based on the model's prior ability before querying and on-policy training observations. Introduces the RA-F1 metric to balance capability and reliability. Absolute precision improves 10.3%, RA-F1 +2.9%.
Directions to Watch
Fine-tuning alignment for generative recommendation is becoming critical for industrial deployment. Both GFlowGR and NONTP point to the same conclusion: after NTP pre-training, a fine-tuning stage adapted to set-level metrics is essential to maximize gains. This raises a new question: how to design fine-tuning objectives that align more directly with recommendation business goals (GMV, retention, etc.), rather than relying solely on NTP or ranking losses.
Decoupled architectures are becoming the standard solution for long-sequence inference. SlimPer's O(N) knowledge base refinement and Yandex's offline encoding + caching share the same core idea — move compute-heavy parts offline asynchronously, keeping only lightweight matching online. For production systems with user histories exceeding tens of thousands, this is the optimal cost-performance path. Future work may yield even more unified decoupling frameworks: online components with just a few MLP layers and attention pooling, offline with scalable Transformers.
Unified indexing for multi-modal retrieval is emerging as a challenge. Three independent works — Apple Music, Vrbo, Walmart — all employ hybrid retrieval (semantic + behavioral/ID) with fusion strategies (quantile matching, union fusion). This shows that pure semantic or pure behavioral approaches are insufficient; the industry needs ways to painlessly incorporate new signals without retraining downstream models. MESH's modular scaling and CwA's joint learning offer promising directions.
Paper Roundup
Generative Recommendation and Retrieval
GUIDE — Alibaba proposes Generator-DT+Q-value+IDM auto-bidding framework; online ad GMV +4.10%.
GFlowGR — Alibaba uses GFlowNet for generative recommendation fine-tuning; Taobao search ad annual revenue +0.4%.
FlashTrie — Microsoft accelerates constrained beam search on GPU; 800M keywords <3ms; online revenue +0.71%.
CRID — Alibaba embeds business value ranking into DocIDs; 300M item catalog GMV +1.06%.
NONTP — Meituan extends NTP signals with temporal contrastive + cross-domain learning; online CTR +1.8%, GMV +2.1%.
Prompt Generation — Taobao uses config-driven framework to decouple feature logic from model architecture; online GMV +0.51%.
TmallGS — Alibaba proposes five-component large-scale ranking architecture; online UCTCVR and GMV improvements on Tmall search.
Ranking Models and Sequential Recommendation
SlimPer — Meta models ranking as iterative refinement of a <user, item> knowledge base; 10k+ history; deployed on Instagram.
Long-History User Transformers — Yandex decouples long history with offline encoding + caching; search ads +2.77%.
Long-term User Engagement — Pinterest proposes model-agnostic downstream reward framework; deployed across multiple surfaces; retention improvements.
OrDA — Ant Group orthogonally decouples access habits from interest; online UCTR +5.64%.
DANet — Alibaba models discount rates with Fourier transforms; pCVR +3.63%, GMV +2.23%.
SAM — Alibaba models interest lifecycles with satiety gating; PPRR reduced by 60%.
MMRM — JD aligns MLLM with multi-task objectives and introduces multi-faceted user representations; deployed on JD search.
RashomonLLM — Kuaishou uses explanation-prediction coupling agent; live-stream CTR AUC +2.3%, explanation quality +8.7%.
RecRec: Latent Interests — Academic proposes multi-interest recursive reasoning; outperforms SASRec, etc.
RecRec: Recursive Refinement — Academic proposes recursive refinement for sequential recommendation; 3.9M-14M parameters match SOTA.
Retrieval Systems and Engineering Optimization
Causal Retrieval at Pinterest — Pinterest uses causal uplift modeling; 85% fewer shopping triggers; total sessions +0.26%.
MESH — Pinterest modular architecture addresses heterogeneous scaling bias; fresh repins +5.5%, retention +0.46%.
SilverTorch — Meta unifies ANN indexing into a model layer; throughput 23.7x, cost efficiency 13.35x.
Apple Music Multi-Language Retrieval — Apple uses 305M two-tower curriculum training; global CR +2.28%, tail CR +7.93%.
Apple TV Incremental Search — Apple mixes text + ID embeddings; short-prefix NDCG +8.63%, tap-through +1.14%.
Walmart EBR — Walmart uses mixed hard negative sampling + warm-start distillation; NDCG@5 +7.34%, revenue +0.50%.
Vrbo LLM Candidate — Vrbo uses training-free LLM synthesized queries for long-tail coverage; recall gap reduced to <1%.
Proximity Features — Airbnb uses geographic clustering for privacy-compliant cold start; substantial booking lifts.
ZoRRO — JP/Politikens Hus proposes zero-weight training-free news recommendation; 600x speedup, CTR near SOTA.
CwA — Meta FAIR jointly learns partitioning and probing; OOD throughput 4.7x.
AGREE — Alibaba uses MLLM attention maps for fine-grained relevance supervision; nDCG@1 +12.82%.
Mutable Sketches — Academic proposes KP-tree for retraining-free recommendation; 1.8% data reaches 0.810 RMSE.
LLM-Enhanced Recommendation
LLM Re-Ranking Real Estate — QuintoAndar uses LLM reranking fused with conversational context; CTR +5.3%.
Agentic Recommendation CTV — Meta/Amazon uses LLM agent orchestrator for heterogeneous context; addresses LLM inference latency.
Tokenizing Numerical for LLM — Meta maps numerical values and embeddings into soft tokens fused into LLM.
AWA-RL — Alibaba uses refusal-aware RL to mitigate search agent hallucination; precision +10.3%.
Other
Accelerating A/B Tests — Academic proposes Δ-Off-Policy Estimation leveraging policy overlap to reduce A/B variance.
KuaiLive Dataset — Kuaishou releases first real-time interactive live-stream recommendation dataset: 23,772 users, 452,621 streamers, 21 days of logs.