跳转至

Browse by topic

简体中文 · English

Browse two levels: research area, then method family. A paper can appear in more than one method family. Each occurrence has its own entry; open the linked paper for the mechanism and its scope.

Presentation generation and personalization

Title and display content optimization

  • GESE: Explore sets of diverse, faithful headlines, then select a personalized headline using current context. The paper's production A/B reports +2.57% CTR and +0.87% dwell time, not local results.

Personalized creative generation

Ranking and long-sequence modeling

  • HELIX: HELIX interleaves SeqFormer and MixFormer while information flows only from reusable user-only sequence state to candidate-conditioned mix tokens. User computation can be shared across candidates while both axes still interact across depths.

Multitask objective isolation

  • Light Heads: Resetting the head at new training windows aligns head training age across upstream models. The paper also compares not resetting, which can help sparse targets, and removing isolation, which can hurt the original task.

Pretraining and parameter-efficient transfer

  • Progressive Learning from Content to Relevance to Personalization for CTR Prediction: Train randomly initialized CTR embeddings rather than copying VAR vectors. Fielded BPE/N-gram text, query-conditioned attention and native IDs feed ordered fusion. Not reproduced due to cost (review complete); research material only, no executable entry.
  • LazFormer: Adapt to ranking using a ranking residual adapter, dense recent and sparse distant tokens, and hybrid sparse attention. A two-week A/B test reports IPV +5.21% and GMV +9.85%.

User state and preference repair

Foundation model capabilities in recommendation

LLM / Foundation model + Recommendation

  • AgentWebRec: Fetch candidate semantics from the platform, then retrieve private memories for the target user by semantic relevance and time decay. Only when local evidence confidence is too low does the method query compact preference summaries from neighboring agents.
  • RPTune: Higher-priority items go near the prompt end. Post-train selection on the curated catalog.
  • PromptShift: Drift and SliceShift measure the resulting shift. Mitigation blends original score and inverse group popularity according to each user's mainstreamness.
  • Google Music LLM Rationales: Candidates get explanations without putting an LLM in the request path. The multi-arm holdback separates new nominations from rationale annotations and finds that rationales drive most of the shelf-engagement lift.
  • ConnectionMind: SFT follows shortest paths to positive feedback; GRPO rewards formatting, recommendation F1 and shortest-path progress with weights 0.2, 0.5 and 0.3. Production routes the most complex 5%–10% of users to the full LLM and uses a distilled GNN for the rest.
  • DREAM: Unified Outlet accepts only typed parameters on the allowlist and within bounds. The offline evaluator explores and scores strategies; online results validate findings before writing them into memory, closing the Reward Dual Loop.
  • Netflix GenRec: Phase 1 continues training an open LLM on Netflix content and user data to create an internal foundation model. Phase 2 post-trains it on more frequently refreshed ranking data, labels, and reward signals.
  • Open Web UFM: Pretrain a browsing-history Transformer with masked modeling and sequence contrast, then reuse its user embedding for CTR and win-rate tasks. An LLM-in-the-loop search improves recipes using auditable lifters.
  • RecoReward: Reward a description by how much more it appeals to target than non-target users. Online serving consumes descriptions without user profiles.
  • Melo: Orchestrate intent, retrieval, recommendations, and explanations across agent nodes. Catalog grounding prevents unsupported entities; failures trigger reflection and retry.
  • EGR: One encoder produces indexed item vectors and user query vectors. Train item representation learning and next-item prediction jointly throughout.
  • RECAP: The LLM handles semantic updates, while deterministic logic handles lifecycle, merging, and eviction. An LLM judge screens for label-consistent behavior pairs, a two-tower evaluator learns from them, and its matching score becomes the GRPO reward for closing the profile-update loop.
  • RecGPT-V3: 1. Memory Hub compresses behavior into traceable memory units with behavior patterns, preference summaries, raw behavior indices, and timestamps. New behavior updates related units or creates new ones. 2. Hybrid-modal foundation model fuses content with CN-CLIP/Q-Former and quantizes products into SIDs with a two-level RQ-VAE.
  • Prompt Generation: The same interpreter serves training and production; mergers fuse and compress inputs, with parameter-free embedding-space mean pooling a strong default. Event tracking captures online source features for offline replay.
  • Apple Music ELISE: A Siamese bi-encoder complements a lexical inverted index. Quantile distribution matching calibrates dense and token scores to one scale and limits semantic results.
  • NEXT: After positive engagement, serving adds high-confidence edges. Training uses perception-enhanced GRPO, distribution-aligned SFT and last-mile preference optimization.
  • TokenMinds: An LFV/SFV context token supports one encoder prefill and parallel context-specific decoding. Production caches asynchronous heavy-model outputs for real-time rankers.
  • Atomic Intent Reasoning: Move expensive LLM reasoning offline and convert individual behaviors into cacheable hierarchical atomic intents. At serving time retrieve intent chains for the target product and aggregate cross-domain evidence with multi-head attention.
  • ToolRec: Augment a KTO-style objective with user calibration and frequency-driven dynamic system calibration. Improve tool-triggering queries while limiting contextual relevance loss.
  • Pinterest Complementary LLM Predictor: First, SFT predicts one next advertiser. Then GRPO generates 20 advertisers and 5 interests, rewarding the hit position and enforcing output lengths.
  • L2Rec: Towards Dual-View Understanding of LLMs for Personalized Recommendation: User-routed dual-view LoRA-MoE adapts a shared LLM to semantic and behavioral views, then adaptive cross-view gates fuse them.
  • MuChator: Continue pretraining, supervised fine-tuning and reinforcement learning in music so the conversational model can infer current intent and actively discover options instead of merely repeating past preferences.
  • AKT-Rec: The stronger InfoNCE direction updates tail vectors against stop-gradient head teachers. One activity gate balances cluster vs individual embeddings; the other fuses instance-level target attention with cluster-level history for CTR.
  • HARNESS-LM: Sponsored Search can encode documents offline but needs a cheap online query encoder. Train a strong teacher, align a small query encoder to its embedding space via L2, then refine with supervised contrastive learning against the frozen teacher index.
  • TubiFM: Unified Item, Carousel, and Search Ranking for Streaming Discovery: Represent cross-page journeys as a shared user story. Task prompts let the same generative foundation model rank items, carousels and search.
  • LLM Retrieval: A retrieval-domain LLM extracts hierarchical category, brand, product, and context attributes from titles, descriptions, and other creative fields. Category-based retrieval is followed by deeper attribute reranking and ad-to-ad semantic graph construction, then graph expansion from user seed ads.
  • FLUID: The two are separate late-fusion tokens. Candidate IDs are removed through three stages: add slice, remove ID, add room.
  • MM-LLM: Item tokens and user-history token profiles join ID and visual features in the production ranker. LLM inference stays off the online critical path.
  • UniVA: 1. The SID has two semantic levels and one commercial token.
  • RecGPT-Mobile: The training mixture is behavior/search logs 60%, co-purchase 20%, GPT rewrites 15%, and human labels 5%. LoRA SFT learns behavior-to-query mapping.
  • Modular Representation Compression: Adapting LLMs for Efficient and Effective Recommendations: MARC separates compression from task adaptation and uses information constraints to decouple the proxy task from recommendation. It retains task-relevant information while reducing dimension.
  • LWGR: Production precomputes the LLM world knowledge on a seconds-to-minutes nearline schedule and stores it in a user vector store. Requests need one lookup and one cross-attention layer.
  • SOLARIS: SOLARIS predicts which pairs will be accessed, asynchronously precomputes their latents with a foundation model, and caches them. Requests use cached latents on hits and fall back on misses.
  • Cross-domain KD: Low-traffic products cannot train a large specialized teacher. Apply a data-rich YouTube multitask teacher directly to YouTube Music examples, and distill both serving and auxiliary logits into a small model without requiring source-domain data or matching interface features.
  • IDProxy: Multimodal semantics differ from long-learned collaborative item-ID space. Contrastive coarse alignment precedes multilayer hidden-state proxies, injected into the ranker through a residual gate.
  • SIGMA: Inference generates prefixes with beam search and performs ANN only within their subcatalogs. Adaptive probabilistic fusion (APF) trades off precision and diversity according to the spread of prefix confidences.
  • Generative Pseudo-Labeling for Pre-Ranking with LLMs: An LLM derives interest anchors from user behavior and labels unexposed retrieval candidates. The pre-ranker learns from real and pseudo-supervision without calling the LLM online.
  • RGAlign-Rec: Ranking-Guided Alignment for Latent Query Reasoning in Recommendation Systems: Generate latent queries from context with an LLM. Use item preferences from a real ranker to build positive and negative queries, then align high-ranked intents through RG-SFT and DPO.
  • S-GRec: Semantic and revenue objectives may conflict. Instead of adding raw rewards at incompatible scales, A2PO standardizes business and semantic advantages separately inside each rollout group and anchors the update to business advantage.
  • Self-Evolving RecSys: The inner agent uses researcher personas, an experiment journal, and coding/training tools to explore optimizers, architectures, and rewards. The outer state machine handles safety, resource constraints, and real A/B promotion, writing online findings back to memory for the next search round.
  • Unifying Ranking and Generation in Query Auto-Completion via Retrieval-Augmented Generation and Multi-Objective Alignment: Retrieve completion evidence from a searchable catalog. SFT and DPO align relevance, safety, engagement, catalog/context grounding and diversity.
  • LLaTTE: A large asynchronous upstream model consumes very long histories from multiple sources and caches a 2,048-dimensional user representation. The online model reads around 400 events and uses target-aware queries, multi-head latent attention (MLA), and pyramidal token reduction to aggregate candidate-relevant information.
  • RecGPT-V2: Fixed V1 chains duplicate work and generalize poorly. V2 uses hierarchical intent agents, compressed mixed behavior representations, dynamic meta-prompts, multi-reward constrained RL, and Agent-as-a-Judge.
  • CRSD: Build a domain reasoning LLM with continued pretraining, supervised fine-tuning and multidimensional preference optimization. Contrastively self-distill between plain and rationale-augmented inputs to the same lightweight student, eliminating the reasoning chain at serving.
  • PLUM: Next-SID generation is then fine-tuned for retrieval. At serving time, beam search generates candidate SIDs directly, without a dot-product retrieval index.
  • PinFM: Downstream candidate early fusion appends candidates to the user sequence. DCAT shares user history across candidates in one request, computing context self-attention once and cross-attending from each candidate to cached context KV.
  • FilterLLM: Per-user judgments do not scale to cold starts for a billion users. FilterLLM maps product text once to a user distribution, compresses the output space with a user vocabulary, and uses historical behavior to guide distribution learning.
  • SERAL: Serendipitous recommendations must be relevant and previously unseen. SERAL compresses very long history into multilevel cognition profiles, uses LLM and human CDI to break the logging feedback loop, aligns SerenGPT with IPO, then generates and caches candidates nearline.
  • LUM: In <condition, item>, the condition describes the scene, search term, or category of the item being predicted. It is not an action attached afterward to the previous item.
  • MIM: ID-only models struggle with fresh content. Reconstruct masked multimodal input, align content and collaborative interests with SFT, and fuse content/ID interests with CiUBM.
  • MSD: MKIM aligns the distilled student to CTR with LoRA and combines user, target, and history features through adapters. Popular items use precomputed embeddings, low-frequency items and cache misses are encoded online, and fusion uses only the history most relevant to the target.
  • PRECISE: PRECISE fuses ID and text at the item level and separates Universal Training (UT), which models interests across scenes, from Targeted Training (TT), which adapts to a sparse target scene or task. To keep strong text representations from starving the ID pathway of learning signal, text tokens are initially frozen while IDs train, then unfrozen for joint training.
  • LEADRE: It quantizes ads to S-IDs, trains ad semantics and commercial-value auxiliary tasks, then aligns with DPO. Online serving combines a long-term generation cache with real-time serving.
  • LEARN: LEARN's CEG freezes the LLM and mean-pools item text. PCH uses a trainable causal Transformer to learn user preferences from historical content embeddings, then projects them into the online recommendation space.
  • BAHE: During training, deeper trainable layers aggregate cached vectors and predict CTR with the candidate item. The upper sequence is shorter and the shallow stage can update less frequently and asynchronously.
  • LSVCR: LSVCR uses an LLM as an offline text encoder and a training-only supplemental recommender. It aligns its personalized preferences to a deployable dual-sequence sequential recommender (SR), which serves without a generative LLM.
  • NoteLLM: NoteLLM uses one compression prompt to distill content into a special token. GCL pulls co-read notes closer together; CSFT generates hashtags or categories from the same compressed state.
  • BEQUE: BEQUE uses rejection sampling and multi-instruction SFT, self-samples rewrites, ranks them on relevance, retrieval increment and hit rate in an offline search system, and learns partial preferences with rank-aware PRO. Serving refreshes rewrites offline and looks them up in a KV store rather than running a 7B LLM per request.
  • KAR: A hybrid-expert adapter gates shared and specialized experts into ID-compatible features; concatenate with original IDs for BCE prediction. Offline generation and caching avoid online autoregressive latency.
  • M6-Rec: Option tuning and per-layer option adapters update around 1% task parameters rather than fully fine-tuning M6. Late interaction, early exit, sharing and pruning yield 10M/2M M6-Edge models.

Generation, ranking, and cold start

  • UNIQUE: Shared representations train both code distributions and several ranking objectives. Inference nominates codes first, then ranks their mapped items.
  • ANGLE: Training joins generation, discrimination, and ranking objectives. Request-time beam constraints allow only valid identifiers.

Search, retrieval, and long-term value

  • MuSeR: The paper aligns BGE-encoded distilled ERNIE content summaries with ID embeddings, trains with next-K contrastive loss and interest orthogonality, and deploys asynchronous long-term caching, online short-term updates, and hierarchical retrieval. See Figures 2–3 and Table I.
  • MIMA: MIMA groups co-occurring items from one request as positives, generates interests with a causal decoder, and uses Hungarian matching to enforce one-to-one supervision. Routing estimates each interest's activation probability, making channel scores comparable.
  • VK Friend-GNN: Use three independent hashes into a shared table, concatenate/project, and retrieve only historical neighbors via timestamp-sorted CSR and binary search. GATv2 and separate query/candidate heads score links.
  • DCEO: The actor outputs context-dependent simplex weights and a user-level critic estimates relative intervention effects. Only the lightweight actor runs online.
  • TransRetrieval: Reduce source noise, target cost, and domain confusion. Aggregate source histories by token norms, compress target tokens to one, and encode domain as position.
  • Policy-Grounded Dynamic Facet Suggestions: An embedding SLM retrieves facets, and a generative SLM scores them pointwise with one token. Prefix caching and batching serve the latency target.
  • HILL: Attention-like soft assignment at the first layer leaves residuals for the next to quantize. Hierarchical retrieval narrows candidates at each level; upper nodes also produce a small set of high-quality test-time-training samples.
  • GenFacet: End-to-End Generative Faceted Search via Multi-Task Preference Alignment in E-Commerce: Unify e-commerce faceted search as contextual facet generation and intent-driven query rewriting. Teacher distillation and multi-task GRPO preference alignment optimize downstream search satisfaction directly, linking facet selection to retrieval.

Generative retrieval and end-to-end recommendation

  • UniSGR-V2 / TacGR: TacGR predicts continuous side features before each semantic code and crosses them inside the decoder. An auxiliary ranking tower prevents collapse of the regression target, and a one-way mask isolates prediction noise. Not reproduced for now due to cost; validation pending, no executable entry.
  • WeGR: CoCo-VAE learns continuous representations from query–item and item–item co-occurrence, refreshes underused codes, and uses residual fidelity and cross-example collision losses to improve codebook utilization. DHPO grows short SID prefixes toward complete SIDs, expanding reward matching and policy gradients in tandem; multi-positive SFT stabilizes training under sparse rewards.
  • Neither Black nor White: Balancing Semantic and Collaborative Signals with Graph-Informed Semantic IDs (GrIS): Nodes carry content semantics and edges collaboration. Empty graphs reduce to content quantization.
  • Generative End-to-end Ad Retrieval at Douyin: GEAR parameterizes BasisVQ/BasisRQ with an orthogonal basis so codebook updates share structure globally and rotate stably. The generator outputs hierarchical tokens; a context-conditioned reranker ranks items with collisions.
  • GRP v0.1: Encoder–decoder outputs multimodal SIDs. MHP ranks from detached encoder/SID states and then serves as frozen RL reward.
  • X-Rec: The computation pattern matters too. User history passes through the first L-1 Transformer layers.
  • TGR: Advancing Industrial Recommendation from Generative-Paradigm Ranking toward Unified Generation and Reasoning: TGR combines three industrial recommendation approaches. CCFormer supplies scalable ranking; BARGE and HiGR generate single items or whole-page lists with hierarchical semantic IDs; TGR-Reason injects reason tokens generated offline into online decoding, avoiding costly request-time rollouts.
  • TAGR: Static semantic IDs, single-timescale histories, and off-policy preference optimization can become stale. TAGR adapts at three levels: LSID periodically refreshes semantic and collaborative assignments without changing the token vocabulary; IAG blends short-, medium-, and long-term intent; IOPO runs on-policy GRPO intermittently and continues NTP between bursts, preventing online RL from overwriting the behavior distribution.
  • From a Static Multi-Level Small Semantic Codebook to a Dynamic Single-Level Large Semantic Codebook for Generative Recommendation: Compress a small multilevel semantic codebook into a large, dynamically updated single-level codebook, keeping separate collision codes. This reduces autoregressive steps and uses impression-weighted updates to counter codebook drift.
  • Sona: Sona represents users with a chronological event encoder, compresses history to limit long-sequence cost, autoregressively generates Semantic IDs, and ranks items using a Ranking Module conditioned on the same state. A high-capacity teacher supplies distillation supervision only during training.
  • MetaStrategy: A deterministic compiler executes the strategy on candidates so free text does not directly control live ranking. Production training also uses replay, a self-competition curriculum, and evaluator-routed reward-augmented OPD to distill a 4B teacher into a 0.8B student.
  • PushDualGen: Enabling LLMs to Generate Semantic IDs with Interpretable Copy for Industrial Push Recommendation: Generate a servable semantic ID first, then optional explanatory copy only when needed. Serving fuses SID and copy representations while preserving explanation consistency.
  • Gryphon-v2: Its generative head predicts Semantic IDs and an item-level head directly ranks generated candidates. A high-capacity offline teacher supervises current policy rollouts as well as logged impressions, so the ranker sees both serving and real impression distributions.
  • HRPO: Topic: generative recommendation post-training. Broadcasting item-level return to all SID tokens gives sparse, high-variance credit. HRPO smooths prefix utility within user clusters, decomposes residual token credit, accumulates credit-to-go, and updates conservatively with RRPO.
  • SnapLGR: Propagate content relevance over a co-engagement graph, then produce semantic IDs by hierarchical residual quantization. Training aligns SID tokens with actual content representations and learns to generate the next SID from the user's sequence.
  • OxygenREC-v2: Low-entropy positions get dense teacher supervision; high-entropy positions get forward KL to retain uncertainty. Inference never consumes future interactions.
  • UniR²: DQ-PCA uses two query types in one layer. Generation queries read the full user prefix and prior SIDs.
  • Dual-purpose Semantic IDs: Hierarchically quantized collaborative SIDs also reconstruct content via a Semantic Decoder (SiDec), so one token sequence supports generative retrieval, ranking and low-I/O online embedding reconstruction.
  • EGR: One encoder produces indexed item vectors and user query vectors. Train item representation learning and next-item prediction jointly throughout.
  • BARGE: HPR contrastively reranks cumulative semantic paths at every level. An orthogonal split produces two complementary decoding paths.
  • TSGR: TSGR uses residual quantization for a semantic prefix and assigns global and query-conditioned value codes in parallel to products sharing that prefix. A Value-aware Ranking Module (VRM), trained jointly with the generative backbone, does fine ranking without appending an independent pre-ranker to the generator.
  • RecGPT-V3: 1. Memory Hub compresses behavior into traceable memory units with behavior patterns, preference summaries, raw behavior indices, and timestamps. New behavior updates related units or creates new ones. 2. Hybrid-modal foundation model fuses content with CN-CLIP/Q-Former and quantizes products into SIDs with a two-level RQ-VAE.
  • NONTP: TCL predicts the EMA teacher's future state at \(j+k\) from the online backbone state at \(j\), with negatives from other users' sequences only. TDL averages hidden states before the target that belong to different domains and predicts the target through the same item head as NTP.
  • Guess Where You Go: Generative Next Point-of-Interest Recommendation in Amap: Encode the next point of interest as a semantic-ID sequence. Spatiotemporal features, curriculum pretraining and SFT, and EAKTO optimize longer-term feedback.
  • Prompt Generation: The same interpreter serves training and production; mergers fuse and compress inputs, with parameter-free embedding-space mean pooling a strong default. Event tracking captures online source features for offline replay.
  • CMSL: Each strand is encoded independently, then the current non-sequential context queries a pooling-by-multihead-attention (PMA) summary. The paper also describes a linear HSTU approximation with a degree-2 polynomial feature map for long industrial sequences.
  • Recommendation as Generation: Decoupled semantic IDs connect generative recommendation with interest-conditioned video generation; cross-domain rewards align candidate selection and content creation.
  • G2Rec: G2Rec builds a sparse item graph, learns soft membership in multiple interest prototypes through differentiable modularity, and alternates item and continuous interest-profile tokens in the decoder. An auxiliary profile loss complements next-item prediction.
  • SSRLive: Static SIDs encode creator content; dynamic SIDs capture current stream state. Fuse user–creator interactions for viewing, transaction and engagement tasks.
  • DeGRe: For each prefix, it distills lookahead values of all remaining candidates into dense soft labels. Online only one greedy generator is retained.
  • From Item-Only to Query-Item: Query-Conditioned Generative Search with QGS in Quark: Serialize query and item jointly for generative search. Linear HSTU encodes history at linear complexity and HFG-Attention fuses sparse crossed features.
  • GrowthGR: An RQ-VAE SID and decoder-only generator retrieve items. MoPO combines purchase, click, exposure, ranking, and uplift labels with clipped inverse propensity for rare valuable items.
  • CQ-SID: Constrain the first semantic-ID level by product category and encode within-category variation with residual codes. Expert-guided RL combines clicks, purchases, semantics and constraints for generative retrieval.
  • UniVA: 1. The SID has two semantic levels and one commercial token.
  • GloRank: A global SID action space turns local reranking into full-catalog generation. Listwise SFT precedes group-relative optimization of slate utility and tail coverage.
  • Beyond Static Collision Handling: Adaptive Semantic ID Learning for Multimodal Recommendation at Industrial Scale: Relax semantic-ID overlap constraints dynamically using collision load, semantic compatibility, and training stage instead of penalizing every collision equally.
  • GenRec: Item-at-a-time generation ignores complements and ordering on a page. Predict the next page, compress its inputs with an asymmetric merger, optimize group reward, and retain NLL to limit drift.
  • LWGR: Production precomputes the LLM world knowledge on a seconds-to-minutes nearline schedule and stores it in a user vector store. Requests need one lookup and one cross-attention layer.
  • UniRec: Bridging the Expressive Gap between Generative and Discriminative Recommendation via Chain-of-Attribute: Generate category, seller, and brand attribute chains before SIDs. Capacity-constrained quantization suppresses hot-token collapse, and task context stabilizes generation across scenarios.
  • MBGR: Ordinary next-item labels cover only the next logged business. A business-conditioned autoencoder retains SID semantics, shared MoE provides business representations, and LDR finds each business's nearest future interaction, masking missing labels.
  • RCLRec: Reverse Curriculum Learning for Modeling Sparse Conversions in Generative Recommendation: Select short conversion-related subsequences as reverse-chronological decoder prefixes. Joint generation supplies intermediate supervision; quality-aware loss suppresses uninformative curriculum samples.
  • GateSID: Adaptive Gating for Semantic-Collaborative Alignment in Cold-Start Recommendation: Shared attention fuses semantic and collaborative behaviors. A cold-start-aware gate adjusts contrastive alignment: new items lean on semantics while popular items retain collaborative signals.
  • AIGQ: An End-to-End Hybrid Generative Architecture for E-commerce Query Recommendation: The Direct path keeps latency low and the Reasoning path handles complex intent. IL-GRPO optimizes whole query sets with online CTR reward, using hybrid offline and online serving.
  • GLIDE: Recent-only histories retrieve habitual content. Generate item IDs conditioned on recent behavior and a long-term user embedding, balancing immediate intent and discovery.
  • Stop Treating Collisions Equally: Qualification-Aware Semantic ID Learning for Recommendation at Industrial Scale: Use business qualifications to set collision-pair margins rather than repelling every pair equally. Items with sparse feedback or cold starts receive more distinguishable semantic IDs.
  • GRC: GRC structures a draft SID, reflection on the first error and attribute consistency, and a corrected SID. It applies GRPO to whole trajectories and allocates a fixed correction budget to the most uncertain beams using reflection-token entropy.
  • GR4AD: GR4AD uses multigranularity UA-SID to encode ads, LazyAR to relax dependencies between autoregressive levels, VSL for business-value-weighted supervision, and RSPO to optimize ranked lists directly. Dynamic Beam Serving adjusts beam width with load.
  • SIGMA: Inference generates prefixes with beam search and performs ANN only within their subcatalogs. Adaptive probabilistic fusion (APF) trades off precision and diversity according to the spread of prefix confidences.
  • MFLI: Offline indexing delays fresh content, and online neighbor search becomes more expensive as candidate counts grow. MFLI learns the index jointly as model parameters, uses facets for semantics, popularity, and freshness, and allocates retrieval budgets per facet at request time.
  • Climber-Pilot: A Non-Myopic Generative Recommendation Model Towards Better Instruction-Following: Time-aware multi-item prediction distills long-horizon consumption into a one-step generator. Condition-guided sparse attention injects category and business constraints without separate filtering or multi-step inference.
  • S-GRec: Semantic and revenue objectives may conflict. Instead of adding raw rewards at incompatible scales, A2PO standardizes business and semantic advantages separately inside each rollout group and anchors the update to business advantage.
  • DiffuReason: Bridging Latent Reasoning and Generative Refinement for Sequential Recommendation: Thinking Tokens form continuous latent intent. Diffusion denoising refines conflicting interests step by step; GRPO aligns continuous generation with discrete ranking rewards end to end.
  • PIT: A Dynamic Personalized Item Tokenizer for End-to-End Generative Recommendation: Jointly train tokenizer and generative recommender, selecting the candidate SID with the smallest user-conditioned loss. A one-to-many beam index preserves alternatives under streaming distribution shifts.
  • DOS: Content-only codebooks may not match generative recommendation. Align behavior and item semantics, then separate primary and secondary signals with constrained orthogonal rotation before residual quantization.
  • OneMall: Unify previously separate product-card, short-video and livestream channels using common product SIDs, Query-Former compression, cross-attention over behaviors, a sparse MoE generator and RL alignment.
  • S2GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation: Insert interpretable thinking tokens before each SID level and align with codebook distributions contrastively. Co-occurrence and load balancing improve the coarse-to-fine hierarchy.
  • PROMISE: Process Reward Models Unlock Test-Time Scaling Laws in Generative Recommendations: It provides dense training feedback and prunes semantically drifting beams early at inference. Wider beams then give controllable test-time scaling.
  • HiGR: PCRQ-VAE produces hierarchical semantic IDs. The hierarchical decoder picks clusters before items, and ORPO aligns whole-page preferences.
  • DualGR: Use separate routers for short- and long-term interests, enforce valid semantic-ID prefixes, and apply an exposure-aware term to prevent concentration on head items.
  • IntSR: Use explicit queries and implicit session intent to generate POIs/items from one search-and-recommendation vocabulary. Temporal context represents intent drift.
  • OnePiece: Combine context tokens, block-level latent reasoning, and progressive multitask objectives in a cascade ranker. Retrieval, clicks, and value share representations but converge in stages.
  • OneRec-V2: Post-training uses actual watch duration and negative feedback. DARS corrects duration bias, and GBPO controls positive and negative gradients with dynamic denominators rather than PPO clipping.
  • GenRank: GenRank uses actions as supervision or features for the same item token and adds position and time biases. This reduces ranking cost without dropping the multi-action generation objective.
  • PinRec: Conventional next-item prediction tends to continue a user's most common behavior, and strict ordering does not fit feed sessions. PinRec injects the desired outcome into the output head and generates several ANN-compatible item vectors at once for the same future window.
  • COBRA: Generate a sparse code, then condition an end-to-end trainable dense vector on it. Use beam search to find codes and refine retrieval by dense neighbors.
  • OneRec: OneRec represents items with three-level RQ Semantic IDs, uses MoE to generate a full session of 5 target items, and trains a personalized reward model for watch and interaction metrics. Iterative Preference Alignment (IPA) chooses self-hard winners and losers from 128 beam-search sessions from the current model and uses just 1% of examples for iterative DPO.
  • SessionRec: Mean pooling beats other paper ISE variants; SSE can be Transformer or HSTU. The negative loss uses genuine next-session exposures without positive feedback.
  • LEADRE: It quantizes ads to S-IDs, trains ad semantics and commercial-value auxiliary tasks, then aligns with DPO. Online serving combines a long-term generation cache with real-time serving.
  • HSTU: HSTU unifies recommendation as generative sequence modeling. A transduction layer for sparse long histories replaces the standard Transformer's softmax attention and separate FFN.
  • TIGER: TIGER quantizes content vectors into hierarchical Semantic IDs. An encoder-decoder generates the next item as tokens, using an extra token to disambiguate items sharing a prefix.

Ranking networks and long sequences

  • RECAP CTR: Scale related estimators rather than only increasing the interaction capacity of one predictor. A shared recursive backbone generates routes, combined with independent-model distillation, training-trajectory EMA, and inference route averaging.
  • KuaFu: A cached item can be reused across users and need not be recomputed when a user adds an event. The original production cache and two-axis projection are shown in Figures 1 and 2.
  • ChronicleRec: Branches mask different amounts of recent behavior; alignment pretraining predicts recent intent from compressed history. Resulting Chronicle Tokens are candidate independent and can be cached and reused.
  • From Language to Behavior: Scaling Sequence Transformers for Industrial Recommendation Ranking with Rec-Native Designs: RoPE and relative time represent both order and elapsed time. A heavy user encoder runs once per request, then a light cross decoder scores each candidate.
  • OneModel: OneModel projects heterogeneous features to unified event tokens and models mixed long histories with a shared causal decoder. SAIM gates condition the FFN on the surface; the hierarchical representation combines global attentive pooling with recent local state, balancing transfer and specialization.
  • DrEM: Map multiple pxtr signals to logits, then filter noisy predictions while preserving preferences and invert pair-label flip risk to correct the training loss. Consistency-based fusion reduces ranking drift under input perturbations.
  • IntHQ: Task-Interactive Hierarchical Query on Dual-Stream Representations for Generative Recommendation: Separate behavior and context streams. The when, where, how, and via task queries retrieve interactively at multiple levels, bringing task signals into the encoder rather than confining them to task heads.
  • Netflix GenRec: Phase 1 continues training an open LLM on Netflix content and user data to create an internal foundation model. Phase 2 post-trains it on more frequently refreshed ranking data, labels, and reward signals.
  • TM20K: The student merges consecutive groups with sum divided by the square root of group size under a fixed token budget, and distills teacher attention and ranking distributions. This preserves long-sequence information at lower serving cost.
  • TransX: The behavior side builds global and local caches. Candidate processing cross-attends only to these caches instead of recomputing full self-attention on every request. Bounded component only; full experiments are not reproduced.
  • CCFormer: Industrial ranking needs field interactions and thousand-event histories. Project ID/content fields separately before cross-attention; widen receptive fields layerwise to compress old history while retaining detailed recent tokens.
  • HA-MoE: HA-MoE controls multi-gate expert mixing with content heterogeneity to separate common signals from domain expertise. LENS tracks expert usage and per-domain DL-AUC so aggregate scores do not hide local regressions.
  • ROCS: User-side features are shared across many candidates. Generalized Layer Masking isolates candidates, Deep Cross Attention defers candidate interaction, and IKBO broadcasts the request representation inside the GPU kernel.
  • ClockRoPE: ClockRoPE samples rotation frequencies from the Fourier transform of a target periodic kernel, strengthening the query-key dot product around the same hour or weekday. It can use half the channels while standard RoPE uses the other half.
  • TWICE: Learn current-status labels on the click clock separately from the conversion-clock delay CDF, then use exposure-window weights to correct labels that take a long time to mature.
  • CORE: Valid tokens use clipped GRPO ratios plus reference KL. For low latency, PostCoT logits distill into two conditional BERT heads with BCE and temperature-\(T\) distillation.
  • Mosaic: As the fleet grows, new embeddings can duplicate existing information. Multi-Task Relation Mining (MRM) forms a Cartesian product of binary labels for highly correlated tasks to learn joint behavior explicitly.
  • SpecFormer: Soften the singular-value spectrum and use spectral residual positional encoding to mitigate embedding and attention collapse in recommendation Transformers.
  • YouTube Freshness: Compare recency, IPS debiasing, a removable bias tower, and SNGP uncertainty exploration. Distinguish training debiasing from serving exploration.
  • WHALE: A Wukong branch recursively models higher-order feature interactions and an HSTU branch models behavior sequences. Bidirectional cross-attention exchanges states each layer rather than scaling separate towers independently.
  • Long-History User Transformers: The online component reads that cache and encodes only recent events. Training starts with auxiliary next-event and feedback tasks, then fine-tunes for the target ad surface.
  • TMallGS: TMallGS learns field-specific Q/K/V and controls interactions with distribution-calibrated saliency and noise-adaptive gates. It injects content through FiLM, corrects final scores with context bias, and assigns progressive supervision by intermediate-layer errors.
  • DANet: IPN models general interest. DPN FFT-filters a 400-day discount series; DCM-User extracts preferences from similar purchases, DCM-Context gates promotion regimes, and positive purchases supervise discount regression.
  • MESH: Separate user, item, and context subtowers and amplify each via residual crossing before global interaction. RGBC gates user–item affinity with context; serving can parallelize subtowers and cache item features asynchronously.
  • SAM: Pointwise cross-attention finds history to suppress; long purchases predict replenishment. ASGU strongly suppresses just after purchase and restores attention near the cycle through a log mask.
  • SlimPer: Each layer revisits raw sparse and sequence tokens, selects relevant evidence, explicitly matches it with knowledge templates, and refines the knowledge base. User tokens are encoded once per request and shared across all candidates.
  • ZoRRO: Without model training, combine time decay, semantic similarity, and category relations among historical articles for high-throughput news-candidate ranking.
  • UniFormer: Separate item-independent and item-dependent features and tasks into semantic tokens. A unified interaction module scales sequences, features, and tasks together.
  • JourneyFormer: Encoding Airbnb Guest Journey with Sequence Modeling: Encode Airbnb's long- and short-term guest journey with event type, time, and listing representations. Caching and sparse computation meet serving latency.
  • OneRank: Put structured features and task tokens in one sequence instead of separating encoder and task predictors. Task-private channels and isolated gradients limit interference.
  • Memento: Data Memento retrieves forgotten but currently relevant training examples for a second replay step. Maximal marginal relevance balances relevance with redundancy; Ember Affine scales and shifts, and Quadratic adds multiplicative interactions.
  • UxSID: IAIC applies cross-attention from learned anchors to the behavior sequence, then independent sigmoid PFFNs for each anchor, a residual, and LayerNorm. An orthogonality term discourages interest collapse.
  • Effective Knowledge Transfer for Multi-Task Recommendation Models: Transfer CTR representations into multiple conversion towers according to CVR-task similarity. A hard-example-aware loss limits negative transfer while keeping multi-task serving cost manageable.
  • Semantic-Native Long Sequence Modeling: Full attention is also quadratic in thousand-event histories. The paper quantizes with RQ-VAE into coarse-to-fine SIDs, truncates to a bounded bigram vocabulary, folds every \(k\) neighboring events into a parameter-free super-token, and adds shared global queries to preserve overall interests at longer histories.
  • RankUp: Towards High-rank Representations for Large Scale Advertising Recommender Systems: Counter representation collapse with randomized grouping, multiple embeddings, global tokens, cross-pretrained embedding injection and decoupled task tokens so added capacity increases representational rank.
  • UniScale: Synergistic Entire Space Data and Model Scaling for Search Ranking: ES³ expands training beyond exposed items; HHSFT hierarchically combines heterogeneous examples. Scale data and model capacity jointly with low-cost deployment.
  • HAP: Route candidates to lightweight or strong pre-ranking by difficulty. Cross-branch harmonization keeps ranking scales comparable across compute budgets.
  • SORT: A Systematically Optimized Ranking Transformer for Industrial-scale Recommenders: Redesign feature tokenization, attention and FFN, with hardware-oriented long-sequence compression so one Transformer replaces a traditional DLRM.
  • OneRanker: Use a fake item token to unify generation, click-value estimation and ad ranking. Distribution consistency aligns generated probability with the value head.
  • Beyond the Flat Sequence: Hierarchical and Preference-Aware Generative Recommendations: Learn hierarchical behavior with session-based masked item modeling. At fine-tuning, preference-guided sparse attention retains only events most relevant to current intent, reducing noise and computation in flattened long sequences.
  • HiSAC: Full attention over long behavior histories is expensive, while fixed compression can discard tail interests. A multimodal RQ-VAE provides hierarchical codes; levelwise voting builds user-specific interest agents, and the query soft-routes to a small relevant subset.
  • ULTRA-HSTU: At 16k behaviors and greater depth, HSTU attention and memory bottleneck scaling. Use local tokens and sparse landmarks, widen windows by layer with LBSL, and combine a mixture of transducers with attention truncation, mixed precision and dynamic topology.
  • MixFormer: MixFormer embeds nonsequential fields and behavior into one Transformer token space, sharing parameters and enabling high-order interaction at every layer so dense capacity and sequence length scale together. It also decouples user and candidate-item computation to reuse one user representation across candidates in a request.
  • An Industrial-Scale Sequential Recommender for LinkedIn Feed Ranking: Unify Feed candidates, interaction history, and context into a very long sequence, with HSTU-style target attention, evaluation without sampled negatives, and incremental serving optimizations.
  • Compress, Cross and Scale: Multi-Level Compression Cross Networks for Efficient Scaling in Recommender Systems: Each layer compresses high-dimensional features, applies dynamic explicit feature crosses in the compact space, and projects them back into the backbone. MC-MLCC uses parallel subspaces to increase expressivity while controlling parameter and FLOP growth.
  • Compute Only Once: UG-Separation for Efficient Large Recommendation Models: An attention/TokenMixer mask blocks backward information flow from candidates to user tokens, allowing the user representation to be reused across candidates in one request. Information Compensation restores user–candidate interactions suppressed by masking.
  • Kunlun: GDPA/PFFN model sequences, HSP summarizes sequence seeds, and Global Interaction exchanges sequence and non-sequence information. CompSkip and expert parallelism maintain deep-training stability and throughput.
  • MDL: Shallow gates underuse shared parameters for multiple scenarios and tasks. Represent features, scenarios, and tasks as tokens and combine feature self-attention, domain-feature attention, and fused heads.
  • MSN: Scale parameters without proportionally increasing ranking computation. Store capacity in two-axis product-key memory, retrieve a few slots per query, and gate with a dense backbone.
  • TokenMixer-Large: After reverting the mix it applies token-wise SwiGLU. Interval residuals and intermediate supervision stabilize scaling of deep fine-ranking models.
  • Zenith: Scaling up Ranking Models for Billion-scale Livestreaming Recommendation: Token Fusion compresses heterogeneous sparse features into fewer high-dimensional Prime Tokens. RSA/TMHSA maintain token diversity, while TSwiGLU/TSMoE independently enhance tokens for better compute/size scaling.
  • Unleashing the Potential of Sparse Attention on Long-term Behaviors for CTR Prediction: TimeChunking adapts to a user's inter-event intervals. EvoAttention combines global chunk, interest-transition and local-window attention; RelTemporal adds relative interval, hour and weekend biases.
  • HyFormer: Each layer cross-attends the current sequence K/V, then mixes decoded queries with non-sequence tokens using RankMixer-style boosting. Boosted queries decode again in the next layer, making co-evolution bidirectional and iterative.
  • OneTrans: The pyramid stack retains only recent S queries at each layer. Production reuses S-side computation through cross-candidate and cross-request KV caching.
  • ARGUS: Direct next-item prediction entangles the feedback type and item. Predict the feedback token first, then condition next-item prediction on it, allowing multiple feedback types to share a sequential backbone.
  • RankMixer: A ReLU router and sparsity regularization allow different numbers of active experts per token. Dense training with sparse inference avoids under-training experts.
  • Click A, Buy B: Last click assumes the same item is clicked and bought. Separate click-A-buy-A and click-A-buy-B tasks, with taxonomy-based collaborative weights across items.
  • MGOE: MMoE shares at example level but does not express stable inter-task structure. Build the Macro Task Merging Graph, propagate via graph bottom and experts, and predict with task towers.
  • TransAct V2: Next Action Loss makes the sequence representation predict the next positive interaction. The paper also includes serving optimizations such as SKUT and request deduplication; local reproduction focuses on the model pipeline.
  • SORT-Gen: At inference, it builds objective queues such as CLICK, PAY, and GMV and reduces the search from all candidates to queue heads. Mask-Driven Fast Generation records selected items and queue state in tensor masks, evaluating several positions and queues in one batched call.
  • LONGER: Later layers self-attend only over compressed queries. Training uses mixed precision and activation recomputation; serving uses a per-user KV cache.
  • FuXi-α: Scaling a recommendation Transformer can bottleneck on shallow feature interactions. Use adaptive temporal, semantic, and popularity channels with staged FFN fusion.
  • AdaF²M²: A single all-feature forward can over-rely on strong features, and different user or item states need different features. AdaF²M² uses repeated forwards under distinct masks to learn complementary representations and fuses them dynamically with a state-aware adapter.
  • TWIN-V2: Offline hierarchical clustering compresses millions of lifetime behaviors into virtual items with cluster-size information. Online, GSU retrieves relevant clusters and ESU fine-ranks with cluster-aware target attention.
  • CWM: Use counterfactual watch-time estimation to remove video-duration bias, distinguishing observed watch time from potential utility under a common duration intervention.
  • PLE: MMoE shares all experts across tasks, which can entangle common and conflicting signals. PLE splits shared and task-specific groups; layered CGC gates progressively extract both kinds of information and reduce seesaw effects.
  • DCN-V2: Model bounded-degree feature interactions explicitly. Low-rank factors and a mixture of cross experts reduce ranking parameter and compute costs.
  • SIM: Use the candidate item as a query. The General Search Unit retrieves a relevant subsequence from lifelong history; the Exact Search Unit attends to that subsequence for the candidate.
  • BST: Use self-attention over behaviors and the candidate together, rather than fixed pooling or single-target attention.
  • DIEN: Unlike DIN's direct interest activation, DIEN extracts stepwise interest with auxiliary supervision, then lets the target item control state evolution.
  • SASRec: SASRec uses causal self-attention to select behaviors relevant to the next item, giving a more flexible receptive field across sparse and dense data. It became a common sequential recommendation baseline.
  • MMoE: Fully independent models cannot exploit commonality. MMoE shares experts while learning an independent softmax gate for each task, so tasks combine experts with different weights.
  • ESMM: ESMM jointly learns CTR and CVR over all impressions and uses the probability chain rule to derive CTCVR. Dense CTR supervision stabilizes representations in the CVR tower.
  • DIN: Dice adapts its activation boundary to the feature distribution through batch statistics. Local click scoring uses BCE on positives and negatives.
  • DeepFM: Wide & Deep needs manual wide features; DeepFM shares field embeddings across both paths end to end. The model adds first- and second-order FM terms to an MLP output, reducing manual crosses and duplicate parameters.
  • YouTube DNN: Aggregate watch-history embeddings and other heterogeneous context with a deep nonlinear user tower. Predict the next video as large-scale multiclass classification and retrieve by nearest neighbor at serving.
  • Wide & Deep: A pure DNN generalizes but may recommend irrelevant combinations. The paper trains both paths jointly and shares the final logit.

Cold start and semantic–behavior alignment

  • CohortMix-TS: Use metadata to build a fixed-strength mixture Beta prior for a new user. At each round, Thompson sampling selects a previously unexposed slate; feedback updates each user's posterior independently.
  • LLM Thompson Priors: Topic: LLM semantic priors for cold-start bandits. New comments have no interaction feedback but their text contains gender/content preference signals. Convert LLM semantic judgments into Beta pseudocounts and maintain Thompson posteriors by gender and age segment.
  • PinEqualizer: A separate exploration corpus uses Thompson posteriors, real-time impression UCB or Neural Linear UCB. Less observed exposure increases the bonus; search relevance suppresses off-topic exploration.
  • Pin-SCALE: Keep residual SID hierarchy via cascading pooling, encode engagement into tokenization, and use query–item and item–item contrastive views to align IDs with engagement prediction for dense retrieval.
  • Proximity Features: Dense locations split by IP hash; sparse locations progressively merge at coarser geographic precision. Cohort behavior is aggregated daily by stable key without persistent personal identifiers online.
  • Taiji: Use reverse-engineered preference reasoning and rejection sampling to activate recommendation chain-of-thought. POPO then balances LLM semantic reward and collaborative-ID reward dynamically, avoiding the nonconvex Pareto-frontier blind spots of fixed weights.
  • LLM Retrieval: A retrieval-domain LLM extracts hierarchical category, brand, product, and context attributes from titles, descriptions, and other creative fields. Category-based retrieval is followed by deeper attribute reranking and ad-to-ad semantic graph construction, then graph expansion from user seed ads.
  • SID-Coord: Coordinating Semantic IDs for ID-based Ranking in Short-Video Search: Rather than treating semantics as another dense feature, build multilevel discrete SIDs. Hierarchical fusion, target-aware HID–SID gating, and historical-interest alignment coordinate head-item memorization and long-tail generalization per example.
  • PinCLIP: General VLM training does not match recommendation, and fresh content lacks behavior. PinCLIP combines multigranular image/text signals with Pin–Board graph-neighbor alignment.
  • Podcast MTL: The paper jointly trains ads and organic promotion, sharing user, content, and context representations to transfer signal from rich tasks to cold-start tasks. Imbalance handling and gradient balancing limit negative transfer.
  • SaviorRec: Zero-initialized Modal-Behavior Alignment (MBA) codebooks learn residuals relative to current behavior. Bidirectional target/history attention feeds the CTR predictor.
  • PRECISE: PRECISE fuses ID and text at the item level and separates Universal Training (UT), which models interests across scenes, from Targeted Training (TT), which adapts to a sparse target scene or task. To keep strong text representations from starving the ID pathway of learning signal, text tokens are initially frozen while IDs train, then unfrozen for joint training.

Semantic identifiers and incremental updates

Long-sequence compression and retrieval

Foundation model capabilities in recommendation

  • PailitaoMMSearch2.0 Technical Report: Scaling Generative Retrieval to Unified End-to-End Multimodal Search: The data process combines cold start, quality calibration and feedback from real traffic. Training proceeds through product-knowledge continued pretraining (CPT), supervised fine-tuning (SFT) with mixed reasoning budgets, and intent-balanced reinforcement learning (RL). Not reproduced due to cost (review complete); research material only, no executable entry.
  • PinGER: Token-Efficient Generative Retrieval with Semantic IDs: Beam diversification penalizes repeated subtrees in proportion to shared prefix depth rather than changing only sampling temperature. Post-training derives high/low-utility preference pairs from ranker logs for the same exposed slate, applies DPO at every SID prefix, and retains a next-identifier anchor. Bounded component only; full experiments are not reproduced.

Training objectives and decision optimization

Sampling, distillation, and reinforcement learning

  • CASI: Broadcasting set reward can still provide a valid score-function estimator of the original set objective. CASI's internal shaping does not additionally claim to preserve that unbiased gradient. Bounded component only; full experiments are not reproduced.
  • LLM Thompson Priors: Topic: LLM semantic priors for cold-start bandits. New comments have no interaction feedback but their text contains gender/content preference signals. Convert LLM semantic judgments into Beta pseudocounts and maintain Thompson posteriors by gender and age segment.
  • KGD: KGD uses BMTP to select collaboratively or semantically related future items. During transfer, freeze the refreshable knowledge encoder, read it with cross-attention, and write task geometry with ACR orthogonal to the anchor.
  • HRPO: Topic: generative recommendation post-training. Broadcasting item-level return to all SID tokens gives sparse, high-variance credit. HRPO smooths prefix utility within user clusters, decomposes residual token credit, accumulates credit-to-go, and updates conservatively with RRPO.
  • RecHarness: A Bandit-Routed Agentic Harness for Self-Evolving Recommender Systems: Define architectural changes as a finite set of arms. A bandit allocates trials from historical gains and uncertainty, then writes winning feedback into the next candidate round.
  • From Understanding to Action: Feedback-Grounded Policy Discovery for Generative Recommendation: Discover policies from user feedback and distill them into a lightweight ranker using relations in two spaces, avoiding LLM calls at serving.
  • ASARL: CriticAgent repeatedly verifies labels and identifies coverage gaps, and GenAgent adds samples in the long tail. The pipeline then runs Social Context Training, Preference-Guided Optimization on interaction signals, and distillation into a small serving model.
  • Reward Guided Decoding: Keep generator probabilities as a prior and exponentially tilt them with an independent reward at decoding time. A closed-form policy increases expected reward while KL temperature controls deviation, without retraining the generator.
  • RAMP: Privacy choices, permissions and platform limits can hide personalized fields for some traffic. Instead of zero-filling, RAMP maintains personalized and non-personalized paths, selects output with an availability mask, and aligns predictions to distill knowledge from the feature-rich path into the restricted one.
  • UAME: Analytical pair probabilities combine pxtr objectives, giving higher weight to more conflicted pairs. Serving uses only the mean, without extending the online path.
  • Downstream Rewards: Pinterest proposes and screens candidate downstream rewards offline for their ability to predict future engagement. Reward heads that pass are added as model-agnostic auxiliary objectives across Homefeed, Search, Related Pins, and Notifications without replacing the existing ranker.
  • Cluster GOOBS: The paper clusters items by LLM/multimodal embeddings, uses the GOOBS update engine to write into real-time OOB pool segments by cluster, and has the sample engine draw hard negatives from the positive item's cluster with high probability. The online model is unchanged and sampling remains O(1).
  • SCALR: SCALR decouples transfer from the model: estimate an item translation distribution from overlapping users, sample target-domain item \(i_k\sim\hat P(\cdot\mid u,j)\) for a source event, then add \((u,i_k,w_k)\) as ordinary synthetic training data. Sampling retains distribution/catalog diversity unlike repeated head-item top-k generation.
  • Rec-Distill: Raw teacher logits first receive student-side sampling correction to align probabilities. Batch training converges before consuming streaming teacher signals.
  • Pinterest Complementary LLM Predictor: First, SFT predicts one next advertiser. Then GRPO generates 20 advertisers and 5 interests, rewarding the hit position and enforcing output lengths.
  • Stochastic Primal-Dual Decoding: Without retraining, add auxiliary objective contributions to relevance logits at each decoding position. Mirror descent updates Lagrange multipliers from remaining constraint slack.
  • MDCNS: Disagreement favors informative hard negatives. The Teacher's distribution over the entire pool is distilled back to Self, using more information than a single selected negative.
  • UniVA: 1. The SID has two semantic levels and one commercial token.
  • AgenticRecTune: Actor proposes, Critic checks risk, Insight summarizes findings, Skill persists knowledge, and Online returns experiment-platform feedback across retrieval, fusion and reranking.
  • LWGR: Production precomputes the LLM world knowledge on a seconds-to-minutes nearline schedule and stores it in a user vector store. Requests need one lookup and one cross-attention layer.
  • Cross-domain KD: Low-traffic products cannot train a large specialized teacher. Apply a data-rich YouTube multitask teacher directly to YouTube Music examples, and distill both serving and auxiliary logits into a small model without requiring source-domain data or matching interface features.
  • SaFRO: Satisfaction-Aware Fusion via Dual-Relative Policy Optimization for Short-Video Search: Satisfaction reward models immediate engagement and long-term retention. Dual-Relative PO normalizes advantages within groups and across tasks; relation gating fuses search objectives.
  • GR4AD: GR4AD uses multigranularity UA-SID to encode ads, LazyAR to relax dependencies between autoregressive levels, VSL for business-value-weighted supervision, and RSPO to optimize ranked lists directly. Dynamic Beam Serving adjusts beam width with load.
  • S-GRec: Semantic and revenue objectives may conflict. Instead of adding raw rewards at incompatible scales, A2PO standardizes business and semantic advantages separately inside each rollout group and anchors the update to business advantage.
  • DRL-PUT: Treat a utility-weight combination as the action. Condition on request/user state and learn from logged propensities and aggregate rewards instead of manually tuning relevance, CTR, CVR, and revenue weights.
  • MPFormer: Objective tokens drive one sequential retriever across behaviors. Allocate capacity and retrieval quota dynamically by task difficulty.
  • OneRec-V2: Post-training uses actual watch duration and negative feedback. DARS corrects duration bias, and GBPO controls positive and negative gradients with dynamic denominators rather than PPO clipping.
  • SERAL: Serendipitous recommendations must be relevant and previously unseen. SERAL compresses very long history into multilevel cognition profiles, uses LLM and human CDI to break the logging feedback loop, aligns SerenGPT with IPO, then generates and caches candidates nearline.
  • LEADRE: It quantizes ads to S-IDs, trains ad semantics and commercial-value auxiliary tasks, then aligns with DPO. Online serving combines a long-term generation cache with real-time serving.
  • BEQUE: BEQUE uses rejection sampling and multi-instruction SFT, self-samples rewrites, ranks them on relevance, retrieval increment and hit rate in an offline search system, and learns partial preferences with rank-aware PRO. Serving refreshes rewrites offline and looks them up in a KV store rather than running a 7B LLM per request.

Causal inference and long-term value

Multitask learning and joint labels

  • Learned Cross-Task Relationships in Multi-Task Models: Marginal heads discard joint information. Paired joint labels or missing unconditional auxiliary heads supply shared gradients and downstream policy features; the paper tests these effects separately. Bounded component only; full experiments are not reproduced.

Ad attribution and weak supervision

  • Statistical Learning from Attribution Sets: Use position priors and binomial tail probabilities to estimate risk and squared-prior weighting for ERM, rather than random or max-prior pseudo-labeling. Implement printed v1 literally and preserve its failed finite-sample identity check. v1 Section 3 Theorem 1 and Section 4.1 Eqs. 6–7. Bounded component only; full experiments are not reproduced.

Baselines and incremental debiasing

Advertising and commercial decisions

  • Hierarchical GTV Estimation: Bridging the Gap between Container-Level Ranking and Item-Level Conversion: The Hybrid Set-Aware Encoder (HSAE) models interactions among the top-\(K\) items before adding user, store, context and query features. Proxy-Label Learning (PLL) separately predicts whether the store converts, the probability of each item's purchase conditional on conversion, and quantity. Residual Correction and Ensemble Fusion (RCEF) accounts for value outside the top-\(K\) and mixes the corrected estimate with generic store-level regression. A learned gate sets that mix from context per example, not by routing on observed conversion labels. Bounded component only; full experiments are not reproduced.
  • Reinforcement Learning for Controller Parameters in the Auto-Bidding System at Pinterest: Separate Q-heads learn relative budget-utilization and CPC outcomes. The system combines them according to the remaining campaign lifetime. Bounded component only; full experiments are not reproduced.

Foundation model capabilities in recommendation

Multi-stage ranking and blending

Real-time sequences and listwise reranking

  • SequenceO1: SequenceO1 compresses distant history to a fixed number of prototypes and retains a recent-behavior branch. STCA combines the two in parallel, controlling cost while retaining long-term interests.
  • POEM: Convert relative ranks from several ranking objectives on the current request into partial-order signals. Construct training sequences that better reflect current interests than chronology alone.
  • PIANO: The Query-Driven Interest Refiner extracts query-relevant interests from history. The Information Aggregation Node summarizes the entire candidate list and conditions item-level CTR/CVR predictions.

Coordinated retrieval, pre-ranking, and ranking

  • OneTrans-V2: History cannot read stage tokens. Each stage token sees only events before the request and not other stage tokens, reusing encoding without future leakage.
  • UniRec: UniRec jointly trains two-stage fusion with partially shared embeddings and separate DCN/Mixer and AutoInt-Lite towers. Downstream preferences supervise upstream training only; they are not online inputs to pre-ranking. v3 retains multi-objective conflicts through CPPA, prunes pairs by confidence with CAPP, and uses within-group AGRR KL to counter distribution drift from attributes such as video length. Bounded component only; full experiments are not reproduced.
  • SPEAR: Generic rewrites can win via path-score shortcuts. Dual embeddings and gradient isolation protect retrieval semantics; multiply rewrite confidence by item relevance, then use a dynamic selector for request-specific weight, scale and bias.
  • STEPS: STEPS jointly decides whether to push and when to wake next. A planning agent uses gated ordinal regression for intervals, an execution agent decides whether to send based on trajectory return, and a lightweight filter controls computation and blocks abnormal plans.
  • ROCS: User-side features are shared across many candidates. Generalized Layer Masking isolates candidates, Deep Cross Attention defers candidate interaction, and IKBO broadcasts the request representation inside the GPU kernel.
  • OneShot: OneShot learns a hierarchical index with ranking: discrete paths on forward pass, straight-through gradients on backward pass, global path balance against hot buckets, and neural scoring beyond dot product after probing paths. Training joins ranking, indexing and balancing objectives.
  • SlimPer: Each layer revisits raw sparse and sequence tokens, selects relevant evidence, explicitly matches it with knowledge templates, and refines the knowledge base. User tokens are encoded once per request and shared across all candidates.
  • CS3: Independent towers enable ANN but limit interactions. CS3 adds cycle revision and knowledge coordination while preserving offline and online deployability.
  • COBRA: Generate a sparse code, then condition an end-to-end trainable dense vector on it. Use beam search to find codes and refine retrieval by dense neighbors.

Reranking, blending, and multi-objective page decisions

  • Once Generated, Ranked: End-to-End Generative Slate Recommendation with Unified Semantic-Collaborative IDs: TUSID combines semantic and collaborative codes. GL2P plans the whole slate, then SPA aligns to list-level feedback with conservative policy optimization.
  • DEGR: Add squared cosine penalties within a cohort to prevent repeated similar exposures. Exploration feedback adjusts ORPO reward weights, and inference selects greedily with diversity constraints.
  • Twitch Multi-Objective Ranking: Topic: Multi-objective livestream ranking. Viewing, interaction, follows and payment arrive at different delays and vary across lifecycle segments. The paper combines immediate and delayed models, segment-specific objective weights and MMoE with shared experts and objective-specific gates.
  • DIRECTOR: Generate a dynamic-index distribution for each slot and constrain all slots jointly with Sinkhorn optimal transport. Global matching rather than independent argmax reduces conflicts and produces the entire slate in parallel.
  • PSG: Generate a half-length pair sequence and expand losslessly. Ordered pair scores and duplicate constraints ensure usable slates.
  • LLM-Based Re-Ranking for Real Estate Search: Pass conversational requirements, structured listing attributes, text descriptions, and candidate-set statistics to an LLM reranker. Cross-candidate comparison complements local relevance from vector retrieval.
  • GenPage: Towards End-to-End Generative Homepage Construction at Netflix: Generate the entire homepage with one model rather than cascading retrieval and ranking. After WBC/SFT, RL post-training uses long-term production rewards while handling business rules and cold start.
  • Pinterest Complementary LLM Predictor: First, SFT predicts one next advertiser. Then GRPO generates 20 advertisers and 5 interests, rewarding the hit position and enforcing output lengths.
  • DeGRe: For each prefix, it distills lookahead values of all remaining candidates into dense soft labels. Online only one greedy generator is retained.
  • Memento: Data Memento retrieves forgotten but currently relevant training examples for a second replay step. Maximal marginal relevance balances relevance with redundancy; Ember Affine scales and shifts, and Quadratic adds multiplicative interactions.
  • A Production-Ready RL Framework for Personalized Utility Tuning with Pareto Sweeping in Pinterest Recommender Systems: A dual-head Q-network estimates Repin and P2P utility. Pareto sweeping filters nondominated policies offline, so serving swaps governed utility weights without replacing the model.
  • GenRec: Item-at-a-time generation ignores complements and ordering on a page. Predict the next page, compress its inputs with an asymmetric merger, optimize group reward, and retain NLL to limit drift.
  • Dual-Rerank: An autoregressive teacher learns item-order dependencies; a non-autoregressive student predicts ranks in parallel. Utility distillation optimizes quality and latency.
  • Constraint-Aware Generative Re-ranking for Multi-Objective Optimization in Advertising Feeds: One network generates the sequence and estimates reward. Encode slot and ad-load rules in autoregressive decoding, and prune infeasible branches early rather than repairing output afterward.
  • Rethinking Multi-objective Ranking Ensemble in Recommender System: From Score Fusion to Rank Consistency: Rewrite AUC as a differentiable rank-sum objective to align tasks in ranking space. Then learn dependencies among objectives such as purchases and comments with self-attention to produce one ranking.
  • DRL-PUT: Treat a utility-weight combination as the action. Condition on request/user state and learn from logged propensities and aggregate rewards instead of manually tuning relevance, CTR, CVR, and revenue weights.
  • SORT-Gen: At inference, it builds objective queues such as CLICK, PAY, and GMV and reduces the search from all candidates to queue heads. Mask-Driven Fast Generation records selected items and queue state in tensor masks, evaluating several positions and queues in one batched call.
  • CDM: A controllable MMR teacher generates lists balancing relevance and diversity. Distill contextual marginal gains into a student suitable for low-latency serving.

Foundation model capabilities in recommendation

Content understanding, moderation, and risk control

Content understanding and semantic representation

  • CAMIE: Co-Engagement-Aware Multimodal Item Embeddings for Snap Dynamic Product Ads Retrieval: Mine jointly engaged items from user journeys, then use symmetric in-batch InfoNCE to shift content similarity toward behavioral co-conversion. One checkpoint supports both multimodal and text-only retrieval.
  • Multimedia Asset Personalization via Multimodal Embeddings at Netflix: Use frozen CLIP/MediaFM representations as shared online features in an asset two-tower model and enhance the search canvas with query similarity. Offline IPS and a winner proxy screen checkpoints.
  • DME: After contrastive pretraining, organize evidence with Evidence-Grounded Typed Latent Reasoning and preserve opposite-side details with Cross-Conditional Reconstruction. Both generation heads are training-only.
  • GALA: Generative Aligned Learning for Adaptive Multimodal Representation in the Taobao Shangou Recommender System: Pretrain multimodal representations on query–image–text triples, align them with GRPO and behavioral rewards, then fuse multimodal and ID representations with an adaptive gate.
  • FLUID: The two are separate late-fusion tokens. Candidate IDs are removed through three stages: add slice, remove ID, add room.
  • MM-LLM: Item tokens and user-history token profiles join ID and visual features in the production ranker. LLM inference stays off the online critical path.
  • TagLLM: A Fine-Grained Tag Generation Approach for Note Recommendation: Multimodal chain-of-thought extracts fine-grained tags by interest dimension. Two-stage knowledge distillation transfers tagging to a smaller model for explainability and serving efficiency.
  • PinCLIP: General VLM training does not match recommendation, and fresh content lacks behavior. PinCLIP combines multigranular image/text signals with Pin–Board graph-neighbor alignment.
  • IDProxy: Multimodal semantics differ from long-learned collaborative item-ID space. Contrastive coarse alignment precedes multilayer hidden-state proxies, injected into the ranker through a residual gate.
  • RQ-GMM: Residual Quantized Gaussian Mixture Model for Multimodal Semantic Discretization in CTR Prediction: Quantize image/text embeddings in residual stages with Gaussian-mixture posterior soft assignments to improve codebook utilization relative to hard clustering. Train semantic discretization separately from CTR preference modeling.
  • SARM: LLM-Augmented Semantic Anchor for End-to-End Live-Streaming Ranking: An offline MLLM generates livestream semantic anchors, which a lightweight SAE injects into end-to-end ranking. Asymmetric deployment updates only the ranker online to avoid MLLM latency.
  • MIM: ID-only models struggle with fresh content. Reconstruct masked multimodal input, align content and collaborative interests with SFT, and fuse content/ID interests with CiUBM.
  • PRECISE: PRECISE fuses ID and text at the item level and separates Universal Training (UT), which models interests across scenes, from Targeted Training (TT), which adapts to a sparse target scene or task. To keep strong text representations from starving the ID pathway of learning signal, text tokens are initially frozen while IDs train, then unfrozen for joint training.

Relevance auditing and data quality

  • LLM-based User-Product Relevance Evaluation for DPA: Clicks are sparse and biased, so offline hit rate alone does not represent user intent. This method compresses user history by semantic ID and event type, trains a student judge using automatically discovered and manually business-aligned relevance criteria, and aggregates scores by user and traffic group. Not reproduced for now due to cost; validation pending, no executable entry.
  • ASARL: CriticAgent repeatedly verifies labels and identifies coverage gaps, and GenAgent adds samples in the long tail. The pipeline then runs Social Context Training, Preference-Guided Optimization on interaction signals, and distillation into a small serving model.
  • CORE: Valid tokens use clipped GRPO ratios plus reference KL. For low latency, PostCoT logits distill into two conditional BERT heads with BCE and temperature-\(T\) distillation.
  • MESH: Separate user, item, and context subtowers and amplify each via residual crossing before global interaction. RGBC gates user–item affinity with context; serving can parallelize subtowers and cache item features asynchronously.
  • Cluster GOOBS: The paper clusters items by LLM/multimodal embeddings, uses the GOOBS update engine to write into real-time OOB pool segments by cluster, and has the sample engine draw hard negatives from the positive item's cluster with high probability. The online model is unchanged and sampling remains O(1).

Privacy, policy constraints, and risk control

Moderation and risk-control coverage

Current implementations primarily cover relevance audits, training data quality, privacy-constrained features, and policy constraints. Content safety, fraud and abuse, and ad compliance remain explicit gaps. Relevance models do not fill them.

  • SIRF: Continued pretraining writes complex trigger/exemption rules into model weights. A single verdict and adjustable threshold support decisions on a seconds-scale serving path.
  • RAMP: Privacy choices, permissions and platform limits can hide personalized fields for some traffic. Instead of zero-filling, RAMP maintains personalized and non-personalized paths, selects output with an availability mask, and aligns predictions to distill knowledge from the feature-rich path into the restricted one.
  • UAME: Analytical pair probabilities combine pxtr objectives, giving higher weight to more conflicted pairs. Serving uses only the mean, without extending the online path.
  • Causal Retrieval: Always triggering the shopping candidate generator wastes resources and can over-distribute to users without shopping intent. The paper learns multi-event outcomes and doubly robust uplift from randomized logs, selects a threshold via offline linear replay, and runs online in parallel with remote retrieval.
  • Proximity Features: Dense locations split by IP hash; sparse locations progressively merge at coarser geographic precision. Cohort behavior is aggregated daily by stable key without persistent personal identifiers online.

Serving and research infrastructure

Feature evaluation and retirement

  • FORTE: This is not one-shot feature selection. Adding individual effects does not estimate a joint removal effect: individually redundant features may jointly carry important information, and different heads can disagree on the direction.

Autonomous research and feedback loops

  • EvoSkillRec: Split a recommender into executable skills with typed inputs and outputs. The controller composes within constraints; only validated innovations advance to the reusable library.
  • EvoPilot: The protocol fixes the shared base, data version, evaluator, effective output depth, and sole permitted treatment difference. It reads actual run state from artifacts, fails closed on mismatches, and still requires human review before accepting a conclusion.
  • AutoLR: Automating the Path from Research to Launch Review in Industrial Recommender Systems: AutoLR links literature review, proposal debate, implementation experiments, offline evaluation, A/B, and Launch Review in a durable state machine. An LLM reasons over proposals and generates code; a deterministic controller retains ownership of budgets, metric extraction, guardrails, and transitions so noisy wins do not compound through KEEP decisions.
  • CORAL: Validate constraints, execute, and write outcomes into the next round's short memory. This is an iterative loop rather than one-shot tuning.
  • RecEvolve: Propose variants from experiment knowledge and select champions on isolated validation. Version state for rollback and reject reward hacking such as extreme temperature, carrying forward auditable knowledge.

Serving / efficiency

  • Optimizing Effective Training Time for Large-Scale Recommendation Systems: Decompose wall time into useful training on new data versus lifecycle costs: initialization, compilation, checkpointing, publishing, and recovery. Attribute lost time by owner so it can be addressed.
  • ROCS: User-side features are shared across many candidates. Generalized Layer Masking isolates candidates, Deep Cross Attention defers candidate interaction, and IKBO broadcasts the request representation inside the GPU kernel.
  • Memory Layer: Memory Layer makes the cache a sparse model table: the tower writes during training with learning rate one, and downstream and serving read the same representation. A cache miss uses a zero embedding plus always-on attributes such as creator/topic.
  • Prompt Generation: The same interpreter serves training and production; mergers fuse and compress inputs, with parameter-free embedding-space mean pooling a strong default. Event tracking captures online source features for offline replay.
  • NOVA: Summarize past changes, diagnostics and outcomes in an architecture gradient. Gate changes through L1 structure semantics, L2 execution, L3 offline effectiveness and L4 online impact; escalate high-risk tasks to humans.
  • RankGraph-2: Online multi-layer GNNs are costly and can overemphasize popular edges. Correct edge popularity, precompute multihop PPR, and compress serving with a jointly learned residual cluster index.
  • EvoRec: EvoRec's Research and Code Agents iterate models. Its Skill Evolver distills reusable methods from persistent memory so both the model and its optimization methods evolve.
  • Pinterest Complementary LLM Predictor: First, SFT predicts one next advertiser. Then GRPO generates 20 advertisers and 5 interests, rewarding the hit position and enforcing output lengths.
  • UniVA: 1. The SID has two semantic levels and one commercial token.
  • CS3: Independent towers enable ANN but limit interactions. CS3 adds cycle revision and knowledge coordination while preserving offline and online deployability.
  • LWGR: Production precomputes the LLM world knowledge on a seconds-to-minutes nearline schedule and stores it in a user vector store. Requests need one lookup and one cross-attention layer.
  • SIGMA: Inference generates prefixes with beam search and performs ANN only within their subcatalogs. Adaptive probabilistic fusion (APF) trades off precision and diversity according to the spread of prefix confidences.
  • Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators: STATIC compresses a trie offline into CSR transitions, with dense lookup available for shallow layers. Candidate gathering happens on the GPU or TPU at each step, avoiding CPU pointer traversal and device round trips. Bounded component only; full experiments are not reproduced.
  • S-GRec: Semantic and revenue objectives may conflict. Instead of adding raw rewards at incompatible scales, A2PO standardizes business and semantic advantages separately inside each rollout group and anchors the update to business advantage.
  • OneTrans: The pyramid stack retains only recent S queries at each layer. Production reuses S-side computation through cross-candidate and cross-request KV caching.
  • OneRec-V2: Post-training uses actual watch duration and negative feedback. DARS corrects duration bias, and GBPO controls positive and negative gradients with dynamic denominators rather than PPO clipping.
  • RADAR: Run a full ranker asynchronously after the current request. Cache its high-value items to supplement real-time retrieval on the next request.
  • LONGER: Later layers self-attend only over compressed queries. Training uses mixed precision and activation recomputation; serving uses a per-user KV cache.
  • MSD: MKIM aligns the distilled student to CTR with LoRA and combines user, target, and history features through adapters. Popular items use precomputed embeddings, low-frequency items and cache misses are encoded online, and fusion uses only the history most relevant to the target.
  • BAHE: During training, deeper trainable layers aggregate cached vectors and predict CTR with the candidate item. The upper sequence is shorter and the shallow stage can update less frequently and asynchronously.
  • BEQUE: BEQUE uses rejection sampling and multi-instruction SFT, self-samples rewrites, ranks them on relevance, retrieval increment and hit rate in an offline search system, and learns partial preferences with rank-aware PRO. Serving refreshes rewrites offline and looks them up in a KV store rather than running a 7B LLM per request.
  • M6-Rec: Option tuning and per-layer option adapters update around 1% task parameters rather than fully fine-tuning M6. Late interaction, early exit, sharing and pruning yield 10M/2M M6-Edge models.

Advertising and commercial decisions

  • BAFF: Training B directly on pooled logs creates treatment interference. BAFF checks where A's ad falls in B's ranking and how its bid differs, retaining A records only when they have approximately no effect on B's decision.
  • PinDCO: PAM penalizes creatives that occupy too many feed pixels by relative aspect ratio. Lightweight filtering, caching and dynamic batching control serving cost, while ε-greedy traffic collects less biased new-creative feedback.
  • CADET: Context-Conditioned Ads CTR Prediction With a Decoder-Only Transformer: A decoder-only Transformer encodes cacheable histories, then injects post-scoring context after candidate scoring to align training with serving and unify ad CTR prediction.
  • ML-DCN: Masked Low-Rank Deep Crossing Network Towards Scalable Ads Click-through Rate Prediction at Pinterest: Replace DCNv2's full-dimensional crosses with low-rank crosses of adjustable internal width. A learned mask selects cross channels, increasing capacity at the same FLOPs.

Retrieval, triggering, and multi-channel routing

Multitask learning and multi-objective optimization

  • FLVM: Measurement confounders enter a restricted baseline; full user/video information enters the latent encoder. Each feedback head reads only relevant factors and adds its logit adjustment to a detached baseline.
  • SMES: Towards Scalable Multi-Task Recommendation via Expert Sparsity: Split each task's active experts into a stable subset shared across tasks and an adaptive private subset. Cross-task global load balancing constrains routing, activation count, and expert skew.
  • Towards End-to-End Alignment of User Satisfaction via Questionnaire in Video Recommendation: Keep sparse survey satisfaction supervision in its own LoRA path so abundant behavioral labels do not drown it out. Online DPO uses a survey expert as a dynamic reference and updates satisfaction preferences while minimizing disruption to the main ranking backbone.

Advertising and commercial decisions

Spatial externalities and allocation

  • Online Advertising with Spatial Interactions: The local component implements the paper's grouped pseudo-disk LP relaxation and two-stage randomized rounding, checking group exclusivity and that no selected disk's center is covered by another selected disk. It does not reproduce full matching, live auctions, or the end-to-end factor-18 approximation. Bounded component only; full experiments are not reproduced.

Other industrial components

Foundation model capabilities in recommendation

  • OneLatent: Latent Reasoning for Efficient Foundation Recommendation Models: Use exploration, evaluation and attribution views to generate adaptive teacher supervision, learn three latent tokens, and serve three-level SIDs without long text. Alignment reconstructs single-view reasons, then multiple views, then predicts next SID from request context. Not reproduced due to cost (review complete); research material only, no executable entry.

Retrieval and matching