Weekly Overview
Industry papers keep coming in dense — 20+ of this week's 37 carry online results or production deployments. Three technical signals stand out. First, the bottleneck in generative recommendation has moved from "can we generate" to "are we generating the right thing." Second, re-ranking and blending are collapsing the whole pipeline into a single autoregressive decoder. Third, agentic recommendation is moving from tool-calling toward self-reflection and reward modeling.
Generative recommendation: quantization error in semantic IDs and decoding drift are now the main battleground. Kuaishou's BRIDGE names the information gap between the RQ-VAE quantization process and the decoder the "semantic quantization gap." It closes three kinds of bias — grounding, neighborhood, and difficulty — with codebook-anchored latent reasoning, privileged EMA self-distillation, and depth-aware compute allocation. On three Amazon benchmarks, Recall@10 improves 10.9% on average; in Kuaishou's online A/B on ads, eCPM rises 0.65%. Meta's CAP-Mix takes the other path — skip hard quantization entirely. Items become a soft mixture of shared prototypes plus an item-specific vector. HR@5 beats a hard SID table with the same tokenizer by 4.45–14.90%, and inference throughput is 1.4–161× the discrete baseline. Both point the same way: SID discretization itself is becoming the SNR ceiling for generative retrieval.
Re-ranking and blending: from hand-tuned rules to end-to-end generation. ByteDance's OneTrans-GSR unifies re-ranking and blending in a single decoder-only Transformer. A reward model scores the full exposure list to guide RL. Online: GMV/u +2.63%, advertiser value +2.06%, repurchase rate +2.03% — and end-to-end latency drops 13%. Taobao's ICE attacks the evaluator instead. It argues that conventional "independent behavior modeling" compresses a continuous browsing journey into per-item CTR predictions, which introduces systematic bias. ICE explicitly models the causal chain of impression → exposure → click. In a 14-day A/B: clicks +0.70%, revenue +3.67%. Both papers answer the same question: list-level objectives can't be assembled from per-item scores.
Agentic recommendation: three branches — reflection, reward, and quality attribution. Kuaishou's EvoRec turns failed trajectories into reusable strategies via a dynamic reflection tree — with no parameter updates at all. Online ad revenue +4.6%, high-intent users +14.5%. Taobao's E-GenRM trains a 27B generative reward model to judge preferences for a shopping agent. Preference accuracy on E-GenRM-Bench hits 87.40% (6.97pp above the strongest baseline); 14-day A/B shows user satisfaction +2.33pp. Kuaishou's WISE points the agent the other way — diagnosing where the system itself is broken.
Generative Recommendation and Semantic ID Optimization
Quantization compresses continuous content into discrete codes. The decoder only sees the codes — not how they were produced. That information asymmetry is the starting point for several papers this week.
BRIDGE (Kuaishou) — The paper splits the gap into three layers: the grounding gap (the decoder lacks semantic anchors when target-side information is limited), the neighborhood gap (token-level supervision discards inter-item similarity and amplifies exposure bias), and the difficulty gap (shallow tokens are easy to generate, deep tokens are hard, yet compute is allocated uniformly). Three mechanisms map to these: codebook-grounded latent reasoning injects frozen quantizer centroids as semantic anchors before generating SIDs; privileged EMA self-distillation transfers neighborhood knowledge from a teacher to an on-policy student, preserving inter-item geometry; effort-aware recurrence allocates more recurrent compute to shallow layers, decreasing with depth. On three Amazon benchmarks, Recall@10/NDCG@10 improve 10.9%/9.7% on average over the strongest baseline, up to 12.6%/10.7%. Kuaishou ran a one-week A/B at 5% traffic: eCPM +0.65%. The gain isn't large, but it elevates "quantization information loss" from an engineering detail to a first-class concern — extending the thread from Gryphon on the mismatch between SID sequence likelihood and relevance objectives.
CAP-Mix (Meta-affiliated) — Directly questions the discrete bottleneck. Each item = a soft mixture of shared prototypes + an item-specific vector. Assignments, prototypes, and item vectors train jointly; inference keeps the soft representation. Two optional interfaces: SID responsibility can serve as an assignment prior, and full content embeddings can initialize item-specific vectors without changing the initial flat table. In matched comparisons on three datasets, with only offline SIDs it beats a hard table with the same tokenizer by 4.45–14.90% HR@5. With full content embeddings it matches or exceeds a strong flat table — while hard assignment degrades on sparse Amazon datasets. Across six benchmarks: HR@5 +6.9%, NDCG@5 +13.6% on average; inference throughput 1.4–161× the discrete baseline. This line of work shares the same tension as DualGR, which got +0.527% video plays online: SIDs buy parameter sharing and pretrained semantics at the cost of information loss.
HERO (Alibaba 1688) — Focuses on a side effect of on-policy distillation: it amplifies popularity bias. Teacher feedback favors popular tokens; student rollouts favor popular items. The two biases reinforce each other, and cold SID paths are rarely generated or supervised. The fix is heterogeneous teacher distillation — two teachers specializing in popular and cold items respectively, routing samples by the ground-truth item's popularity. Diversity-aware item rollout injects ground-truth SIDs of cold items into training to cover rare paths. Inference uses only the student. In a 15-day A/B on the 1688 homepage "Guess You Like": GMV +3.412%. Compared with the two-stage alignment in OpenOneRec (OPD + Rec-RL), HERO's increment is splitting "what to distill" by popularity.
GRC (Alibaba International Digital Commerce) — Targets error accumulation in single-pass decoding: draft generation → multi-granularity reflection → reflection-guided correction. GRPO optimizes the entire GRC trajectory under token-level and trajectory-level rewards. Online serving uses Entropy-Guided Reflection Scheduling to dynamically allocate more correction budget to high-uncertainty trajectories. Up to +15.74% over six SOTA baselines; online ad revenue +1.79%, with manageable latency overhead. This decoding-time correction approach shares the same problem awareness as OneLoc, which uses RL to balance multiple objectives in local services.
ReuseGR (Alibaba) — Makes offline-online decoupling a system-level solution. A 7B offline planner generates a shared SID prefix set for each query; a 0.5B online model selects a prefix using each request's history and completes the remaining tokens. ReuseCut sets cache depth per query based on history coverage. Path KD transfers the 7B's suffix scoring knowledge to the 0.5B. BudgetCut uses GRPO to select branches under a fixed branch budget, optimizing top-100 recall. On industrial search ads data (~2M query-item records, 3.38M items): HR@100 +3.83pp; online latency 38.9% lower than a fully online 0.5B, and comparable HR@100 to a fully online 7B at 57.7% lower latency. 14-day A/B: CTR +0.93%, RPM +1.05%. The "large model plans offline, small model completes online" idea complements the serving-system perspective of xGR.
WeSearch (WeChat) — Challenges the implicit assumption that each document follows a single semantic path: multi-topic documents get flattened under a fixed hierarchical SID. Instead, retrieve query-relevant clusters first, then rank atomic documents within the candidate space — decoupling candidate pruning from fine-grained relevance modeling. Combined with relevance-aware training, compact document representations, and incremental index updates, it supports weekly or even daily model refreshes. Fully deployed on WeChat Search (100% traffic). One-month full-traffic A/B: CTR +0.68%, article clicks per user +2.27%, retrieval resource cost -41%, image-text retrieval latency -19ms.
UAGR (Kuaishou) — Injects downstream multi-objective utility into both SID construction and decoding. Utility-Enriched Semantic Codes encode item-level utility into codes without increasing SID length or vocabulary size, and support incremental updates. Utility-Guided Decoding combines multiple post-exposure behavior predictions into a utility estimate and folds it into beam scoring, preventing high-utility paths from being pruned early during beam search. Fully deployed at Kuaishou; online video watch time +0.206%. The gain is small, but the problem is real — generation likelihood and business utility are not the same thing. Gryphon made the same observation on a music service.
SourceFlow (Tencent-affiliated) — A structural flaw in flow-matching recommendation: the source distribution is sampled independently from global item frequencies, unrelated to target history, so the predictor gets no user-specific source cues. SourceFlow puts collaborative structure into the source distribution itself. Neighborhood Coupling replaces history items with unobserved collaborative substitutes to construct a shifted binary source. Collaborative Consensus refines replacement probabilities using interaction evidence from other users who share the same item. Stretched Restoration preserves these replacements late in the flow to increase the chance of recovering the original interaction. Across five datasets, Recall@10 improves 15.7% on average relative. Compare with Interests Burn-down Diffusion, which replaces Gaussian noise with an interest burn-down process — both modify the "noise/source distribution" stage rather than piling more conditioning onto the predictor.
Re-ranking and Ad Ranking/Bidding Optimization
List-level objectives can't be assembled from per-item scores — three industrial papers this week validate this from different angles.
OneTrans-GSR (ByteDance) — Merges re-ranking and blending into a single learning problem. Blending must handle heterogeneous content — product cards, short videos, livestreams, ad cards. The conventional approach is hand-tuned rules plus low-capacity models per business line, with the combination space and trade-offs artificially locked. OneTrans-GSR uses a single decoder-only policy (extending the deployed OneTrans backbone) to autoregressively generate the exposure list. During training, a reward model evaluates the full list and guides RL toward list-level and long-term objectives. During decoding, business constraints are enforced explicitly. Offline, list quality rises monotonically with reward-model and policy compute — the most important property, because it means this approach can eat scale. Online A/B: GMV/u +2.63%, advertiser value +2.06%, repurchase rate +2.03%; re-ranking + blending end-to-end latency drops 13%. The latency drop is worth noting: replacing a rule engine with generation doesn't just change the modeling approach — it also eliminates multi-stage serial overhead.
ICE (Alibaba Taobao) — In a Generator-Evaluator framework, evaluator fidelity directly determines recommendation quality. But existing evaluators generally use independent behavior modeling (IBM), compressing a continuous browsing journey into per-item CTR predictions. This ignores two physical facts: a click can only happen after an item is displayed and exposed; and items compete for limited user attention. ICE switches to joint behavior modeling (JBM): first predict the browsing-depth distribution to estimate per-position exposure probability, then use predicted exposure to suppress spurious context from unexposed items, and finally introduce a click-distribution matching objective to capture intra-sequence dependencies. Plug-and-play across backbones. Taobao 14-day A/B: clicks +0.70%, revenue +3.67%; now at full traffic. Compared with HPGR's hierarchical preference-aware generative recommendation on APP Gallery, ICE's value is that it doesn't touch the generator — it corrects the evaluator from a "scorer" into a "causal process model."
UniMVT (Kuaishou) — Marketing interventions like coupons introduce confounding bias into CTR estimation: observed clicks are a mixture of intrinsic preference and intervention uplift, so conventional models misestimate base CTR, distorting ranking and billing. Worse, intervention intensity is a multi-valued continuous variable. UniMVT decouples confounders from treatment-sensitive representations. A full-space counterfactual inference module jointly reconstructs debiased base CTR and the intensity-response curve. An auxiliary intensity estimation task captures treatment propensity. A unit-uplift objective normalizes intervention effects for comparability across the continuous coupon-value spectrum. Validated on synthetic + industrial data + online A/B. Compared with multi-task fusion networks like AdaTT, UniMVT brings explicit causal-inference structure into the multi-task framework.
ChoirRec (Alibaba Taobao) — Low-activity users pose a triple challenge for CVR prediction: noisy behavioral signals, insufficient user-level information, and training that systematically favors high-activity users. ChoirRec uses LLMs for semantic grouping. Semantic Group Generation forms reliable user clusters across activity levels to filter noise. Group-aware Hierarchical Representation enriches sparse user embeddings with group priors. A Group-aware Multi-granularity Module uses a dual-channel architecture with adaptive fusion to learn group knowledge. Offline GAUC +1.16%; online order volume +7.24%. The magnitude gap between 7.24% and 1.16% itself shows there's structural opportunity in the low-activity user segment — something offline metrics can't capture.
HOB (Alibaba) — Unified bidding across heterogeneous auction environments. Channels differ in ranking rules (pure eCPM vs UE-enhanced), pricing (first-price vs second-price), and bidding conventions — but the advertiser's budget and ROAS constraints are shared. HOB's core is making marginal cost computable and alignable across heterogeneous channels. In particular, when first-price auctions coexist with organic/paid traffic, existing bidding formulas can't derive an alignable MC form. The global layer derives channel-specific MC forms and coordinates them with a shared MC objective. The local layer models free-win probability and winning-price uncertainty with a zero-inflated exponential distribution. Any interior optimum satisfies cross-channel MC equalization. Deployed on a commercial DSP: GMV +3.0% while maintaining ROAS constraints. Compared with Uncertainty Quantification for Autobidding, which uses Bayesian posteriors to correct noisy CTR/CVR estimates, HOB pushes uncertainty modeling further downstream to the bidding decision itself.
Auxiliary Supervision for Ad Ranking — CVR training labels are sparse. The paper systematically introduces multi-source user events and metadata beyond conversion as auxiliary supervision. Three engineering treatments: linearized labels improve auxiliary signal precision; a split architecture decomposes training tasks; Label Combination (LaCo) handles noisy auxiliary labels. The abstract doesn't disclose online numbers, but the decomposition — turning auxiliary supervision from "add another loss" into "label precision + task isolation + noise robustness" — is directly useful for teams working on CVR.
DCD — Post-click landing experience is generally treated as a static routing problem: which landing page to show, web vs app, native UI vs external site — one-size-fits-all. DCD groups these under post-click destination automation, selecting a destination for each eligible outbound click to maximize advertiser value (AV, value-weighted conversion outcome). The learning objective evolves from posterior destination inference, to two-task destination-conditional conversion prediction, to a single-task policy model that directly judges which of two destinations is better — trained with A2C-style soft labels from a high-capacity critic. Large-scale online A/B on the web-vs-app case shows per-click personalization beats static routing, with the single-task objective gaining the most. This thread sits in the same problem domain as ChorusCVR's handling of ambiguous negatives in full-space CVR — different facets of the same issue.
Can LLM Labels Replace User Clicks? (Baidu co-authored) — Systematically compares two supervision signals. Findings on TianGong-ST and Baidu-Click: click supervision is stronger on high-frequency queries and document-level signals (authority, quality); LLM labels are stronger on mid/low-frequency queries and query-doc semantic matching, relevant/irrelevant discrimination. Based on this complementarity, the authors try two fusion strategies — data scheduling and frequency-aware multi-objective learning — with the latter more effective, improving ranking across all query frequencies. The value here isn't a new method but empirical evidence for "when to use which label." Scaling Dense Retrieval got +1.4% CTR on Walmart sponsored search using LLM cascade labeling — essentially pushing this lesson to larger-scale industrial practice.
OSMF (JD) — Ad image generation has long been optimized for overall CTR, ignoring preference differences across user groups. OSMF first does product-aware adaptive grouping (dynamically organizing users by attributes and product characteristics, each group represented by collective preference features), then uses a Group-aware MLLM to generate customized images per group, and finally applies Group-DPO for group-level preference alignment. It also open-sources the GAIP dataset: ~600K groups, 40M users. Online CTR improves (specific values undisclosed). Compared with UniRank, which uses RLHF to align multimodal re-ranking, OSMF cuts preference granularity at the crowd level rather than the individual level.
LLM and Agent-Driven Recommendation and Search
The main thread for agentic recommendation this year has shifted from "can call tools" to "can learn from failure" and "can be reliably evaluated."
EvoRec (Kuaishou) — A unified agentic framework. A Need Agent infers user needs from long-term, short-term, and working memory. A Planning Agent adaptively allocates retrieval budget between "behavior continuation" and "need satisfaction," selects retrieval tools, and sets ranking fusion weights. An Acting Executor executes and records decision trajectories. When user feedback arrives, a Dynamic Reflection Tree locates whether the failure occurred in need inference, planning, or a prior fix, proposes a personalized correction, and validates it via replay. Validated fixes update working memory and are distilled into cross-user reusable strategies — with no parameter updates throughout. On Kuaishou local-services ad data, Hit@50 is +15.7% over the strongest baseline; on user segments with fix history, hierarchical reflection adds another 13.1%. 7-day online A/B (nearline deployment): ad revenue +4.6%, CTCVR +3.5%, high-intent user revenue +14.5%. The line of work comes from the same team's self-evolving EvoRec, which iterated models via a Research Agent + Code Agent and distilled methodologies from a memory pool via a Skill Evolver — +5.54% offline, +1.85% online revenue. This version moves "evolution" from the model layer to the strategy layer.
E-GenRM (Alibaba Taobao) — A generative reward model for open-ended shopping tasks. On the data side, it constructs diverse, long-horizon dialogue trajectories from naturally occurring user behavior in the shopping app — rather than being framed by existing agent usage patterns and known failure modes. On the modeling side, it first does SFT via a behavior-anchored evaluation chain, then criterion-guided RL — rewarding correct preference judgments and evidence-supported factual claims under each criterion. Personalized hard-negative training makes it more sensitive to implicit (rather than explicitly stated) user preferences in shopping behavior. An active evidence retrieval extension fetches missing e-commerce information for evaluation. E-GenRM-27B hits 87.40% preference accuracy on the authors' benchmark, 6.97pp above the strongest baseline. Taobao 14-day A/B: effective shopping assistance rate +1.39pp, user satisfaction +2.33pp. The bottleneck for shopping agents is often judgment, not generation — this paper puts resources on the judgment side.
WISE (Kuaishou) — Proposes the System Quality Diagnosis (SQD) task: cases where the system runs healthily but produces poor results are common, and existing monitoring covers only operational reliability, not quality attribution. WISE is an end-to-end quality diagnosis agent that traces user-perceivable problems back to specific algorithmic components. Two training supports: ReTrace constructs evidence-consistent trajectory supervision; AHPO (Adaptive Harness-Policy Optimization) uses failure-guided optimization of the agent harness and model policy. The authors also build WISE-BENCH: a real-world multimodal search benchmark with user-side evidence, internal pipeline traces, and expert diagnosis annotations. On Qwen3.5-35B-A3B, end-to-end accuracy is 75.20%; already used in Kuaishou production for quality audits, online quality assurance, and model optimization. Compare with ContextSim, which uses LLM agents to simulate users for offline evaluation — two ends of the same problem: one generates data for evaluation, the other diagnoses production.
CADRE (Tencent) — The difficulty in e-commerce search relevance judgment: evidence from different perspectives may support opposite conclusions, and single-pass reasoning can't reconcile conflicts. CADRE uses three expert agents sharing a frozen backbone to analyze product attributes, user intent, and business rules in independent contexts. A trainable Verifier evaluates each piece of evidence against matching criteria and decides relevance. When experts disagree, a Verifier-guided debate triggers targeted clarification and correction of disputed interpretations. The Verifier is optimized with multi-turn GRPO plus successful-experience replay — trajectories that produced correct final judgments can be used alongside fresh rollouts for subsequent updates. Online serving distills CADRE as an offline teacher into a lightweight student. A/B on an e-commerce search ads platform shows improvements in relevance classification and satisfaction metrics. Same cost structure as An Efficient Long-Context Ranking Architecture's "distill LLM teacher into efficient re-ranking architecture."
LEAPS (Alibaba Taobao) — User search behavior is shifting from discrete keywords to natural-language multi-constraint queries. Existing architectures face a dilemma: precise queries often trigger zero results, while forced simplification yields noisy, overly broad results. LEAPS uses a "Broaden-and-Refine" plugin paradigm with plugins at both ends. Upstream, a Query Expander generates adaptive complementary query combinations to maximize the candidate set, with three-stage training (inverse data augmentation → posterior-knowledge SFT → diversity-aware RL). Downstream, a Relevance Verifier uses CoT to fuse multi-source signals — OCR text, reviews — to filter noise. Non-invasive architecture; doesn't touch the existing short-text retrieval pipeline. Fully deployed on Taobao AI Search since August 2025, serving hundreds of millions of users monthly.
SPADE (Meituan) — Proactive query recommendation on the local-services homepage requires open-set generation over a million-scale vocabulary, with outputs that must simultaneously satisfy preference alignment, fine-grained spatiotemporal matching, long-tail intent coverage, and diversity within a few candidate slots. Four components each address one axis. Objective-Aligned BPE aligns subword merges with business value signals to raise generation probability for long-tail queries. STAR replaces coarse positional tokens with continuous multi-scale encoding plus soft gating. MARS-DPO does preference alignment on real exposure-click data with composite rewards and adaptive hard negatives. SDBS uses sigmoid diversity beam search to improve intent coverage without sacrificing recall. Offline: HR@5 +5.6%, GeoC@5 -40.3%, Exec@5 +1.9%. One-week A/B: module CTR +1.88%, search volume +0.177%, search DAU +0.105%. Compared with PROMISE, which uses a process reward model for test-time scaling, SPADE encodes diversity constraints into beam search itself.
MILES-Rec (Amap) — Utility-driven interactions in a maps app are only indirect, incomplete evidence of consumption needs. MILES-Rec infers implicit service intent from heterogeneous map behaviors and expresses it as prioritized natural-language queries. It first learns progressive reasoning demonstrations (synthesizing user profiles and map histories, doing situational understanding, intent expansion, query ranking), then optimizes with intent-aligned GRPO — rewarding query relevance and post-ranking service coverage with subsequent user interactions, penalizing redundancy and overly long responses. Two-level credit assignment combines list-level shared advantage with query-level advantage, coordinating individual quality with overall complementary coverage. The authors build MILES-Bench covering diverse services under varying behavior-history richness. Stable gains across four backbone sizes.
SSR (Microsoft) — Generative AI search makes "click status" completely obsolete: a click may lead to an unsatisfactory result, and a non-click may be successful answer consumption (good abandonment). SSR (Session Success Rate) trains human-annotated judges using privacy-preserving interaction evidence and screen recordings, then trains an online scoring model on query events, viewport behavior, result-page interactions, and semantic reformulations. On 6,904 annotated queries, balanced accuracy is 0.75; within abandoned queries, 0.68. The samples themselves are telling: 53.2% of abandonments are successful, 6.5% of clicks are unsuccessful. Across seven ranking experiments and five non-ranking experiments, SSR agreed with Sessions per User in 7 cases, was inconclusive in 4, and contradicted in 1. Across ~4,000 production experiments, SSR's binary sensitivity is about 2.5× that of Sessions per User.
BrowseFlow (Alibaba Taobao + Tsinghua) — Two difficulties in user simulation: the same user may produce different browsing responses to the same recommendation feed, and logs record only one realization; and key events like clicks and likes are extremely sparse among many scroll actions. BrowseFlow uses conditional discrete flow matching: each response is represented as a browsing range plus an ordered sequence of key events, directly modeling counts and positions rather than densely predicting each step's action. Random Event Retention defines probability paths over partial event sequences; transition rate matching learns stochastic insertion dynamics. Response fidelity and generation variability are evaluated with energy scores (a proper scoring rule), supplemented by Wasserstein distance on key-event count distributions. The authors also release the TaobaoBrowse benchmark. Best distributional performance on three datasets. Compared with context-aware simulators like KuaiSim, BrowseFlow reframes sparsity from "predict what action" to "model event counts and positions."
GraphMind (Microsoft) — Three stages. Offline, extract action-centric workflow graphs from human-resolved traces (problems, actions, and causal links). Online, a multi-agent traversal engine combines graph retrieval and LLM reasoning to dynamically construct and execute workflows. Adaptive Traversal Reinforcement strengthens successful traversal paths so the graph evolves with execution. Deployed on troubleshooting for four cloud database services. On 93 held-out incidents, blind review beats an Agentic Summary-RAG baseline, requires 8× less retrieval context, and the ATR layer cuts hallucination rate by 26%. In a 12-week field study, 97% of scored conversations produced actionable results within interactive latency. The problem domain is ops rather than recommendation, but "learn workflow graphs from execution feedback" is relevant to long-term memory design in agentic recommendation.
Multimodal Representation and Training System Optimization
Half of this group is squeezing storage and latency; the other half is opening the black box of embeddings.
PersonaAgent (Visa) — Entity interaction patterns in payment networks change faster than production model refresh cycles, and entity embedding tables suffer from high cardinality and sparse interactions. PersonaAgent represents each entity's latest behavioral signals as persona memory refreshable weekly. At inference, a lightweight adapter retrieves the corresponding entity's behavioral signals from memory to enhance the base anomaly classifier's scores — without changing base weights, using retrieval to fill the knowledge-latency gap. Visa deployment: 1,000 transactions batch-processed on CPU in 1–2ms; PR-AUC +0.65–6.58% relative; production business metrics +8.93–14.24% relative. This "memory + retrieval augmentation + lightweight adaptation" structure shares engineering motivation with HyMiRec, which compresses user history embeddings with residual codebooks.
SparseLift (Kuaishou) — In large-scale recommendation training, sparse paths (sparse operators + collective communication) often exist as opaque nodes in the computation graph, blocking whole-graph optimization. SparseLift is a low-overhead compiler for sparse paths: it analyzes model structure and runtime information, refactors opaque nodes into analyzable subgraphs, and recovers resource semantics for sparse compute and communication. On the recovered semantics it designs an incremental cost model and a constraint-aware planner that automatically selects and organizes subgraph optimizations into an execution plan, reused during steady-state training. On an industrial video recommendation task with 100B sparse parameters, steady-state throughput +16.5%; on DLRM, 1.60× and 1.26× faster than Baseline and TorchRec respectively. For 100B-parameter training clusters, 16.5% steady-state throughput means substantial cost savings.
MUSE (Alibaba Taobao) — Lifelong interest modeling has long relied on ID features, generalizing poorly on long-tail items with limited semantic expressiveness. Existing work applies multimodal representations to the GSU (General Search Unit) for behavior retrieval but often neglects fusion at the ESU (Exact Search Unit) stage. MUSE draws two systematic conclusions: in the GSU, simple is enough — lightweight cosine similarity with high-quality multimodal embeddings outperforms complex retrieval mechanisms; in the ESU, richer multimodal sequence modeling and effective ID-multimodal fusion are needed. Deployed on Taobao display ads, supporting 100K-length user behavior sequences with negligible latency overhead. The authors open-source the first large-scale dataset combining ultra-long behavior sequences with high-quality multimodal embeddings. Compared with DMT's topology-aware tower splitting, MUSE's increment is turning "where and how much to use multimodal" into reproducible engineering conclusions.
SA-RSQ (Meituan) — High-dimensional multimodal features are expensive to store and serve. Hard quantization is compact but distorts boundaries; dense soft quantization couples representation quality with storage budget. SA-RSQ uses Top-K sparse routing with softmax weights to store (Index, Probability) tuples, decoupling per-item storage from codebook dimension. Under a fixed support set, gradients propagate through routing weights and weighted reconstruction, without a straight-through estimator. On proprietary food-delivery ad data, within an 8–48 bytes/item storage budget, it achieves a better reconstruction-performance-CTR trade-off. One-week online A/B: CTR +2.51% relative, CPM +3.66%. Compared with Gwhere, which uses contrastive residual quantization tokenizers to align text/visual/spatial/collaborative signals, SA-RSQ focuses more on the quantizer's own design space.
User Embeddings Are Superpositions of Behavioral Patterns — Brings the superposition hypothesis from mechanistic interpretability to user embeddings: dense user embeddings can be understood as sparse superpositions of interpretable behavioral patterns. For linear embeddings, this structure is directly predicted by a user behavior generative model; for deep representations, the superposition hypothesis gives an approximation. Superposed patterns can be naturally recovered with sparse autoencoders; recovered patterns can be automatically named and queried in natural language. On synthetic data, SAE outperforms classical sparse baselines (PCA/ICA/NMF/K-means) — also on Transformer-encoded embeddings. Validated on two real datasets with both classical and Transformer user embeddings; natural-language labels support zero-shot audience retrieval and feature characterization — even when embeddings were never trained with text. The framework is post-hoc and model-agnostic; no changes to the upstream embedding system. Compared with SAERec, which uses SAEs to build fine-grained intent priors from LLM text embeddings, this paper shifts the interpretability target to user-side embeddings. The practical value: interpretability is no longer after-the-fact consolation — it becomes a queryable, reusable retrieval entry point.
UNITE (Kuaishou-affiliated) — Multimodal retrieval is constrained by heterogeneous data sources and complex cross-modal alignment. UNITE attacks from two angles: data curation (the first systematic analysis of how modality-specific data properties affect downstream task performance), and modality-aware training configuration — Modal-Aware Masked Contrastive Learning (MAMCL) mitigates competition between instances of different modalities. SOTA on multiple multimodal retrieval benchmarks; the 7B model surpasses the 26B IDMR. Compared with HIVE, which uses LLMs for query synthesis and verification to improve reasoning-intensive retrieval, UNITE takes the more fundamental path of training data and loss functions.
Directions to Watch
Semantic ID information loss is now treated as a system-level bottleneck. BRIDGE, CAP-Mix, UAGR, and SA-RSQ attack the same problem from four angles — quantizer design, soft representations, utility injection, and sparse soft quantization: discretization itself is capping the information ceiling. CAP-Mix's 1.4–161× inference throughput advantage hints at something: soft representations don't necessarily cost latency. What to watch next: whether soft/semi-soft SIDs can replace hard SIDs at scale online, and whether Gryphon's "joint training of SID generation + item-level scoring" becomes a standard paradigm.
Decoding-time correction and test-time compute are becoming an independent research layer for generative recommendation. BRIDGE's effort-aware recurrence, GRC's reflection-correction trajectory, UAGR's utility-guided beam scoring, ReuseGR's BudgetCut branch selection — all modify the decoding process without touching the backbone. The upside: it eats compute directly, decouples from the backbone, and is cheap to deploy. The risk: if the correction signal comes from a biased reward model, correction amplifies bias rather than fixing it.
Agent evaluation and self-evolution are forming a closed loop. EvoRec turns failed trajectories into reusable strategies (no parameter updates). E-GenRM trains a generative reward model to judge agent behavior. WISE diagnoses system quality problems from the other direction. SSR defines a session-level success metric for generative search. These four point to the same architecture: an agent generates behavior, another model judges behavior, and judgments flow back as training signals or policy updates. Most worth tracking is whether E-GenRM's "reward model as infrastructure" idea can spread from shopping agents to the main search and recommendation pipelines.
Paper Roundup
Generative Recommendation and Semantic IDs
BRIDGE — Kuaishou proposes the semantic quantization gap concept and three bridging mechanisms; Recall@10 +10.9% on three Amazon benchmarks, online eCPM +0.65%.
CAP-Mix — Meta-affiliated proposes prototype soft-mixture item representations to replace hard SIDs; HR@5 +6.9%, NDCG@5 +13.6% across six benchmarks, inference throughput 1.4–161×.
HERO — Alibaba 1688 uses heterogeneous teacher distillation and diversity-aware rollout to mitigate OPD popularity bias; online GMV +3.412%.
GRC — Alibaba International Digital Commerce proposes a generate-reflect-correct decoding framework; up to +15.74% over six baselines, online ad revenue +1.79%.
ReuseGR — Alibaba decouples offline 7B planning from online 0.5B completion; HR@100 +3.83pp, latency -38.9%, online CTR +0.93%, RPM +1.05%.
WeSearch — WeChat proposes a unified retrieval framework of query-cluster retrieval + atomic document ranking; fully deployed, CTR +0.68%, clicks per user +2.27%, resource cost -41%.
UAGR — Kuaishou injects downstream utility into SID construction and decoding; fully deployed, video watch time +0.206%.
SourceFlow — Tencent-affiliated introduces collaborative structure into the flow-matching source distribution; Recall@10 +15.7% relative across five datasets.
Re-ranking and Ad Ranking/Bidding
OneTrans-GSR — ByteDance unifies re-ranking and blending with a single decoder, reward-model-guided RL; GMV/u +2.63%, advertiser value +2.06%, latency -13%.
ICE — Alibaba Taobao replaces independent behavior modeling with joint behavior modeling in the evaluator; 14-day A/B clicks +0.70%, revenue +3.67%.
UniMVT — Kuaishou unifies multi-valued treatment networks to jointly optimize debiased CTR and uplift; online A/B shows business metric gains.
ChoirRec — Alibaba Taobao uses LLM semantic grouping to enhance CVR for low-activity users; offline GAUC +1.16%, online order volume +7.24%.
HOB — Alibaba aligns marginal cost across heterogeneous auction environments; deployed on a commercial DSP, GMV +3.0% while maintaining ROAS.
Auxiliary Supervision for Ad Ranking — Introduces multi-source user events as CVR auxiliary supervision, with linearized labels, split architecture, and LaCo.
DCD — Post-click destination personalization framework, A2C soft-label training of a single-task policy model; online A/B beats static routing.
Can LLM Labels Replace User Clicks? — Baidu co-authored, systematically compares two supervision signals and proposes frequency-aware multi-objective fusion; ranking improves across all query frequencies.
OSMF — JD does group-level ad image generation with Group-DPO alignment; open-sources GAIP (600K groups, 40M users), online CTR improves.
LLM and Agent-Driven Recommendation and Search
EvoRec — Kuaishou uses a dynamic reflection tree for a parameter-update-free self-evolving agent; Hit@50 +15.7%, online ad revenue +4.6%, high-intent users +14.5%.
E-GenRM — Alibaba Taobao trains a 27B generative reward model for a shopping agent; benchmark preference accuracy 87.40%, 14-day A/B satisfaction +2.33pp.
WISE — Kuaishou formalizes the system quality diagnosis task and builds WISE-BENCH; 35B model end-to-end attribution accuracy 75.20%, used in production for quality audits.
CADRE — Tencent uses multi-expert agents + Verifier-guided debate for e-commerce relevance judgment; A/B shows relevance classification and satisfaction gains, distilled student serves online.
LEAPS — Alibaba Taobao uses a Broaden-and-Refine plugin to handle natural-language multi-constraint queries; fully deployed since August 2025, serving hundreds of millions monthly.
SPADE — Meituan's four-component end-to-end generation framework for local-services proactive query recommendation; offline HR@5 +5.6%, online module CTR +1.88%.
MILES-Rec — Amap infers implicit intent from map behaviors and generates prioritized query lists; builds MILES-Bench, stable gains across four backbone sizes.
SSR — Microsoft proposes the session success rate metric to identify good abandonment; balanced accuracy 0.75 on 6,904 annotated queries, ~2.5× production experiment sensitivity.
BrowseFlow — Alibaba Taobao + Tsinghua use conditional discrete flow matching for user browsing simulation; builds TaobaoBrowse, best distributional performance on three datasets.
GraphMind — Microsoft builds self-evolving workflow graphs from ops traces; deployed on four cloud database services, 8× less retrieval context, hallucination rate -26%.
Multimodal Representation and Training Systems
PersonaAgent — Visa uses refreshable persona memory for transaction anomaly scoring enhancement; 1,000 transactions CPU batch in 1–2ms, PR-AUC +0.65–6.58% relative, business metrics +8.93–14.24% relative.
SparseLift — Kuaishou proposes a sparse-path compiler; 100B sparse-parameter task steady-state throughput +16.5%, 1.26× faster than TorchRec on DLRM.
MUSE — Alibaba Taobao systematically analyzes multimodal usage in two-stage lifelong interest modeling; deployed supporting 100K-length sequences, open-sources dataset.
SA-RSQ — Meituan proposes sparse-activated residual soft quantization; better reconstruction-CTR trade-off at 8–48 bytes/item, online CTR +2.51%, CPM +3.66%.
User Embeddings Are Superpositions of Behavioral Patterns — Uses SAEs to recover interpretable behavioral patterns from user embeddings and name them in natural language; supports zero-shot audience retrieval, post-hoc and model-agnostic.
UNITE — Kuaishou-affiliated proposes a general multimodal embedding framework with modality curation + MAMCL; SOTA on MMEB and other benchmarks, 7B surpasses 26B IDMR.