RecSys Weekly 2026-W46
2026-11-14
| 2026-11-14
字数 4375阅读时长≈ 11 分钟

Weekly Overview

Recommendation systems research this week clusters around three technical threads: how to reuse sequence-encoding compute, how to make generative retrieval generate compliant outputs, and how to model reachability under long-tail and multi-objective constraints.
Thread 1 — the foundation-model paradigm shifts from "stack bigger end-to-end" to "separate and reuse": ByteDance's SepLRM reorganizes multiple end-to-end large recommendation models into one RFM (recommendation foundation model) plus several vertical models, using the encoder's last-layer KV cache as the shared interface. The authors report roughly 100% capability transfer and have deployed it in two large-scale industrial scenarios. CascadeRec, deployed at an FP&A SaaS provider, uses a two-tier LLM reranking cascade: the router escalates only requests likely to benefit from extra inference to the expensive tier. Over a month-plus of production A/B, inference cost dropped about 10% with no notable quality regression. Both share the same judgment: sequence encoding should not be run separately by every recommendation stage.
Thread 2 — constrained decoding becomes a first-class citizen of generative retrieval: Uber Eats' LEGIT turns real-time eligibility and dedup rules into interval queries over two tries, cutting per-token overhead from 5× (a direct GPU implementation) to about 1.7×. RoutingTrie (Google) lifts Hit@100 from 67.5% to 96.1% in production replay at a cost of 1.1 ms latency. Taobao Shangou's TS-GEAR writes geographic reachability directly into the semantic-ID generation objective rather than filtering after the fact.
Thread 3 — evaluation standards for long-tail reachability and diversity are tightening: ReSID (Alibaba) is the first to break long-tail reachability into three stages — SID construction, decoding, and matching — and concludes that at equal code length, parallel SID consistently outperforms sequential SID on tail items. Kuaishou's KuaiRecV2 folds diversity into a unified rubric: a method only counts as progress if accuracy, diversity, and multi-task metrics all improve together.

Recommendation Foundation Models and Scaling Training

The core question for the foundation-model paradigm isn't "can it get bigger" — it's "can one sequence encoding be shared across multiple downstream tasks."
SepLRM (ByteDance) splits this question at the interface level. A conventional E2E large recommendation model encodes each behavior sequence once, then every stage and business line repeats that computation. SepLRM switches to an encoder-decoder structure: the encoder's last-layer state is exported as a candidate-agnostic KV cache, and downstream vertical models build candidate-conditioned queries on top of that cache, preserving the same query-KV interaction form as the E2E version. The key design choice: the KV cache retains fine-grained behavior information, which neither aggregated user embeddings nor candidate-dependent logits can do. To make it servable, the authors co-design cache compression, the training and serving pipeline, and a distributed KV cache store called LG-Store. It reaches SOTA on three large-scale datasets with roughly 100% transfer, and is deployed in two large-scale industrial scenarios. This line extends prior work that explicitly models cross-task structure in multi-task modeling — for example, Learned Cross-Task Relationships models pairwise relationships via the joint distribution of task labels, while SepLRM pushes "cross-task sharing" down to the cache layer.
CascadeRec (FP&A SaaS provider) tackles the cost problem of LLM reranking. The system runs on AWS EKS and GKE, serving content recommendations to multiple enterprise tenants. A Tier-1 lightweight model handles all requests; the router escalates only potentially beneficial requests to Tier-2 (self-consistency sampling, multi-path reasoning, or a larger model). The finding worth recording comes from segment analysis on public benchmarks: for users with sparse history, uniformly applying LLM reranking drops NDCG@10 by up to 30%, and the only statistically significant gain appears in the longest-history segment; escalating cold-start users to the expensive tier makes the loss even larger. This directly supports their segment-gated rollout strategy. Uneven multi-tenant load is the other driver of the cascade design. Related work: BracketRank uses tournament-style inference reranking to reach 26.56 nDCG@10 on BRIGHT, and LHF recovers 17-61% of the oracle coverage gap from the retriever-fusion angle — CascadeRec addresses the compute-allocation side of the same problem.
Feature-invariant scaling laws makes this observation: even under different feature changes, model performance as a function of training scale still follows a predictable scaling law. This lets you use small-scale experiments plus statistical inference to predict full-scale A/B effects with confidence intervals, filtering out neutral and degrading changes early. Validation covers three model families (Wukong, HSTU, LLM-based) and 40+ architecture, optimization, and data changes, and finds several changes on MovieLens-20M that further improve HSTU. Prior work on this line includes LUM's three-step paradigm that pushed parameters to 7 billion, and OneRec-V2's lazy decoder-only design that cut training resources by 90%. This week's paper moves scaling laws from "how to get bigger" to "how to run fewer experiments."
KuaiRecV2 (Kuaishou) is a benchmark, not a method. It contains million-scale multimodal item information, 30 days of long-term interaction logs, and a dynamic video candidate pool, packing three previously separately-evaluated challenges into one rubric: large-scale continual learning (long-term effectiveness under a dynamic item pool and distribution drift), compositional diversity (handling the accuracy-diversity tradeoff at both token and item level), and multi-task reward balancing (generative models can model sequence distributions but have poor controllability over multi-task requirements). The evaluation criterion is strict: a new method only counts as progress if it improves all metrics simultaneously. This criterion hits generative recommendation hard — work like EGR, which unifies item and user representations with a shared LLM and gets +2.91% conversion on Snap DPA, and R4T, which uses RL-compiled diffusion to cut fan-out latency by an order of magnitude, both need their diversity costs re-examined under the new standard.

Semantic IDs and Constrained Decoding for Generative Retrieval

The shared backdrop for this batch: generative retrieval in industry has moved past "can it work" into "is what it generates compliant." Compliance covers three things — whether IDs point to real, servable items; whether the full page of results satisfies business rules; and whether the value of the candidate set as a whole is optimized.
SID-LLM (Shopify) unifies search and recommendation into one thing: generating product semantic IDs. The catalog of billions of products changes continuously under strict latency constraints. The authors' approach: jointly pretrain on catalog data and behavior data so search and recommendation share representations and training objectives; then use RL to improve labeled search relevance and task-expert distillation to improve recommendation quality (relative to joint SFT, with no detected search quality loss); on the serving side, multi-token prediction paired with a lightweight ranker beats autoregressive baselines on retrieval quality while cutting latency. As a standalone search system, its quality matches Shopify's existing multi-model search stack and meets production latency; integrated into the existing search and recommendation stack, it improves both offline evaluation and online A/B. The historical reference on this line is DualGR, which uses a dual-branch long/short-term router and search-based SID decoding to gain +0.527% plays and +0.432% watch time online; SID-LLM's difference is fully merging the search and recommendation objectives.
TriGen handles trigger-induced recommendation (TIR): a single user click immediately triggers a follow-up feed, and the recommendation must trade off instant triggered intent against continuity with historical preferences. Existing generative recommendation mostly works under the user-induced (UIR) setting, treating the target item as a passive autoregressive continuation of a long history, with no mechanism to absorb this structural tension. TriGen fills the gap with three components: calibrated tail injection re-encodes local history, dynamic gating adaptively arbitrates between persistent taste and immediate impulse, and position-specific transformation maintains hierarchy-aware trigger guidance at each decoding step. It consistently outperforms baselines offline, improves revenue and GMV in online A/B across six country markets, and is fully deployed. On the serving side, compare xGR's approach — separate KV caches to unify prefill and decode, support early ranking termination — which achieves at least 3.49× throughput improvement under strict latency.
RoutingTrie (Google) addresses invalid generation under metadata constraints. A catalog trie only guarantees SID validity; business filtering happens after generation, which both wastes compute and may leave too few candidates. RoutingTrie brings metadata in before beam selection: it maintains a shared SID trie, builds a compact index for each metadata field (Dense Bitset or Sparse List, chosen by a storage-cost formula based on field distribution), and at decode time probes fields to determine whether the requested value appears beneath candidate edges, synthesizing a mask to guide the beam. In production replay, Hit@100 rises from 67.5% to 96.1% with only 1.1 ms added latency. It pairs with LEGIT (Uber Eats): LEGIT handles dynamic eligibility plus dedup plus the hierarchical grammar of "the item must belong to the store above," by turning depth-first-numbered trie leaves into interval queries so every generated page passes independent catalog validation. The engineering key is putting trie traversal on CPU and the model on GPU, avoiding kernel launches and CPU-GPU waits eating the gains; after comparing 36 implementations, per-token overhead is about 1.7× the unconstrained case. Gryphon fills in the other side — the mismatch between SID sequence likelihood and relevance objectives — by jointly training item-level scoring, gaining 3.7% Recall@1000 on an industrial music service.
TS-GEAR (Alibaba, Taobao Shangou) writes geographic constraints into the ID itself. In instant delivery, merchants outside the delivery range aren't ranked lower — they're simply invalid. Existing semantic IDs treat geography only as a soft signal, so reachability depends on hard truncation after generation. TS-GEAR's Geo-Semantic ID splits the merchant identifier into a geographic prefix and a semantic suffix, making reachability a prediction target; Query-Geo Target Attention aggregates historical behavior with the real-time query and location as dual targets, while encoding each behavior's own location into the attention key; GMPO rewrites the advantage inside GRPO using deduplicated group rewards and marginal advantage, avoiding an extra diversity auxiliary objective. Online it runs a 0.1B backbone serving 150 million queries per day, with ad impressions +3.8% and revenue +4.2%. The geography-aware path already has validation in local-services scenarios via OneLoc (GMV +21.016%, orders +17.891%); TS-GEAR's advance is turning reachability from post-filtering into a generation objective and replacing per-item scoring with set-level rewards.
RedMall (Alibaba) approaches generative e-commerce recommendation from another angle: SIDs are treated as one-dimensional sequences, ignoring hierarchical dependencies across layers; next-token prediction can't explicitly optimize session-level recommendation; and few RL alignment methods explicitly optimize the ranking position of positives. RedMall builds hierarchical SIDs via residual quantization over multimodal representations, adds a lightweight SID encoder to explicitly model inter-layer dependencies, then stacks sparse MoE, dual-gated attention, and multi-objective prediction to expand capacity. For training it designs an exposure-aware loss that treats exposed-but-unclicked as negative feedback, and uses GRPO plus a LambdaRank auxiliary loss to jointly constrain user preference and session-level ranking. It leads across the board on Recall and NDCG on four public datasets; online, ads gain ADVV +2.24% and CTR +0.59%, while recommendation gains GMV +1.08% and DAB +0.95%. The exposure-aware loss shares its origin with DualGR's ENTP-Loss; RedMall ties it to a listwise ranking objective.
ReSID (Alibaba) deserves separate treatment because it breaks the long-tail problem apart. The authors measure reachability separately along SID construction, decoding, and matching, yielding two conclusions: at equal code length, parallel SID consistently outperforms sequential SID on tail items; and even when full generation fails, the decoded sequence still carries locally consistent information that can recover relevant tail items via partial matching — but this recovery is capped by upstream search: once beam search prunes all prefixes completable within a fixed mismatch tolerance, recovery fails. ReSID therefore builds on long parallel SIDs, pairs them with popularity-debiased and user-semantic-aware beam search to preserve promising paths, then recovers items from surviving tokens via partial SID matching. The result is SOTA long-tail retrieval while staying competitive on overall accuracy, with lower latency and higher throughput under narrow beams, and it's deployed serving main traffic. It complements Gwhere, which uses contrastive residual quantization in POI scenarios to align text, visual, spatial, and collaborative signals — both trade more structured IDs for reachability.

Embedding Retrieval and Reranking Model Optimization

The shared proposition here: how can dual-mode or multi-stage embeddings capture the fine-grained semantics of generative embeddings at unchanged inference cost?
AntEmbed puts discriminative and generative embeddings into the same vector space. Discriminative embeddings encode raw multimodal input efficiently but have weak fine-grained semantics; generative embeddings capture query intent and relevance evidence through reasoning but are computationally expensive. AntEmbed first constructs retrieval-oriented reasoning content (characterizing query intent and confusion-aware semantic evidence), treats the raw candidate and the reasoning-augmented candidate as two semantically equivalent views, and aligns them with multi-positive contrastive learning plus a single embedding token; then cross-mode self-distillation transfers the fine-grained information of the reasoning-augmented representation to the raw representation via candidate similarity distribution distillation and query representation alignment. This way the discriminative mode needs no extra inference online, and both modes share one candidate index. AntEmbed-8B reaches 81.5 overall on MMEB-v2 with reasoning-augmented embeddings, 3.1 points above the strongest baseline. For the latest progress on this line, see HIVE, which uses an LLM for query synthesis and verification to lift MM-BRIGHT nDCG@10 from 32.2 to 41.7.
Yuanbao-Reranker is a family of general dense rerankers at 0.6B / 4B / 8B, built on Qwen3 backbones. On the data side it uses answer-first and query-type-guided document-grounded synthesis, a three-stage filtering funnel (positive support verification, semantic deduplication, source and length quota balancing), then filters out potential false negatives by teacher-score margin and samples negatives by Hard / Medium / Moderate tiers, yielding 2.75M 1:3 ranking groups. Training jointly optimizes pointwise relevance cross-entropy and a bounded minimum-margin hinge objective, acting on both positive-negative and negative-negative pairs — transferring fine-grained teacher preferences without over-constraining predictions that are already well-separated. The result leads at every tier: 0.6B reaches 70.96 average, beating 1B-4B effective-parameter baselines; 4B reaches 73.63, beating the 8B teacher (72.68); 8B reaches SOTA 74.21, with gains across five benchmarks — CMTEB-R, MMTEB-R, BEIR, MLDR, MTEB-Code — over first-stage retrievers. The data-side approach is comparable to ARHN's relabeling of hard negatives with open-source LLMs and STAR's token-level interaction learning to mitigate semantic shortcut bias.
e-means is the only one in this batch that directly changes the clustering algorithm. It replaces Lloyd iterations with exponential moving average plus mini-batch updates, and derives scaling laws for how the decay rate and batch size affect final loss — so on multiple problem sizes and embedding datasets these two hyperparameters can be set directly by formula, no tuning sweep needed. Across a wide range of embedding datasets, it reaches the loss Lloyd k-means needs 25 rounds to hit in 1.4-2.6 rounds, corresponding to 10-45× GPU wall-clock speedup; on DINO10B (10 billion DINO embeddings) it partitions into 16 million clusters in just 2.5 epochs, with lower loss than Lloyd at 25 rounds. The retrieval-side use is direct — TACHIOM uses token-distribution-aware centroid assignment to cluster at million-centroid scale, achieving 247× clustering speedup and 9.8× retrieval speedup on MS-MARCOv1, and centroid scale is exactly where clustering cost bottlenecks.
Two more are systems-engineering oriented. DcLBC studies document-local decision boundaries in late chunking: shared document context makes semantically similar negatives within the same document hard to distinguish. Starting from InSeNT's intra-document gradients, the authors prove that conditional negative assignment entropy decreases monotonically with the scaling factor at fixed logits, and on that basis apply adaptive margin scaling plus chunk-aware routing, improving contextual chunk retrieval on ConTEB and NanoBEIR while keeping passage-level retrieval quality competitive. Hybrid Product Search puts search directly on an Iceberg data lake: it extends Apache Pinot, does BM25 scoring over Lucene text indexes, adds corpus-level term statistics so scores are comparable across files, and implements reciprocal rank fusion in pure SQL. On Amazon ESCI (1.2M US products, 22,458 test queries) it reaches nDCG@10 of 0.361, on par with exact retrieval; on a 277M-product, 0.8 TB lake, a single node with 16 vCPU answers hybrid retrieval combining brand-scope vectors, BM25, and client-side fusion at p50 of 6.5-9 ms, searchable 4.3 s after submission. Compare Adaptive Re-Ranking's query routing — both are squeezing the latency budget of hybrid retrieval.

Ad Bidding and Marketing Decision Optimization

Three papers with online results, different paths, all making the coupling between "prediction" and "decision" real.
TUGB (JD) targets three pain points of Decision-Transformer-style auto-bidding policies: insufficient exploration in static offline data, fragility to dynamic environment drift, and reliance on heuristic return-to-go specifications. TUGB runs two stages: pretrain a generative bidding model, then RL-finetune with an ensemble-based uncertainty-aware environment model that quantifies epistemic uncertainty to penalize high-risk trajectories, while evolving the policy from RTG-conditioned generation to direct bid optimization. Fully deployed on JD search ads, with cost +2.32% and target cost +3.57%, and subsequently adapted to mechanism upgrades and JD food-delivery ads. Prior work in this series: PRO-Bid handles constraint satisfaction with Pareto representations plus counterfactual regret optimization; GUIDE gets ad GMV +4.10% and ROI +3.52% on Taobao; and DRIVE combines distribution modeling with retrieval-augmented candidate generation.
TPO (TrueGradient AI) builds promotion optimization into an end-to-end chain: analogous SKUs, point elasticity, elasticity curves, price recommendation, promotion lift, net lift. Elasticity estimation has two stages — post-promotion, measured against rolling regular prices using directional markdown events; during promotion, a causal estimator residualizes trend, seasonality, and promotion, uses a Tobit likelihood to handle stockout censoring, and self-updates starting from the first clean promotion week. Products with weak evidence borrow point elasticity and curvature from analogous SKUs, with analogy relationships selected by a cross-validated ridge model over category, brand, subclass, price line, and description. The optimizer separates own-price lift from substitution and post-promotion fallback, and adjusts for inventory and trade funding. On a 39-SKU consumer-goods brand with average elasticity -3.06, a four-week promotion experiment yields sales +20.6%, revenue +6.9%, and flat margin; analogy pooling halves elasticity error relative to hierarchical averaging.
SkinAdvisor (Tencent) addresses production decisions for game skins: which to make, when to release, how to price. The core model, SkinCast, is a multimodal Transformer coupling four encoders — a pretrained Informer for player preference time series, a pretrained Player-Hero-Skin graph embedding for relational priors, and two lightweight Transformers for skin attributes and release context — fused through a shared backbone with local-plus-global day-x positional encoding, decoding a 180-day daily sales trajectory in a single non-autoregressive pass through N Informer layers. Predictions feed a Select→Fill→Infer→Sales→ROI→Ranking serving pipeline that ranks candidate skins by predicted ROI. The system has completed a full release cycle: offline RMSE is about 20% lower than the strongest baseline (Informer) and about 60% lower than classical time-series methods, with p99 latency under 8 ms on 16-core CPU and peak throughput above 1100 rps. The authors record two deployment lessons: for production decisions, order-of-magnitude accuracy is enough; and pretrained embeddings leak temporal information, requiring season-based splits. On the sequence-modeling side this belongs to the same problem domain as DSIN's hierarchical intra-/inter-session interest modeling and RAT's retrieval-augmented feature representations — except the output here is a long-range sales trajectory, not CTR.
DR-CATE-CBwK handles promotion target selection under budget constraints: a web platform picks costly interventions for users under a global budget to maximize incremental outcomes. The authors note that while an action-independent outcome baseline cancels out in the treatment contrast, modeling a nonlinear baseline adds statistical complexity — so they instead learn the treatment contrast directly with an action-centered causal bandit. The method combines offline doubly-robust learning for warm start, an online action-centered estimator for sequential contrastive learning, and a primal-dual controller with an affordable-action rule to guarantee the effective cost budget pathwise. Under the stationary i.i.d. setting with linear CATE, they give a high-probability Õ(Kd√T) regret bound. On a semi-synthetic job-market benchmark, a practical Thompson-restricted variant has the lowest regret among feasible-budget methods, about 20% below the strongest reward-based baseline, and never exceeds budget. For online-bandit improvements in the same direction, see Calibration-Gated LLM Pseudo-Observations (19% cumulative regret reduction on MIND-small) and C3 Thompson sampling.

Directions to Watch

Generalizing constrained decoding. Three papers this week (LEGIT, RoutingTrie, TS-GEAR) tackle the same thing from three angles: moving business rules forward into the decoding process instead of truncating after generation. The evidence: RoutingTrie lifts production-replay Hit@100 from 67.5% to 96.1% at only +1.1 ms latency, and LEGIT cuts per-token overhead from 5× to 1.7×. The gap: each adapts to one constraint form (hierarchical grammar, metadata fields, geographic reachability), with no unified abstraction for constraint expression. If "additive constraints" were abstracted into composable operators over index structures, this could be reused across inventory, timeliness, compliance, and more dimensions. The drivers are teams with strong physical constraints — Uber, Google, and Alibaba International/Shangou.
A shared interface for sequence encoding. SepLRM uses the native KV-cache interface for cross-stage transfer, reporting roughly 100% transfer; CascadeRec uses a router to decide which requests deserve the more expensive inference tier; feature-invariant scaling laws use scaling laws to decide which changes deserve full-scale experiments. All three share one assumption: compute should be allocated by marginal return, not statically partitioned by pipeline stage. The clearest near-term application is multiple business lines sharing one sequence encoding — ByteDance has already shown transfer across model iterations, recommendation stages, and business scenarios. The open question is the operational cost of cache consistency and version evolution; SepLRM handles part of it with LG-Store but hasn't published cache miss rates or backfill latency numbers.
Joint evaluation of long tail and diversity. ReSID demonstrates that long-tail reachability can be decomposed and localized (parallel SID beats sequential SID; recovery is capped by beam pruning), while KuaiRecV2 requires all metrics to improve together. Together these will change the optimization objective for generative retrieval: you can no longer hide tail collapse behind overall Recall, nor trade diversity for accuracy. Kuaishou's 30-day dynamic candidate pool benchmark is currently the closest public resource to this standard. For teams working on semantic IDs, one directly actionable check is to switch SID construction from sequential to parallel and measure the change in tail-item Recall.

Paper Roundup

Recommendation Foundation Models and Scaling Training
SepLRM — ByteDance proposes a recommendation foundation model paradigm, reorganizing multiple E2E large recommendation models into one RFM plus several vertical models, with the encoder's last-layer KV cache as the shared interface; roughly 100% transfer, deployed in two large-scale industrial scenarios.
CascadeRec — An FP&A SaaS provider proposes a two-tier LLM reranking cascade with adaptive routing; over a month-plus of production A/B, inference cost drops about 10%; finds that uniform LLM reranking drops NDCG@10 by up to 30% for users with sparse history.
Feature-invariant scaling laws — Proposes feature-invariant scaling laws, using small-scale experiments plus statistical inference to predict full-scale A/B effects; covers 40+ changes across three model families and improves HSTU on MovieLens-20M.
KuaiRecV2 — Kuaishou releases an industrial-grade benchmark combining continual learning, compositional diversity, and multi-task balancing, with million-scale multimodal items and 30 days of interactions; requires all metrics to improve together to count.
Semantic IDs and Constrained Decoding for Generative Retrieval
SID-LLM — Shopify unifies search and recommendation into generating product semantic IDs, with joint pretraining plus RL plus expert distillation plus multi-token prediction; as a standalone search system its quality matches the existing stack and meets production latency, and it improves online A/B after integration.
TriGen — Proposes a generative paradigm for trigger-induced recommendation with calibrated tail injection, dynamic gating, and position-specific transformation; improves revenue and GMV in online A/B across six country markets and is fully deployed.
ReSID — Alibaba is the first to analyze long-tail reachability along SID construction, decoding, and matching; uses long parallel SIDs plus debiased beam plus partial-match recovery to reach SOTA long-tail retrieval with lower latency under narrow beams, deployed on main traffic.
RoutingTrie — Google probes metadata indexes before beam selection to generate compositional masks; production replay lifts Hit@100 from 67.5% to 96.1% at +1.1 ms latency.
TS-GEAR — Alibaba Taobao Shangou proposes Geo-Semantic ID, Query-Geo Target Attention, and GMPO; a 0.1B model serves 150 million queries daily, with ad impressions +3.8% and revenue +4.2%.
RedMall — Alibaba proposes hierarchical SID plus a lightweight SID encoder plus sparse MoE plus exposure-aware loss plus GRPO+LambdaRank; online ads gain ADVV +2.24% and CTR +0.59%, recommendation gains GMV +1.08% and DAB +0.95%.
LEGIT — Uber Eats turns real-time eligibility and dedup into interval queries over two tries; a heterogeneous CPU-trie + GPU-model implementation cuts per-token overhead from 5× to about 1.7×.
SteerGR — Uses sparse autoencoders for concept-level control in generative recommendation; a three-stage approach amplifies or suppresses target concepts without retraining, validated on two industrial models and real game data.
PHiT — Step-level diagnosis finds that only the history-analysis step contributes positively in recommendation CoT; constructs semantic supervision targets directly from observed history, with all thought tokens in a single causal prefill, no teacher CoT or recurrent state feedback needed.
Embedding Retrieval and Reranking Model Optimization
AntEmbed — Proposes dual-mode multimodal embeddings, aligning discriminative and generative representations with multi-positive contrastive learning and transferring fine-grained semantics to the raw representation via cross-mode self-distillation; AntEmbed-8B reaches 81.5 on MMEB-v2, 3.1 points above the strongest baseline.
Yuanbao-Reranker — Proposes 0.6B/4B/8B dense rerankers on Qwen3 backbones, building 2.75M ranking groups via document-grounded synthesis plus a three-stage filter, jointly optimizing pointwise cross-entropy and a bounded minimum-margin hinge; the 8B version averages 74.21 across five benchmarks for SOTA.
e-means — Replaces Lloyd iterations with EMA plus mini-batch updates and derives scaling laws for decay rate and batch size; needs only 1.4-2.6 epochs for equivalent loss, 10-45× GPU speedup, and partitions DINO10B into 16 million clusters in 2.5 rounds.
DcLBC — Analyzes document-local decision boundaries in late chunking, proposing adaptive margin scaling and chunk-aware routing; improves contextual chunk retrieval on ConTEB and NanoBEIR.
Hybrid Product Search — Extends Apache Pinot to do BM25 scoring, vector retrieval, and in-SQL RRF fusion directly on an Iceberg data lake; reaches nDCG@10 of 0.361 on Amazon ESCI, with p50 of 6.5-9 ms on a single node over 277M products.
Ad Bidding and Marketing Decision Optimization
TUGB — JD proposes two-stage generative auto-bidding, pretraining a Decision Transformer then RL-finetuning with an ensemble uncertainty environment model; fully deployed on search ads with cost +2.32% and target cost +3.57%, and adapted to mechanism upgrades and food-delivery ads.
TPO — TrueGradient AI builds an end-to-end promotion optimization chain with two-stage elasticity estimation plus analogous-SKU pooling; a four-week experiment on a 39-SKU brand yields sales +20.6%, revenue +6.9%, and flat margin, with elasticity error halved relative to hierarchical averaging.
SkinAdvisor — Tencent proposes a multimodal Transformer predicting 180-day daily skin sales, with four-encoder fusion and non-autoregressive decoding; offline RMSE is about 20% lower than Informer and about 60% lower than classical methods, and it has completed a full release cycle.
DR-CATE-CBwK — Proposes an action-centered causal bandit that learns treatment contrasts directly, with doubly-robust warm start and primal-dual budget control; on a semi-synthetic job market it cuts regret about 20% relative to the strongest reward-based baseline.
  • Recommendation Systems
  • Weekly
  • Papers
  • AI Weekly 2026-W46RecSys Daily - 2026-11-14
    Loading...