RecSys Weekly 2026-W32 | Recsys Frontier

type

Post

status

Published

date

Aug 8, 2026 05:35

slug

rec-weekly-en-2026-W32

summary

This week's recommendation systems research clusters around three technical threads: generative recommendation moving from proof-of-concept to end-to-end engineering, LLMs stepping from ranking assistance into core decision-making, and the pretrain-continuous refresh paradigm redrawing the boundary between knowledge and geometry. Industrial papers account for over half of the output — Yandex, Kuaishou, ByteDance, Tencent, Snap, Shopee, LinkedIn, JD, Microsoft, and Huawei all published deployment papers, most with online A/B data attached. Thread 1: Generative recommendation moves beyond the "generate-as-recall" prototype toward end-to-end single models. Yandex's Gryphon-v2 replaces a full cascade of 15+ candidate generators, coarse ranking, and fine ranking with a single model — active users +1.41%; Snap pushes LLM generative recall into short-video scenarios, View Time +0.37%. Both point to the same conclusion: the engineering bottlenecks of generative architectures (ranking objective transfer, inference cost, eligibility constraints) are being dismantled one by one. Thread 2: LLMs move from ranking assistance into high-stakes decision-making. Tencent's SeqLLM injects behavior sequence modeling into payment risk control, merchant screening precision up from 92.0% to 97.5%; Kuaishou's HOBA uses LLM inference for hyperparameters, SARSA for expert selection, and an expert pool for execution — a three-layer structure that makes bidding decisions adaptive online, target cost +3.6%. Baidu's QDET matches DeepSeek-R1-671B on timeline summarization with a 7B model, CTR +5.5%. Thread 3: The pretrain-continuous refresh paradigm begins redrawing the boundary between "knowledge" and "geometry." Shopee's KGD uses behavior multi-token prediction to clean pretrained knowledge and anchored calibration residuals to decouple task geometry — GMV/user +1.75%, validated over 90 days of production traffic with no degradation. This thread points to a judgment: the next battleground for pr

Weekly Overview

Thread 1: Generative recommendation moves beyond the "generate-as-recall" prototype toward end-to-end single models. Yandex's Gryphon-v2 replaces a full cascade of 15+ candidate generators, coarse ranking, and fine ranking with a single model — active users +1.41%; Snap pushes LLM generative recall into short-video scenarios, View Time +0.37%. Both point to the same conclusion: the engineering bottlenecks of generative architectures (ranking objective transfer, inference cost, eligibility constraints) are being dismantled one by one.

Thread 2: LLMs move from ranking assistance into high-stakes decision-making. Tencent's SeqLLM injects behavior sequence modeling into payment risk control, merchant screening precision up from 92.0% to 97.5%; Kuaishou's HOBA uses LLM inference for hyperparameters, SARSA for expert selection, and an expert pool for execution — a three-layer structure that makes bidding decisions adaptive online, target cost +3.6%. Baidu's QDET matches DeepSeek-R1-671B on timeline summarization with a 7B model, CTR +5.5%.

Thread 3: The pretrain-continuous refresh paradigm begins redrawing the boundary between "knowledge" and "geometry." Shopee's KGD uses behavior multi-token prediction to clean pretrained knowledge and anchored calibration residuals to decouple task geometry — GMV/user +1.75%, validated over 90 days of production traffic with no degradation. This thread points to a judgment: the next battleground for pretrained recommendation models isn't larger models, but how to hold onto learned knowledge amid continuous distribution drift while keeping downstream adaptation free from gradient interference.

Generative Recommendation: From "Generate-as-Recall" to End-to-End Single Models

Gryphon-v2 (Yandex) — unified generation-ranking architecture. User history is encoded once; an autoregressive decoder generates Semantic ID candidates, mapped back to category items, then fine-ranked by a Ranking Module that reuses the shared encoder state. The key design is Rollout Distillation: a Teacher Ranker exists only at training time, scoring two candidate distributions — candidates rolled out by the current decoder (so the Ranking Module sees generation outputs identical to serving-time sources) and logged exposure samples (covering items users actually saw). Distilled scores serve as the Ranking Module's sole ranking supervision; no second model is added to the serving path.

In online A/B, a single Gryphon-v2 replaces a production cascade of 15+ candidate generators — active users +1.41%, latency on par with the cascade. This extends the joint training approach of the previous-generation Gryphon — which used an item-level scoring component to resolve the mismatch between SID sequence likelihood and relevance objectives — but swaps the supervision source from "sequence likelihood" to "ranking preference." DualGR handled long/short-term interests with dual-branch routing and exposure-aware loss; Gryphon-v2's answer is more direct: since the ranking objective defines user value, distill the Teacher's ranking scores directly into the generator. This path shows the core tension in generative recommendation has shifted from "can it generate correctly" to "how to rank what's generated" — distillation is just the means; the landing point is the Ranking Module reusing the generator's encoding state.

SeqLLM (Tencent) — injecting behavior sequences into LLMs without losing language capability. The difficulty: LLMs excel at text but can't natively model long behavior sequences, and direct fine-tuning causes catastrophic forgetting. The solution is three components: a compact discrete vocabulary (encoding behavior events as native tokens), a lightweight projector trained with two-stage aligned curriculum (mapping behavior tokens into LLM semantic space), and prefix-guided capability injection (using task prefixes to constrain SFT rather than continual pretraining). In WeChat Pay production, merchant screening precision rose from 92.0% to 97.5%; behavior token embeddings lifted production-grade fraud detection Precision@Top-0.01% by 26.8 percentage points. On public benchmarks, Recall@5 is 32% higher than the User-LLM baseline, and Pass@32 on RecIF beats the full OneRec-8B pipeline by 14.2% — using only one-fifth of the GPU days. Complementing xGR's serving cost optimization, SeqLLM answers "how does an LLM learn behavior sequences without losing its native capabilities" — pointing at high-false-positive-cost scenarios like risk control and anti-abuse.

SnapLGR (Snap) — LLM generative recall enters short-video serving for the first time. Three design blocks: multimodal SIDs build semantic identifiers from video embeddings, with PPR co-participation and contrastive learning injecting collaborative signals (improving codebook utilization, reducing collisions); continued pretraining (CPT) grounds SID tokens before SFT; TensorRT-LLM CUDA backend beam search plus decentralized worker-loops guarantee latency. Online comparison against a TIGER-style baseline: View Time +0.37%, Time Spent +0.09%. The team then froze the tokenizer and decomposed the offline gap into the individual contributions of model architecture, scale, and pretraining — this kind of "which step contributed what" decomposition is rare in industrial deployment papers, which usually report only final numbers. DAS used dual-stream alignment for semantic IDs at Kuaishou; SnapLGR uses PPR contrastive learning to inject collaborative signals. Both acknowledge pure semantic SIDs aren't enough — the collaborative dimension must be added.

RecHarness (Kuaishou) — LLM-driven automated model optimization. Rather than letting the LLM generate modifications directly (fast but unstable ideas), it splits into two steps: a bandit router selects the next modification direction based on historical validation feedback, and the LLM generates concrete hypotheses and code changes within that direction. A basin-hopping mechanism activates a structural jump arm when local edits stall, maintaining long-range exploration. Over 7 days of online A/B: ADVV +2.084%, Revenue +0.534%, Exposure +0.559%. Compared to Design Once, Deploy at Scale's templated development, RecHarness is a hybrid of "exploration strategy + LLM execution" — binding the LLM's generation capability inside the bandit's exploration framework, limiting the entropy of the search space.

DEGR (JD) — dual-exploration-driven generative re-ranking. Re-ranking is constrained by fixed upstream supply; under low-quality supply there's no headroom for gains. DEGR uses an exploration reward model to balance immediate and exploration value — under low-quality supply it prioritizes exploration exposure, preserving browsing potential and creating serendipitous conversions. The hybrid optimization integrates supervised learning, exploration diversity constraints, and adaptive reward-weighted ORPO. Online: UCTR +1.22%, PV +0.20%. Gwhere used RL objectives for generative POI recommendation; DEGR explicitly models "exploration" into re-ranking — not re-arranging combinations, but bridging context across requests.

Academic-side generative recommendation work centers on distillation efficiency and cold start. SmartGR targets uneven difficulty in hierarchical SID distillation and beam search prefix pruning errors — hierarchical-aware SID distillation plus beam-aware ranking distillation, averaging 8.6% improvement across four benchmarks with 2.39× inference speedup. Exp-RSFT directly optimizes log rewards with exponential weights (exp(r/λ)), theoretically decomposing suboptimality into coverage cost and noise cost, with temperature λ balancing the two — an inverted-U curve across three public benchmarks and industrial datasets, avoiding PPO/DPO reward over-optimization. UnpairGR learns a unified semantic ID space, letting paired and unpaired multimodal observations share a Transformer and residual codebook — no feature imputation or fallback mapping needed. OMEGA compresses user sequences into compact representations with learnable query tokens, builds a collaborative memory bank, and fuses via target-aware retrieval plus gated cross-attention.

EvoReason (Kuaishou) — the distillation supervision problem for implicit reasoning. Raw CoT trajectories are redundant with unstable paths; direct distillation into latent space performs poorly. EvoReason first extracts "reasoning primitives" (reusable reasoning behaviors), has the teacher generate structured CoT based on primitives, then uses self-evolving on-policy distillation where the student's implicit reasoning results feed back to the teacher — a closed-loop iteration aligning supervision signals with the student's reasoning space. LLM-Derived Priors (NAVER WEBTOON) converts semantic signals extracted by LLMs from comment text into Bayesian priors, warming up Thompson sampling for cold start. Real A/B/C testing validates that gains are largest under sparse feedback, with different effects across funnel stages. RRC addresses the mismatch between generative reward models and RL's scalar scoring — building rewards from relative preference rankings, AlpacaEval2 35.8%→41.3%. The Position Bias Audit proposes the InvariRank framework, quantifying LLM re-ranker sensitivity to candidate permutations across three layers: pairwise preference instability, global preference inconsistency, and list output consistency — the results show that reducing exposure bias is insufficient to restore ranking effectiveness; improving relevance or flattening exposure doesn't guarantee recovery of stable preference structure.

Retrieval and Multimodal Representation: From "Can Retrieve" to "Can Discriminate"

DME (ByteDance/Douyin) — two-stage multimodal embedding. Stage 1: large-scale contrastive pretraining establishes a unified multimodal space (broad coverage, multi-task). Stage 2: two mechanisms add "semantic sufficiency" — the concept positioned as: embeddings must not only be similar to the counterpart but also carry retrieval evidence and preserve fine-grained semantics from the counterpart side. Evidence-grounded latent reasoning organizes retrieval evidence in latent space; cross-condition reconstruction forces counterpart-side semantics through cross-directional autoregressive reconstruction. Both operate only at training time; serving uses a standard contrastive encoder. On MMEB-v2, 2B/9B variants reach 74.8/78.4, strongest on video and visual document tasks; Douyin internal offline eval +2.92%, online search LT +0.1%. Complementary to HIVE's query-hypothesis-verification framework — DME internalizes "hypothesis-verification" into latent space without relying on explicit generation.

GALA (Alibaba/Taobao Flash Sale) — three-stage multimodal alignment. Behavior-aware triplet pretraining on search logs captures early user intent; a middle generative RL alignment stage uses GRPO to dynamically optimize multimodal embeddings by conversion reward, bridging the pretrain-finetune divergence; finally, adaptive gating (mixed loss) fuses multimodal and ID embeddings, preventing long-term ID-dominated training from eroding multimodal contributions. Serving 200M+ DAU, offline AUC +0.12/+0.20, order volume +0.55%. TRM's semantic-token-replaces-item-ID approach is inverted in GALA: not removing IDs, but keeping multimodal consistently contributing under ID dominance.

PCR-CA (Microsoft) — parallel codebooks for multi-category app semantics. Traditional RQ-VAE is hierarchical residual quantization — when category boundaries are fuzzy, it can only approximate layer by layer. PCR-CA learns multiple discrete codebooks in parallel, independently encoding different semantic aspects (gameplay, art style, genre), with contrastive alignment losses bridging semantic and collaborative signals at both user and item levels, and dual attention fusing ID features with semantic features. Long-tail app AUC +2.15%, online CTR +10.52%, CVR +16.30%, fully deployed on Microsoft Store. Compared to DAS's dual-stream alignment, PCR-CA's differentiator is "parallel codebooks" — handling multi-label semantics without relying on prefix-dependent hierarchical structures.

SPEAR (Dewu) — end-to-end rewriting and retrieval for community search. Porting the path-based architecture to e-commerce search produces a "generic-term dominance" effect: the model prefers rewrites that score well but drift from query intent. Three components map to three failure modes: a dual-embedding backbone with auxiliary loss and gradient isolation protects recall-side semantics from CTR-driven ranking signal erosion; a multiplicative gating aggregator requires both rewrite confidence and item relevance to be strong simultaneously, eliminating the generic-term shortcut; a dynamic rewrite selector generates rewrite weights per request with user-query-conditioned scaling bias. Offline semantic similarity@10 +18.2, click recall@10 +99.5; online query-view CTR +0.259, average reading depth +0.733.

RCBS (Karrot) — region-constrained batch sampling. Local community platforms have geographically constrained exposure; standard in-batch negatives treat "geographically impossible negatives" as negatives, diluting the contrastive signal. RCBS constructs region-homogeneous mini-batches where users only contrast against items they can actually reach — naturally producing harder, more useful negatives. Consistent offline and online gains across home feed ranking, retrieval, and display ad ranking; user embeddings deployed to production. Unlike Gwhere's contrastive residual quantization tokenizer approach, RCBS doesn't touch model architecture — only the sampling distribution. Minimal intervention, consistent gains.

PaletteID — prototype-composition semantic identifiers. SQ-DPP selects "prototype palettes" (considering both local content density and global semantic diversity); each item retrieves a set of semantically related prototypes aggregated into a representation. Addresses codebook assignment losing continuous signals and residual code paths over-relying on prefixes. Consistent gains on Amazon/TikTok, with larger improvements on long-tail items.

Two audit papers this week are worth noting. The Modality Weighting Audit unifies six per-user modality weighting implementations onto a shared collaborative backbone, finding that a single global weight captures most of the content gain (+1.9/+3.6/+3.5pp), with per-user weighting showing no consistent utility (gains ≤0.9pp and flipping across corpora). The Gender Sensitivity Mechanism Analysis localizes gender signals in dense retrieval as originating in input embeddings and propagating through a small number of late attention heads — embedding-level intervention non-specifically neutralizes score differences, while attention-level intervention produces directional shifts.

Four solid works on retrieval efficiency. EXCISE handles exclusion-query failures in late-interaction retrievers ("want X not Z" actually promotes Z documents) — two query-side modules totaling 1.5M parameters identify excluded topics and re-embed a shortlist of 100 documents, ExcluIR exclusion success@10 from 0.058 to 0.691. MarginMerge does coverage-aware multi-vector compression: 97-99% nDCG@5 retention, 90-95% storage vector reduction, ~41% fewer ranking flips. Hierarchical BM25 uses a coarse index for group selection plus in-group exact scoring — 4.4GB fixed memory for a billion documents, ~300ms per query. CeQe uses cross-encoder token-level attribution for query expansion — every expansion term is copied verbatim from retrieved passages, never introducing vocabulary absent from the corpus, NQ Recall@100 from 0.32 to 0.47.

Industrial Recommendation Systems Engineering: Knowledge Flow, Serving Flow, and Experiment Infrastructure

KGD (Shopee) — knowledge-geometry decoupling for pretrain-continuous refresh. Answers two questions: what to learn from behavior sequences, and how to transfer under continuous refresh. BMTP predicts only collaborative or semantically related future items, avoiding spurious transfer across unrelated sessions; ACR writes task geometry in residuals orthogonal to pretrained embeddings — the encoder holds refreshable behavior knowledge, and the task learner reads encoder state through read-only cross-attention, untouched by task gradients. 4-12% improvement across eight public benchmarks; in 90 days of production traffic, the baseline shows no gains while KGD maintains its advantage. GMV/user +1.75%, ad revenue +1.53%. Viewed against GenRec's page-level NTP training objective, the supervision signal for pretrained recommendation models is refining from "predict the next behavior" to "predict the parts of behavior that actually carry learning value."

QDET (Baidu) — multi-task timeline fine-tuning + RL concise summarization. Three auxiliary tasks (temporal ordering, causality judgment, timeline completion) bring a 7B model to 76.2 F1 on timeline summarization, matching DeepSeek-R1-671B (76.1) — 1% of the parameters, zero-shot gap eliminated. RL summarization achieves 88.2% length compliance, 7.7 points above the 671B model. Online: CTR +5.5%, dwell time +4.6%, exploration depth +4.4%. Similar to HiGR's list-level planning approach, QDET treats "timeline" as a list in recommendation — but supervision comes from event relations rather than user feedback.

TransX (LinkedIn) — behavior stream and serving stream separation. Reconstructs recommendation as sequence-to-sequence action transduction: a nearline behavior encoder handles the long-term behavior stream, online real-time representations form the serving stream, and decoding conditions on both via scalable cross-attention. Amortized serving (incremental behavior encoding + per-request KV caching) makes latency insensitive to behavior sequence length, cutting online computation by ~80%. CTR +6.0%, conversions +4.4%. xGR does KV-cache-separated serving optimization; TransX architecturally splits "long-term behavior" and "real-time events" into two streams — the division of labor in sequence modeling is this week's clearest architectural judgment.

STEPS (ByteDance/Douyin) — self-triggered push. Reconstructs "whether to push + when to push" as a self-triggering agent pipeline: a planning agent uses gated ordinal regression to schedule the next system call, an execution agent decides whether to push based on trajectory rewards, and a lightweight filtering agent controls compute overhead and prevents unreasonable plans. Fully deployed on Douyin covering 1B+ users: active days +0.2843%, push permission opt-out rate -1.9089%, compute overhead -79.42%. A concrete instantiation of the Aligning Large Language Models for Controllable Recommendations two-stage alignment framework in the push scenario — planning and execution separated, making "when to trigger" a learnable decision rather than a fixed interval.

SITA (Kuaishou) — target-aware compression. Parallel semantic quantization learns semantic interest tokens, conditioned aggregation over the target item's semantic ID builds user representations. The long-standing debate in long-sequence recommendation is "target-awareness vs. efficiency" — dynamically retrieving relevant behaviors is target-aware but compute depends on the target; compressing the full sequence is efficient but target-agnostic. SITA organizes compressed interests with semantic structure, and conditioned aggregation unifies both. Complementary to AdaSID's collision handling — that work solves ID assignment, this solves ID consumption.

Twitch Live Ranking — delayed-window extended feedback collection, multi-model architecture combining freshness and delayed signals, segment-aware targeting optimizing ranking by user lifecycle stage, MMoE integration cutting 41.9% of parameters. DAV +0.09%, high-activity user ARPU +0.56%, new follows +0.27%; Twitch mobile validated positive interactions +1.12%. VLM Relevance Evaluation (Pinterest) uses vision-language models as relevance labelers for online experiments, substantially reducing minimum detectable effect (MDE).

Experiment infrastructure signals are dense. WatchLens is an open-source video recommendation experiment platform that associates recommendation policies and ranking positions with every event at logging time. UpliftBench evaluates 12 uplift estimators across 7 dataset families, finding Qini coefficient rank-correlates with effect accuracy at only +0.07 while AUUC aligns better — metric choice determines conclusions. This isn't methodological purism; it's the factual basis for engineering decisions. Trajectory-to-Evidence (Kuaishou) is an audit framework for research agents — completed experiment trajectories don't equal evidence; you need verified artifacts, scoped claims, and auditable records.

Academic systems: Cost-Aware Multi-Objective Bandit gives a budget regret bound of O(Σ log B/Δ) for hypervolume-based UCB; TimeRLM uses recursive language models for long-context anomaly localization, AnomalyXL localization IoU 0.682 vs. baseline 0.329; ATLAS uses Gromov-Wasserstein alignment + adversarial objectives + RVQ codebooks for recommendation domain generalization — trained on five source domains, directly evaluated on ten unseen domains, HitRate +24%; CILER brings conditional identifiability theory to OOD recommendation. TabDPT-Turbo replaces retrieval with long-context retrieval-free row-wise attention — performance on par with TabDPT v1.1 but orders of magnitude faster.

Advertising and Bidding Optimization: Online Adaptation and Paradigm Thinking

HOBA (Kuaishou) — hierarchical on-policy bidding agent. Three layers map to three time scales: the LLM top layer uses a Think-Act-Observe-Reflect loop to infer hyperparameters from context (including historical experience retrieval); the middle SARSA layer selects among expert models, with causal adjustment removing selection bias; the bottom expert pool (PID, MPC, IQL, Decision Transformer) executes bidding under upper-layer constraints. Key constraint design: online learning is restricted to discrete expert selection rather than continuous bid optimization, substantially reducing exploration risk. Validated on AuctionNet and large-scale online A/B: target cost +3.6%. Isomorphic to STEPS's planning-execution separation — high-level low-frequency strategic decisions, low-level high-frequency execution, with a middle selection layer isolating exploration risk.

GRACE (Meta) — serving system for generative ad recall. Two challenges: eligibility constraints (every generated ad must satisfy advertiser targeting rules) and compute cost (thousands of ads generated per request under wide beams). GTM encodes targeting rules as bitmask/Bloom filter matchers over SID prefixes for constrained decoding — SID-level GTM lifts ad-level target matching pass rate from 23.55% to 40.42%. The encoder-decoder Transformer gets redesigned attention kernels, KV caches, and beam search adapted for wide-beam short-sequence scenarios. On GH200: cross-attention latency down 68×, self-attention down 23.4-25.8×, overall decoding latency down 11.1×.

Competition-Aware Request Distribution (Huawei) — distributional bid prediction plus probabilistic forwarding deciding whether to send requests to each DSP, with lightweight policy optimization adaptively adjusting per-DSP thresholds to track non-stationary markets. On a production platform handling 20B+ daily requests, full multi-DSP deployment: DSP request volume -34.2%, net revenue +4.6% (p<0.001). Segment analysis shows aggregate metrics are misleading — the policy is actually selecting beneficiaries by "comparative advantage" across DSPs, not simple traffic throttling.

GOAL (Kuaishou) — constraint-aware generation for incentivized ads. Models continuous incentive amount allocation as conditional sequence generation, integrates a hierarchical causal state encoder capturing local behavior dynamics and long-range dependencies, and SCPO learns a single generation policy that generalizes across ROI constraints — no retraining per constraint. Validated on real data plus synthetic fatigue environments: improved long-term revenue and retention, reduced ROI violation rates.

One-Shot Pricing (academic) — theoretical framework. In HOTW (hands-off-the-wheel) ad markets, all information needed for optimal pricing resides on the exchange side. The market is equivalent to a Fisher market; the Eisenberg-Gale convex program computes market-clearing prices and allocations in one pass, satisfying both budget and ROI constraints. This price is revenue-optimal among uniform pricing mechanisms and equivalent to the equilibrium outcome of sequential first-price auctions with pacing — one convex program replacing millions of real-time auctions. ProductWebGen is a multimodal benchmark for product webpage generation: 500 test samples across 13 categories, systematic comparison of edit-based and unified-model workflows.

Directions to Watch

"Ranking distillation" is becoming standard in generative recommendation. Gryphon-v2's Rollout Distillation, SmartGR's hierarchical-aware distillation, and DEGR's exploration reward model — three paths pointing to the same judgment: the serving bottleneck for generative recommendation isn't parameter scale, it's how the ranking objective is transferred. Yandex, JD, and Kuaishou are all doing it. The next thing worth watching is "distillation objective unification" — the trade-off between Teacher objective functions still being generation probability with ranking preference as fine-tuning, versus ranking preference as primary supervision.

LLMs entering high-stakes decisions is this week's clearest new scenario. Tencent risk control, Kuaishou bidding, NAVER WEBTOON cold start, Baidu search timelines — the common pattern is that LLMs don't do final scoring; they provide "semantic priors" (semantic landing points for behavior tokens, hyperparameter inference, Bayesian priors, event relation judgments), while lightweight models execute the actual decisions. This division of labor positions LLM cost-benefit at "warm-starting high-false-positive-cost scenarios" rather than per-request inference.

The audit turn in experiment infrastructure deserves attention. UpliftBench reveals metric choice determines conclusions, the modality weighting audit rejects the general value of per-user weighting, the gender mechanism analysis localizes debiasing intervention points, and RecHarness trajectory audits expose non-monotonic trajectory evolution — four independent works jointly demanding stricter falsification standards. The direct implication for industrial teams: before adding complex modules, run "global simple baseline + permutation control" first. This may be the week's lowest-cost, highest-leverage advice.

Paper Roundup

Generative Recommendation and LLM Enhancement

Gryphon-v2 — Yandex proposes a unified generation-ranking architecture; Rollout Distillation transfers Teacher Ranker ranking preferences; a single model replaces a 15+ candidate generator cascade; active users +1.41%.

SeqLLM — Tencent injects behavior sequence modeling into LLMs (discrete vocabulary + projector + prefix-guided injection); WeChat Pay merchant screening precision 92.0%→97.5%, Precision@Top-0.01% +26.8pp.

SnapLGR — Snap deploys LLM generative recall; multimodal SID + PPR contrastive learning + CPT + TensorRT-LLM; View Time +0.37%.

RecHarness — Kuaishou uses bandit routing + LLM generation + basin-hopping for automated model optimization; online ADVV +2.084%.

DEGR — JD dual-exploration-driven generative re-ranking; exploration reward model + adaptive ORPO; UCTR +1.22%.

SmartGR — Hierarchical-aware SID distillation + beam-aware ranking distillation; average +8.6%, inference 2.39× faster.

Exp-RSFT — Exponential reward-weighted fine-tuning for generative recommenders; temperature λ balances coverage and noise costs; outperforms PPO/DPO.

UnpairGR — Beihang/Meituan shared semantic ID space leveraging unpaired multimodal observations.

EvoReason — Kuaishou reasoning-primitive-guided on-policy distillation aligning CoT supervision with student implicit reasoning space.

Think2Go — SFT+RL unified for generative POI recommendation; two advantage-weighting mechanisms implement implicit curriculum learning.

OMEGA — Collaborative memory-augmented generative recommendation; target-aware retrieval + gated cross-attention.

GARDRec — Decision-level graph grounding; KG enters decision branches from prompt evidence.

LLM-Derived Priors — NAVER WEBTOON warms up Thompson sampling cold start with LLM semantic priors; largest gains under sparse feedback.

RRC — Building rewards from relative preference rankings resolves generative reward model vs. RL scalar scoring mismatch; AlpacaEval2 35.8%→41.3%.

Position Bias Audit — InvariRank framework quantifies LLM re-ranker position sensitivity; reducing exposure bias is insufficient for ranking effectiveness.

AgentCF Attack Defense — How connectivity (candidate count + catalog concentration) modulates multi-agent CF attack/defense effectiveness.

RAG Knowledge Extraction Benchmark — First systematic knowledge extraction attack/defense benchmark; unified protocol for cross-language evaluation.

Retrieval and Multimodal Representation

DME — ByteDance two-stage multimodal embedding; contrastive pretraining + evidence-grounded latent reasoning + cross-condition reconstruction; MMEB-v2 74.8/78.4, online LT +0.1%.

GALA — Alibaba three-stage multimodal alignment; middle generative RL alignment (GRPO) bridges pretrain-finetune gap; order volume +0.55%.

PCR-CA — Microsoft parallel-codebook VQ-AE for multi-category semantics; long-tail AUC +2.15%, CTR +10.52%; deployed on Microsoft Store.

SPEAR — Dewu dual-embedding backbone + multiplicative gating + dynamic rewrite selector; online CTR +0.259, reading depth +0.733.

RCBS — Karrot region-constrained batch sampling; feasible negatives replace geographically impossible ones; consistent gains in ranking and retrieval.

PaletteID — Prototype-composition semantic identifiers; SQ-DPP builds prototype palettes.

Modality Weighting Audit — Global weights capture most gains (+1.9/+3.6/+3.5pp); per-user weighting shows no consistent utility (≤0.9pp and flipping).

Gender Sensitivity Mechanism Analysis — Gender signals originate in input embeddings propagating through late attention heads; attention-level intervention produces directional shifts.

EXCISE — Query-side exclusion mechanism; ExcluIR exclusion success@10 0.058→0.691.

MarginMerge — Coverage-aware multi-vector compression; 97-99% nDCG@5 retention, 90-95% vector reduction.

Hierarchical BM25 — Billion-document scale at 4.4GB fixed memory, ~300ms per query.

CeQe — Cross-encoder token-level attribution query expansion; NQ Recall@100 0.32→0.47.

Align-RAG — Closed-form magnitude scaling + phase shifting aligns retrieval windows; training-free outperforms trained adapters; MSE down 3.75% on average.

DocRetriever — Layout-aware sparse embeddings without OCR + inference-enhanced re-ranking; builds MultiDocR benchmark.

HyperAgent4POI — Multi-agent hypergraph dynamic semantic message passing; NDCG@20 +8.2% at 60% modality missing rate.

DCL — Decoupled semantic/language subspaces; cross-lingual zero-shot dense retrieval.

Bayesian Data Reweighting — Latent variable modeling of query-document importance; closed-form posterior adaptive downweighting of false negatives.

Capability Pages — Cluster-contrastive skill representations (positive triggers + negative boundaries + discriminative subjects); SRA-Bench average Recall@10 +2.94.

Industrial Recommendation Systems Engineering

KGD — Shopee knowledge-geometry decoupling; BMTP cleans pretrained knowledge + ACR decouples task geometry; GMV/user +1.75%, ad revenue +1.53%.

QDET — Baidu multi-task timeline fine-tuning + RL concise summarization; 7B matches DeepSeek-R1-671B (76.2 vs 76.1 F1); CTR +5.5%.

TransX — LinkedIn behavior/serving stream separation + amortized serving; CTR +6.0%, conversions +4.4%, compute -80%.

STEPS — ByteDance self-triggered push agent; planning + execution + filtering three-layer decisions; active days +0.2843%, compute overhead -79.42%.

SITA — Kuaishou semantic interest tokens for target-aware compression; unifies target-awareness and efficiency in long-sequence modeling.

Trajectory-to-Evidence — Kuaishou research agent audit framework; verified artifacts + scoped claims + auditable records.

WatchLens — Open-source video recommendation experiment platform; modular and configurable; policies and events logged with associations.

Cost-Aware Multi-Objective Bandit — HV-UCB budget regret bound O(Σ log B/Δ); CAGE exponential error probability.

Twitch Live Ranking — Delayed windows + segment-aware targeting + MMoE; DAV +0.09%, ARPU +0.56%.

CILER — Conditionally identifiable latent environment modeling; improvements across all twelve OOD recommendation metrics.

ATLAS — Sony Gromov-Wasserstein alignment + adversarial objectives + RVQ codebooks; five source domains → ten unseen domains, HitRate +24%.

TimeRLM — Recursive language models for long-context anomaly localization; localization IoU 0.682 vs. baseline 0.329.

VLM Relevance Evaluation — Pinterest replaces human relevance labeling with VLMs; substantially reduces MDE.

UpliftBench — Qini rank-correlates with effect accuracy at only +0.07; AUUC aligns better (+0.49).

TabDPT-Turbo — Retrieval-free long-context row-wise attention tabular foundation model; performance on par with TabDPT v1.1 but orders of magnitude faster.

Advertising and Bidding Optimization

HOBA — Kuaishou hierarchical on-policy bidding agent; LLM hyperparameter inference + SARSA expert selection + expert pool execution; target cost +3.6%.

GRACE — Meta generative ad recall serving system; GTM target matching pass rate 23.55%→40.42%; decoding latency down 11.1×.

Competition-Aware Request Distribution — Huawei distributional bid prediction + probabilistic forwarding; DSP request volume -34.2%, net revenue +4.6%.

GOAL — Constraint-aware generation framework for incentivized ads; SCPO generalizes across ROI constraints.

One-Shot Pricing — HOTW markets equivalent to Fisher markets; Eisenberg-Gale convex program computes revenue-optimal equilibrium prices.

ProductWebGen — Multimodal product webpage generation benchmark; 500 samples across 13 categories; edit-based vs. unified model comparison.