type
Post
status
Published
date
Aug 2, 2026 03:23
slug
rec-weekly-en-2026-W31
summary
This week's recommendation systems research runs along three interwoven technical threads: generative recommendation has hit a new peak in industrial deployment density, with multiple companies disclosing online gains; LLM recommendation is shifting from explicit reasoning to latent reasoning, with inference cost emerging as the primary constraint on scale; and industrial infrastructure papers are converging on training-serving inconsistency, inference compute reuse, and cold start. The common thread: recommendation systems are moving from a "model capability race" to a "systems engineering race." Thread 1: Generative recommendation enters a multi-objective, controllable industrialization phase. Kuaishou's Multi-Decoder OneRec uses a multi-decoder architecture to decouple shared representations from objective-specific specialization — online app time +0.37%, cold start +2.09%. JD's OxygenREC-v2 internalizes discriminative signals into a 3B-parameter MoE generative backbone, lifting GMV 2.8%-6.8%. The competitive focus has shifted from "can it retrieve" to "can it steer direction and tune objectives." Thread 2: LLM recommendation reasoning is moving from explicit CoT to latent reasoning. Kuaishou's WhisperRec compresses teacher CoT into latent tokens — SID@64 +17.44%, online inference throughput up over 10x. LaRec samples reasoning starting points from personalized Gaussian mixture distributions, exploring multi-path latent reasoning. The quality ceiling of explicit reasoning still stands, but inference cost determines who survives online. Thread 3: Engineering depth in industrial recommendation systems. Meta's ROCS extends request-side compute sharing from feature interactions to sequence models — retrieval model QPS up 3x. Memory Layer uses a key-value cache co-trained with the model to unify training-serving representations — NE gap reduced 86%. Structural alignment between training and serving is becoming a bigger optimization lever than model architecture.
tags
Recommendation Systems
Weekly
Papers
category
Rec Tech Report
icon
📚
password
priority
1
Weekly Overview
This week's recommendation systems research runs along three interwoven technical threads: generative recommendation has hit a new peak in industrial deployment density, with multiple companies disclosing online gains; LLM recommendation is shifting from explicit reasoning to latent reasoning, with inference cost emerging as the primary constraint on scale; and industrial infrastructure papers are converging on training-serving inconsistency, inference compute reuse, and cold start. The common thread: recommendation systems are moving from a "model capability race" to a "systems engineering race."
Thread 1: Generative recommendation enters a multi-objective, controllable industrialization phase. Kuaishou's Multi-Decoder OneRec uses a multi-decoder architecture to decouple shared representations from objective-specific specialization — online app time +0.37%, cold start +2.09%. JD's OxygenREC-v2 internalizes discriminative signals into a 3B-parameter MoE generative backbone, lifting GMV 2.8%-6.8%. The competitive focus has shifted from "can it retrieve" to "can it steer direction and tune objectives."
Thread 2: LLM recommendation reasoning is moving from explicit CoT to latent reasoning. Kuaishou's WhisperRec compresses teacher CoT into latent tokens — SID@64 +17.44%, online inference throughput up over 10x. LaRec samples reasoning starting points from personalized Gaussian mixture distributions, exploring multi-path latent reasoning. The quality ceiling of explicit reasoning still stands, but inference cost determines who survives online.
Thread 3: Engineering depth in industrial recommendation systems. Meta's ROCS extends request-side compute sharing from feature interactions to sequence models — retrieval model QPS up 3x. Memory Layer uses a key-value cache co-trained with the model to unify training-serving representations — NE gap reduced 86%. Structural alignment between training and serving is becoming a bigger optimization lever than model architecture.
Generative Recommendation & Semantic Indexing
The Design Space for Semantic IDs Reopens
LGRID (Kuaishou) — Existing SID generation frameworks mostly follow a "single-representation quantization" paradigm, where entangled heterogeneous attributes cause quantization information loss and severe collisions. LGRID proposes an Encode→Disentangle→Align→Quantize pipeline: first a joint LLM encoder preserves cross-attribute geo-semantic dependencies, then structured disentanglement blocks route hidden states to attribute-aligned slots, and finally dual-stream residual quantization discretizes geographic and semantic factors separately. The result: SID positions carry explicit attribute and semantic meaning — interpretability comes from design, not post-hoc attribution. Relative AUC +5.44% on Kuaishou and Foursquare; overall SID collision rate drops from 97.0% to 39.9%. This continues the signal-organization focus of DualGR for generative retrieval, but pushes disentanglement granularity down to the attribute level.
TopoGR (Academic) — Identifies a structural problem in generative recommendation: the tokenizer learns a structured code space with semantic neighborhood relations, but the generator consumes SIDs as independent categorical symbols. Item relatedness is thereby reduced to exact SID overlap. TopoGR uses bit-decomposable Binary SIDs to expose explicit Hamming geometry — maintaining Hamming proximity at the input layer, injecting topology-aware supervision via Hamming soft targets during training, and applying Hamming-consistent reranking at inference. The deeper implication: SIDs aren't just IDs — they're geometric structures the generator can exploit. Compared to TIGER's fixed-length discrete codes, TopoGR lets SID geometry participate in generation decisions.
VaLiDRec (Academic) — Challenges fixed-length SIDs from another angle: items with different semantic complexity need different identifier lengths. VaLiDRec builds variable-length SIDs directly from information tokens in the LLM's native vocabulary, via three steps: token importance estimation, semantic-quality-aware pruning, and collision-aware refinement. It reformulates recommendation as token-set prediction, eliminating autoregressive decoding and beam search — inference is 87.49x faster than LC-Rec. When SID length adapts to item semantic complexity rather than uniform truncation, information retention improves, and zero-shot cold start benefits as a result.
The common thread across all three: SID evolves from "a usable encoding" to "an interpretable, exploitable, scalable structure." The design quality of the encoding space is becoming a new ceiling for generative recommendation performance.
Multi-Objective & Controllability Become the Deployment Mainline
Multi-Decoder OneRec (Kuaishou) — Industrial recommendation systems allocate explicit quotas to separate retrieval paths per objective; more paths mean fragmented modeling, training, and serving. Multi-Decoder OneRec keeps a shared user context module + a universal decoder for unification, adding one parameter-efficient LoRA expert per objective. Training uses gradient routing to isolate updates: exposure-sample NTP updates the shared backbone, objective-filtered NTP updates event experts, and KL-regularized policy optimization updates the duration expert. At inference, explicit path quota allocation fixes the budget, and Multi-Decoder Constrained Beam Search reduces cross-path overlap. Recall@512 +1.69%-5.62% under a 512-item retrieval budget; online app time +0.37%, cold start +2.09%. They also open-sourced the Kwai26 benchmark — 1.31 billion item-level records, 31.85 million Item-IDs, 25.03 million valid SIDs. This is one of the few works that publishes both a complete multi-objective generative retrieval solution and its benchmark.
RGD (Kuaishou) — Beam search in generative recommendation is dominated by generation likelihood; high-value but low-probability candidates can be pruned early. RGD formulates value-guided decoding as a KL-regularized reward maximization problem, derives a closed-form decoding distribution, and injects reward signals at each step to reshape the search trajectory — no generator retraining required. This effectively adds a test-time controller to generative recommendation: when business preferences change, swap the reward model, not the generator. RGD is deployed on Kuaishou's platform with consistent gains. This "test-time control" approach complements xGR's inference-time item filtering: one handles efficiency, the other handles objective alignment.
OxygenREC-v2 (JD) — A 3B-parameter, 1B-activation MoE generative recommendation model. The core is IDGR (Internalizing Discrimination into Generative Recommendation): pretraining conditions generation on behavior instructions (click, add-to-cart, purchase); post-training uses future interaction behavior as privileged knowledge in an entropy-aware trajectory optimization self-distillation framework, enabling reward-model-free policy optimization. A single backbone is maintained throughout — no additional discriminative objectives. Online UCTCVR +1.6%-4.4%, GMV +2.8%-6.8%. Compared to GRLM's structured Term ID route, OxygenREC-v2 takes a different path — instead of changing the ID format, it fuses behavioral signals into the generation process itself.
The Boundary Between Generative Recommendation and Other Tasks Is Dissolving
UniR² (Kuaishou) — Unifies generative retrieval and multi-objective ranking into a single decoder-only sequence containing user context, SID trajectories, and item features. The generation trajectory serves as a representation bridge between retrieval and ranking. Dual-Query Prefix-Causal Attention provides different information visibility for the two tasks — shared attention weights but independent optimization boundaries; ranking-side LoRA doesn't disturb the generative backbone. Online A/B tests show consistent long-term positive results. This responds to Gryphon's joint training problem for generative recommendation and item-level scoring, but solves it with a single backbone rather than an encoder-decoder structure.
FDP (Kuaishou) — Proposes the Understanding-Action Gap: LLM reasoning can be linguistically plausible without leading to effective recommendation decisions. FDP distinguishes intent knowledge (the user's current need) from policy knowledge (recommendation direction and rejection boundaries under that need), using a feedback-driven agent framework that first induces task-oriented intent, then discovers policy via incremental utility. Two latent tokens are distilled into a lightweight Semantic-ID generator via dual-space relation distillation, enabling LLM-free online inference. Online Revenue +4.506%, ADVV +4.621%. The deeper value: repositioning the LLM from "recommender" to "knowledge distillation source" — continuing OneLoc's geo-aware SID route, but raising the structural granularity of knowledge transfer from "distill to model" to "distill to token."
The other side of boundary dissolution is the redefinition of generation units. Kuaishou's PSG elevates the generation atom from individual items to ordered item pairs. Given n candidates, PSG operates on a pair vocabulary of size n(n-1), generating only L/2 tokens per request. The paper provides three theoretical guarantees: bijective with item-space generation and inducing an equivalent distribution family (no expressivity loss); theoretical speedup of 2-4x, 1.83x in practice; worst-case suboptimality bound of O((L/2)²ε̄), roughly 4x better than item-space generation. Online user dwell time +0.178%, serving over 400 million DAU. The jump from items to pairs as generation atoms is like swapping in coarser-grained Lego bricks for generative reranking.
Memory & Attribute Modeling for Generative Recommendation
LoopMemGR (Academic) — Points out a flaw in the "history-as-context" paradigm of generative recommendation: the system remembers what the user did, but not what it recommended or what it learned from feedback. LoopMemGR maintains a recommendation experience log, extracting evidence from three views — recent view (short-term interaction dynamics), frequency view (repeated recommendation patterns), and global view (cross-user transferable regularities) — compressed into a fixed number of experience tokens. The system's own recommendation-feedback loop supplies the other half of collaborative signals. Validated on Taobao industrial datasets; one of the few works that explicitly models "system-side memory."
SPARC (Academic) — In industrial settings, every behavior carries heterogeneous attributes (category, brand, price, behavior type, timestamp). Expanding them all directly blows up input length; compressing into a single representation discards contextually relevant information too early. SPARC first models sequence dependencies per field type, then routes raw/contextual/identity representations to multiple slots, and finally compresses via lightweight cross-item interaction into a single token. The principle: "contextualize first, then compress." Outperforms both traditional and generative baselines on Taobao and Amazon datasets. The choice of when to compress is itself an information retention strategy.
CogRec (Academic) — Anchors the reasoning process on SID topology. CogRec overlays intra-layer semantic graphs and item-level neighborhoods on the vertical SID hierarchy, expressing recommendation reasoning through three SID Routing operations (Match/LateralJump/Explore): exact matching for fast semantic localization, lateral jumps and exploration for slower structural navigation. Experiments show structure-anchored reasoning is most effective when prefix matching is insufficient but SID-space transitions are learnable. Reasoning and generation happen in the same topological space, rather than writing natural-language reasoning before generation — a directional correction to the explicit CoT paradigm.
LLM & Agent Recommendation
From Explicit to Latent Reasoning
WhisperRec (Kuaishou) — Explicit CoT reasoning produces verbose rationales, and fixed templates struggle to model diverse, dynamic user interests. WhisperRec uses Multi-View Adaptive CoT (MV-ACoT) to build high-quality supervision from complementary perspectives — lightweight analysis for simple samples, targeted multi-factor reasoning for hard ones. A three-stage latent reasoning alignment then compresses teacher CoT into learnable latent tokens. On Kuaishou industrial data: SID@64 +17.44% over explicit CoT, +9.33% over the No-Think variant, online inference throughput up over 10x. This is a further compression of SCoTER's structured CoT transfer route — not distilling to text, but distilling to latent space.
LaRec (Academic) — Identifies two challenges in latent reasoning: lack of fine-grained supervision (only final label feedback) and single-path reasoning (deterministic latent reasoning can't explore diverse interests). LaRec's latent pretraining provides rich supervision via step-level alignment and process-direction alignment; personalized RL fine-tuning builds a Gaussian mixture distribution per user, sampling different reasoning starting points during training to guide the LLM through diverse reasoning paths in latent space. Substantially outperforms baselines across multiple datasets with efficiency comparable to existing methods.
HiLaR (Academic) — A hierarchical latent reasoning framework that builds time-guided hierarchical user preference representations, aligned with multiple latent reasoning states of the LLM — reasoning proceeds from broad preference organization to fine-grained current intent. Layer-aware process rewards (per-layer marginal target likelihood gain) optimize the reasoning trajectory. Outperforms multiple strong baselines on four Amazon datasets. The shared judgment across all three papers: explicit reasoning's information content is the ceiling, but inference cost determines deployment feasibility. Latent reasoning is the path that compresses information while preserving decision-relevant content.
Agent Systems Engineering: Error Recovery Matters More Than Reasoning Ability
Melo (NetEase Cloud Music) — An LLM-driven music recommendation agent that orchestrates heterogeneous tools via a deterministic five-node state graph, rather than fine-tuning a controller. The core judgment: at industrial scale, the bottleneck isn't how smart the brain is, but how the system discovers and recovers from the brain's mistakes. Two production failure modes drove the design — entity hallucination (the agent promises explanations for items not in the catalog) and long-tail degeneration (over-constrained requests collapse into generic popular fallbacks). The countermeasures: inference-time entity grounding (production search index as a verification primitive, gating entity decisions before they propagate) and reflective retry (verbalizing failure reasons back into the next planning step, relaxing or modifying constraints rather than blind fallback). Online A/B: main playlist retention +2pp, play time +1 minute, entity misidentification -7.8pp; reflective retry triggers in 5.8% of sessions with 59% process-level recovery. Cross-referenced against the Autonomous Information Seeking agent taxonomy, Melo is a canonical industrial implementation of "agent-as-recommender" — trading degrees of freedom for controllability via state graphs.
ASARL (Tencent QQ Search) — Social search queries and titles use informal community language, creating a contextual gap with general LLM semantic understanding. ASARL uses three collaborating agents for data curation: ReasonAgent generates explainable relevance labels from social attributes, CriticAgent verifies logical consistency, and GenAgent synthesizes long-tail query-title pairs. Three-stage training (Social Context Training, Preference-Guided Optimization, Social Distillation) distills improvements into a compact deployed model. Significant gains in both offline relevance and online user engagement.
Safety & Trustworthiness of LLM Recommendation
Ranked by Position (Academic) — Reframes position bias in LLM listwise rerankers as an attack surface. An attacker who only reorders candidate positions — without modifying content, labels, or model parameters — can push label-0 targets into the top-k. promo@k quantifies the vulnerability: with R=50 ranking budget, promo@5 reaches 0.57. More critically, ordinary permutation stability predicts vulnerability without requiring actual attacks. Mitigations include bidirectional T5 encoders, permutation-consistency regularization, and architectural invariance. When LLM rerankers are used in industrial recommendation, candidate serialization order is a security-relevant variable — and it's not in most system designers' threat models.
CALMRec (Academic) — LLM recommenders often conflate persistent preferences, transient intents, and exposure-induced behaviors into a single profile. CALMRec uses a frozen multimodal LLM to convert item content and feedback into evidence-anchored semantic atoms, maintaining separate short-term, long-term, and exposure memories. Propensity weighting reduces policy-induced exposure bias; a conservative offline critic reranks candidates for delayed gratification under behavioral support constraints. Across three simulated environments (e-commerce, news, short video), discounted long-term value improves 6.1%-7.6% over the strongest alternative. Exposes LLM recommendation's fragility in feedback loops — repeated exposure is mistaken for preference, and immediate clicks dominate delayed gratification.
Industrial Architecture & Full-Funnel Optimization
Inference Compute Sharing: Request-Level Reuse Becomes a New Paradigm
ROCS (Meta) — Recommendation inference has a unique property: one user request scores many candidates, but request-side features are shared across candidates. ROCS delays request-candidate interaction as much as possible, isolating candidate-dependent representations so the model computes once per request rather than once per candidate. Three components: Generalized Layer Masking (GLM) enforces candidate isolation in feature interaction architectures, Deep Cross Attention (DCA) extends to sequence architectures, and In-Kernel Broadcast Optimization (IKBO) accelerates GPU execution. At production scale: retrieval model QPS up to 3x with no quality loss; short-video ranking model shows 0.5% LogLoss improvement with 50% QPS gain. Deployed across ads and organic content, retrieval and ranking stages. Complements Disaggregated Multi-Tower's data-center locality approach: DMT optimizes communication topology, ROCS optimizes compute reuse structure.
OneShot (Meta) — The long-standing misalignment between learned indexing and ranking objectives: ranking optimizes prediction-behavior alignment, indexing optimizes structural grouping of item representations. OneShot learns the index end-to-end within the model, natively aligning index learning with ranking objectives, and uses neural scoring to break the dot-product bottleneck. Fully deployed on Instagram short-video recommendation: recall +20% at equivalent operational displacement, 10x efficiency at equivalent recall. Unlike xGR's serving-system perspective, OneShot eliminates the "index as a posteriori structure" assumption from the modeling side.
Training-Serving Consistency: From Engineering Patches to Architectural Built-ins
Memory Layer (Meta) — Early ranking stages precompute item embeddings and cache them inside the model. This cache exists only at serving time, not in the training loop, so training and serving use different representations. Memory Layer is a key-value embedding cache co-trained with the model: the item tower writes embeddings during training, the model reads them at serving — a natural single source of truth. Always-on embeddings cover uncached items. Deployed on Instagram Reels: prediction coverage from 96% to 100%, embedding freshness from O(5min) to O(20s), training-serving NE gap reduced 86%, fresh-content recall up over 2x, cold-start engagement +5-6%. Training and publishing compute costs down 30%, serving compute flat. This is one of the few works that directly eliminates the source of training-serving representation divergence.
CCFormer (Tencent) — Balances compute cost against feature interaction granularity under strict latency constraints. Feature-domain separated cross-attention + long-sequence subspace token mixing; a hierarchical sequence compression strategy uses progressively expanding receptive fields to reduce information loss. Launched on Tencent Video and Ads: CTR +3.57%, ad revenue +1.71%, training 2.21x faster (vs. HSTU). Compared to HSTU's pure sequence-scaling route, CCFormer demonstrates that cross-domain feature interaction and sequence modeling can be unified in a single Transformer backbone.
Systematic User Representation
Mosaic (Meta) — User representation is a high-leverage modeling problem that propagates across retrieval, ranking, and integrity tasks. Mosaic learns user embeddings with an architecturally diverse set of expert models: memory-driven, dense-heavy, sequential, and CoTrain. MRM (Multi-task Relations Mining) and CRL (Cosine Redundancy Loss) maximize each new expert's marginal information contribution. CoEval and User Tower Zero-Out provide a log-free embedding evaluation framework; a hybrid CPU/GPU online-offline serving stack lets each expert choose a strategy matching its freshness, latency, and compute requirements. The "user model" shifts from a single model to an "expert fleet," where new models are valued by marginal information contribution rather than standalone performance.
Engineering Answers for Cold Start, Debiasing & Exploration
PinEqualizer (Pinterest) — A full-funnel solution for content cold start, covering retrieval, pre-ranking, ranking, and reranking, applicable to both search and recommendation. The core is reducing bias toward incumbent content without the short-term cost of large-scale explicit exploration. A companion scalable measurement framework supports rapid short-term experiments and validates long-term impact. After two years of deployment: significant improvement in fresh-content exploration, user engagement, and content ecosystem health. Contrasts with Warmer for Less's lightweight cold-start strategy (only 5% added parameters) — PinEqualizer is a system-level approach.
ConAlign (Kuaishou) — A streaming debiasing framework using a small amount of unbiased random traffic. The core is discrete-gated conditional alignment: selectively transferring knowledge from the biased tower to the unbiased tower rather than global correction. This avoids the "sacrifice factual recommendation performance" cost common to debiasing, and supports real-time streaming adaptation. First large-scale industrial deployment of streaming debiasing with small unbiased random traffic; significant gains in long-term user engagement and interest diversity online, with negligible latency overhead.
Breaking the Loop (YouTube Music) — Off-policy online A/B tests of six feedback-loop intervention strategies across serving, training, architecture, and exploration layers. Three key findings: serving-layer interventions get canceled out by the continuously trained learning loop; architectural debiasing (e.g., SNGP) substantially improves new releases and diversity but carries hidden ensembling costs; uncertainty-driven exploration produces the largest new-release gains but sacrifices engagement or diversity. "Which layer you intervene in" matters more than "which intervention you use" — serving-layer heuristics can be swallowed by the learning loop. Complements PinEqualizer's full-funnel approach: YouTube's conclusion is about layer trade-offs, Pinterest's practice is about layer synergy.
SpecFormer (Alibaba) — Attribution analysis reveals the spectral cause of Transformer failures in recommendation: heterogeneous and long-tailed data causes severe spectral collapse, with a few dominant singular values dominating and forward-backward propagation forming a vicious cycle. SpecFormer's three components each address one aspect: learnable spectral softening dynamically smooths the singular value distribution, spectral-softened attention models feature interactions in a more uniform spectral space, and spectral residual positional encoding provides spectral inductive bias via Taylor expansion. Deployed in commercial systems; stacking layers consistently improves attention's effective rank and recommendation performance.
Deployment Experience & Infrastructure
Entity Resolution (Walmart) — Three lessons from a self-service entity resolution system across six benchmarks (864 to 5M records): no single matching algorithm works for all datasets — train multiple algorithm families and let an automatic bake-off pick the winner; precision and recall need separate fixes — precision via hard-rule vetoes, recall via more diverse candidate retrieval; a single false-positive link can silently merge unrelated entities, and transitive closure lets one bad link chain together hundreds of records — every cross-entity merge must be actively re-verified. A rare first-hand record from a production system of "which paths don't work."
Embedding Items at Scale (Yandex) — First industrial-scale comparison of GNN-pretrained item embeddings vs. end-to-end trained embeddings, across two mature production systems: Yandex Market and Yandex Music. Conclusion: pretraining helps when data is limited; large models pretrained on large data show no substantive gains. They release the Yandex Lavka low-resource dataset for reproduction. This gives industrial-grade evidence for the "is pretraining worth it" debate — it depends on data scale.
KuaiLive-M3 (Kuaishou) — A multimodal, multi-domain, multi-feedback live-streaming recommendation dataset. 21.9K users, 35M live interactions, 111M short-video interactions, ~88M timestamped segment-level multimodal embeddings, 25.4K questionnaire feedback records. Supports three benchmark tasks: cross-domain recommendation, live-stream highlight prediction, and questionnaire-enhanced recommendation. Questionnaire feedback bridges implicit behavior and explicit preference — a signal type commonly missing from existing live-streaming datasets.
Advertising & Bidding Optimization
User Foundation Models Enter the Open Web
User Foundation Model for the Open Web (Teads) — The structural difference between open-web RTB and closed ecosystems (e-commerce, social): fragmented user identity that doesn't persist across sessions, sparse historical data, often short sessions. This work pretrains a Transformer encoder with masked language modeling + sequence-level contrastive learning, then fine-tunes for click prediction. The same encoder representation delivers RIG +1.197% on the ad bid-rate model and RIG +1.354% on the CTR ranker; a 7-day online A/B test shows CTR +2.13%, eCPC -1.13%. The pretraining pipeline introduces LLM-as-optimizer: LLM search over a curated catalog of auditable code-level edits (lifters). "LLM as optimizer rather than reasoner" is the most noteworthy methodological contribution — it uses LLMs for offline, auditable process optimization, sidestepping online inference cost.
Dual-Clock Modeling for Delayed Feedback
TWICE (Kuaishou) — Long-horizon conversion prediction under delayed feedback creates a dual-clock, dual-window problem. The short base observation window releases recent clicks promptly on the click clock, but conversions keep arriving on the conversion clock through a longer target window. TWICE decomposes long-horizon CVR into target-window conversion probability + grouped delay CDF. The click clock records training target-window CVR heads; newly arriving conversions train the delay model. Fixed click-time pCVR quality serves as queue exposure volume, resolving the mixture of traffic and conversion propensity across click queues. A single learned CDF produces monotonic predictions across all request time horizons — no history lookup or convolution at serving. Online A/B: expected revenue +2.486%, revenue +1.858%, conversions +2.061%; fully deployed. Compared to DFM's delayed-feedback modeling, TWICE explicitly models the "two-clock" structure rather than approximating with a single timeline.
Auto-Bidding: From Single-Episode Optimization to Cross-Window Planning
SWAG-Bid (AliExpress) — Existing auto-bidding treats each day as an independent episode, but many advertisers' value generation is sparse enough that daily-level efficiency ratios are statistically unreliable — the platform switches to W=7-day sliding-window evaluation. This creates cross-episode coupling: each day's bids affect up to W overlapping windows. SWAG-Bid uses a Masked Trajectory Model to predict the market and generate candidate plans, Multi-Window Model Predictive Control Sampling (MWMS) to score across all overlapping windows, and Per-Step Gated Adaptive Layer Normalization (PSG-AdaLN) for state-adaptive gating. Validated on AliExpress for constraint satisfaction and value capture under sliding-window evaluation. "When the platform changes how it evaluates, bidding optimization objectives must follow" — a canonical case of ad systems evolving with business rules.
PlatformBid (Kuaishou) — The first auto-bidding benchmark designed from a unified ad-platform perspective (integrating SSP, DSP, and Ad Exchange simultaneously). Defines three competitive scenarios: homogeneous competition (all advertisers use the same algorithm), heterogeneous competition (diverse algorithm strategies), and promotional competition (budget surges during sales events). Systematically evaluates classical control, RL, and generative methods. Proposes BidFlow, a flow-matching-based auto-bidding method using expressive policy representations for dynamic competitive environments. Online target cost +0.68%, validating offline-online consistency. The benchmark design itself is a contribution to problem definition — when the platform is both referee and player, the bidding optimization objective must include both advertiser conversions and platform revenue.
Cost-Gated LLM Retrieval
SMART (Snap) — Dynamic product ads need to balance retargeting (resurfacing known interests) and exploration (discovering new categories). Controlled experiments reveal a key decomposition: rule-generated queries excel at retargeting on lexical BM25 indexes; LLM-generated queries excel at exploration on dense ANN indexes. SMART uses a lightweight quality gate to identify coverage gaps in initial keyword results, routing only ~10% of users who benefit from semantic exploration to the LLM path. Offline evaluation captures most of the semantic exploration gains at 90% lower LLM cost. Online A/B: ad conversion rate +27.6%. LLM value is suboptimal for most users most of the time — routing mechanisms determine ROI more than the model itself.
Directions to Watch
Direction 1: The "systems engineering" race in generative recommendation. Six deployment papers this week (Multi-Decoder OneRec, OxygenREC-v2, FDP, PSG, RGD, UniR²) share a common feature: they no longer chase a single metric of SID generation capability, but solve multi-objective conflicts, inference latency, task unification, and business controllability. Generative recommendation has moved from "can it retrieve the target item" to "can it control the trade-offs across multiple business objectives within budget." A question worth tracking: when generative recommendation simultaneously handles retrieval, ranking, and reranking, how does the evaluation system keep up?
Direction 2: Latent reasoning compresses LLM recommendation cost. WhisperRec, LaRec, and HiLaR independently converge on the same judgment — explicit CoT's information content is worth preserving, but its inference cost is unacceptable. Latent reasoning distills CoT into continuous representations, eliminating per-token generation overhead while preserving decision-relevant information. WhisperRec's 10x throughput gain is the strongest deployment evidence to date. This direction continues the lineage of 2025's SCoTER and Coconut, but has advanced from "can it converge" to "can it ship."
Direction 3: Architectural built-ins for training-serving consistency. Memory Layer brings serving-time caches into the training loop, ROCS builds request-side compute reuse into the model structure, OneShot aligns index learning with ranking objectives. Three papers from the same company (Meta) point to a system-level judgment: recommendation quality bottlenecks increasingly come from structural misalignment between training and serving, not from single-model capability ceilings. Engineering tools (caching, distillation, indexing) are being redesigned as modeling problems.
Paper Roundup
Generative Recommendation
Multi-Decoder OneRec — Kuaishou proposes multi-decoder controllable generative retrieval: shared representations + LoRA experts isolate objectives; online app time +0.37%, cold start +2.09%; open-sources Kwai26 benchmark (1.31B records).
OxygenREC-v2 — JD proposes a 3B-parameter MoE generative recommender; behavior-instruction conditioning + entropy-aware trajectory self-distillation internalizes discriminative signals; GMV +2.8-6.8%.
FDP — Kuaishou proposes feedback-driven policy discovery; distinguishes intent knowledge from policy knowledge, distills into latent tokens; online Revenue +4.506%, ADVV +4.621%.
PSG — Kuaishou proposes pair-space generative reranking; generation atoms upgraded from items to ordered pairs; 1.83x practical speedup; online dwell time +0.178%.
UniR² — Kuaishou unifies generative retrieval and multi-objective ranking into a single decoder-only sequence; Dual-Query Prefix-Causal Attention + LoRA separation.
RGD — Kuaishou proposes reward-guided decoding; KL-regularized reward maximization derives closed-form decoding distribution; aligns with business value without generator retraining.
Gwhere — Amap proposes contrastive residual quantization tokenizer + EAKTO reinforcement learning; solves token conflicts and heterogeneous signal fusion for POI generative retrieval; P-CTR +5.83%, U-CTR +6.20%.
ClockRoPE — YouTube proposes a random Fourier rotation theoretical framework; improves RoPE's temporal periodicity modeling; deployed in production-grade generative retrieval.
Dual-purpose Semantic IDs — Google uses hierarchical quantization to compress continuous embeddings into discrete semantic IDs serving dual roles of collaborative identity and content reconstruction; replaces large-scale vector storage.
EGR — Snap proposes Embedding-Native generative retrieval; shared LLM learns item and user representations jointly; items indexed directly as dense vectors; online conversion +2.91%.
LGRID — Kuaishou proposes LLM-driven generative disentanglement; interpretable SID positions; collision rate from 97.0% to 39.9%; AUC +5.44%.
TopoGR — Academic: bit-decomposable Binary SIDs expose Hamming geometry; three-stage topology preservation; surpasses SOTA on four datasets.
VaLiDRec — Academic: variable-length LLM-aligned semantic IDs; token-set prediction eliminates autoregressive decoding; inference 87.49x faster than LC-Rec.
LoopMemGR — Academic: closed-loop recommendation experience memory; three-view evidence extraction compressed into experience tokens; validated on Taobao industrial datasets.
SPARC — Academic: sequence-aware progressive attribute routing and compression; contextualize-then-compress; enriches historical representation without increasing input length.
CogRec — Academic: structural cognitive fast-slow reasoning; SID Routing (Match/LateralJump/Explore) anchors reasoning on SID topology.
Grevo — Academic: evolutionary item indexing; SID assignment as an evolvable discrete variable; trained recommender as posterior evaluator.
Restoring Collaborative Signals — Academic: personalized natural language injects hierarchical collaborative cues; no backbone changes, no SID retraining; restores collaborative signals missing from SIDs.
LLM & Agents
Melo — NetEase Cloud Music deploys an LLM music recommendation agent; five-node state graph + inference-time entity grounding + reflective retry; retention +2pp, play time +1 minute.
WhisperRec — Kuaishou proposes a latent reasoning framework; MV-ACoT + three-stage latent alignment; SID@64 +17.44%, throughput up over 10x.
ASARL — Tencent QQ Search uses three-agent data curation + three-stage training to solve social search relevance.
NEXT — Meta proposes an 8B vision-language model; Item-to-Intent-to-Item reasoning drives video recommendation; online watch time +0.53%.
HiLaR — Academic: hierarchical latent reasoning; layer-aware process rewards optimize reasoning trajectories.
LaRec — Academic: latent pretraining + personalized RL fine-tuning; Gaussian mixture distribution samples reasoning starting points; multi-path exploration of user interests.
CALMRec — Academic: causally aligned language memory; distinguishes short-term/long-term/exposure memories; discounted long-term value +6.1-7.6% in simulated environments.
Ranked by Position — Academic: reveals position-bias attack surface in LLM listwise rerankers; promo@5 reaches 0.57.
JKO-RAG — Academic: Wasserstein gradient flows for reranking; linear response theory explains stability advantages; surpasses cross-encoders on five BEIR benchmarks.
REPREC — Academic: lightweight user representation alignment; trains only an MLP injector; 1.51x training speedup while retaining 85-100% of LoRA performance.
Hypothesis-Driven Shelf — Spotify replaces fixed templates with natural-language hypotheses to generate personalized shelves; four-stage decomposition planning and retrieval.
SharpRec — Academic: sharpness-aware model merging + saliency recovery; solves cross-domain knowledge conflicts and fusion saturation.
RecoReward — Kuaishou uses recommender affinity scores as rewards; RL trains MLLM to generate recommendation-oriented descriptions; content-only inference preserved.
Two Views, One Voice — Third-place solution for RecSys Challenge 2026; decouples retrieval and generation; PAS framework for structured evidence allocation.
GGR/SAR — Xiaohongshu proposes graph-guided retrieval + scene-aware ranking; data asset discovery Hit@10 from 19.1% to 96.6%.
Industrial Architecture
ROCS — Meta proposes request-oriented compute sharing; delayed interaction + candidate isolation; retrieval model QPS +200%.
OneShot — Meta: end-to-end index learning aligned with ranking objectives; fully deployed on Instagram; recall +20%, 10x efficiency.
CCFormer — Tencent unifies feature interaction and long-sequence modeling; CTR +3.57%, ad revenue +1.71%, training 2.21x faster.
Memory Layer — Meta uses in-model key-value cache to unify training-serving representations; NE gap reduced 86%; cold start +5-6%.
Mosaic — Meta: user embedding expert fleet; MRM + CRL maximize marginal information contribution; log-free evaluation framework.
ConAlign — Kuaishou: discrete-gated conditional alignment; first streaming debiasing industrial deployment; significant gains in long-term engagement and interest diversity.
PinEqualizer — Pinterest: full-funnel content exploration and debiasing system; two years of deployment with significant cold-start and ecosystem health improvements.
HA-MoE — Google: heterogeneity-adaptive multi-gated MoE; explicit heterogeneous context into gates and expert representations; DL-AUC evaluation metric.
Bumblebee — Meta: interleaved hybrid layer blocks; micropipeline combines sequence personalization + attention encoding + feature crossing.
Breaking the Loop — YouTube Music: online comparison of six intervention strategies; serving-layer interventions canceled by the learning loop; SNGP exploration yields largest new-release gains.
SpecFormer — Alibaba: spectrum-aware Transformer; mitigates embedding and attention collapse; deployed in commercial systems.
Entity Resolution — Walmart: three practical lessons from self-service entity resolution; automatic bake-off selects matching algorithms.
Embedding Items at Scale — Yandex compares GNN pretraining vs. end-to-end embeddings; no substantive pretraining gains with sufficient data.
KuaiLive-M3 — Kuaishou releases multimodal multi-domain multi-feedback live-streaming dataset; 21.9K users + 88M segment-level multimodal embeddings + 25.4K questionnaire feedback.
Tight Sample Complexity for LoRA — Academic: first tight sample-complexity bounds for LoRA; proves a rank-selection dichotomy.
DIRECTOR — Academic: optimal-transport-guided parallel reranking; entropy-regularized OT conflict-aware training; duplicate-free slate generation.
Improving Item Discoverability — Instacart uses LLM-generated related intents to expand retrieval; SLM distillation covers tail queries; coverage from 60% to 80%.
Privileged Self-Distillation — Academic: future interactions as privileged information for self-distillation; single-stage training without changing the deployed model.
RAGuard — Academic: two-layer RAG poisoning defense; adversarial retrieval training + ZKIP zero-knowledge inference patch; attack success rate reduced to 0.000.
PEA-CAE — Ontario Power Generation: evolution from naive RAG to deep multi-agent retrieval; cost-aware progressive evidence acquisition.
IIMRec — Academic: single high-quality item-item graph reused across three stages; NCER reduces spectral noise-to-signal ratio.
Discrepancy-Rounded Fair Bandits — Academic: converts fair exposure constraints into a discretization problem; BDQ-UCB deterministically satisfies block-level constraints.
MAPLE — Academic: multi-alpha position-aware listwise ensemble; unified prediction head + diversity regularization; 55x fewer parameters.
LOCKS — Academic: page-local spectral summary KV cache selection; 2048-token budget matches full-cache quality of 100K+ context.
Graph Coarsening — Academic: two-stage diffusion framework — graph coarsening + label propagation; NDCG@5 +24%-50% on telecom datasets.
Cross-Domain OPE/L — Academic: cross-domain offline policy evaluation and learning; CD-IPS + CD-PG address few-shot and deterministic policy problems.
Advertising & Bidding
User Foundation Model — Teads trains a user foundation model for open-web RTB; Transformer pretraining + LLM-as-optimizer; CTR +2.13%, eCPC -1.13%.
TWICE — Kuaishou: two-clock two-window CVR decomposition; click clock + conversion clock complementary supervision; revenue +1.858%, conversions +2.061%.
SWAG-Bid — AliExpress: sliding-window-aware generative auto-bidding; Masked Trajectory Model + MWMS cross-window scoring.
PlatformBid — Kuaishou: unified ad-platform auto-bidding benchmark; three competitive scenarios + BidFlow flow-matching method; online target cost +0.68%.
SMART — Snap: semantics-aware adaptive retrieval; quality gate routes only 10% of users to the LLM path; conversion +27.6%, LLM cost -90%.
SAGE — Academic: constrained-gated LoRA + position enhancement + text alignment; under 10% trainable parameters reaches full fine-tuning performance.
DASH — Tencent: decision-aware user simulator; context engineering + rule-based reward model; action prediction accuracy +12.4% across five domains.