RecSys Weekly 2026-W33 | Recsys Frontier

type

Post

status

Published

date

Aug 15, 2026 05:33

slug

rec-weekly-en-2026-W33

summary

This week's recommendation systems research is led by industrial deployment papers. Netflix, Yandex, Meta, LinkedIn, Kuaishou, Alibaba, and ByteDance each published online A/B results — a density rarely seen in a single year. If there's one trend to watch: generative recommendation is moving from lab validation to the systems engineering phase of "replacing the entire production cascade with a single model." Thread 1 (Two directions in generative recommendation): Yandex Music's Sona replaces a full cascade of 15+ candidate generators plus pre-ranking/ranking with a single generative model — Active Users +4.53%. Netflix's GenRec takes a different path — instead of replacing the cascade, it uses an LLM ranker as the final ranking layer, achieving statistically significant gains over the production ranker in A/B tests. Two routes validated in parallel within the same week is the strongest signal in this week's papers. Kuaishou's PushDualGen tackles explainability in generative recommendation: after generating SIDs, it attaches a skippable copy as an explanation — effective play rate +8.50%, dissatisfaction rate -37.70%. Thread 2 (Causal inference moves from ideas to deployment): LinkedIn's decision-centric causal optimization framework delivers +7.20% long-term value on Feed marketing traffic, unifying causal effect estimation, Bayesian bandits, and linear programming allocation under a single objective. Meta's MARCO operates at a finer grain — using click types as free behavioral labels to decompose click intent, conversion per click +2.80%. The shared takeaway: causal recommendation is no longer just a debiasing technique in papers — it's a deployable source of revenue in production systems. Thread 3 (Systems engineering for multi-task and full-funnel optimization): Alibaba's IntHQ deploys on Amap, addressing three collapse problems in multi-task learning for generative recommendation — UVCTR +1.60%. Alibaba's DREAM stacks an agent-based meta-control layer atop the e

Weekly Overview

Thread 1 (Two directions in generative recommendation): Yandex Music's Sona replaces a full cascade of 15+ candidate generators plus pre-ranking/ranking with a single generative model — Active Users +4.53%. Netflix's GenRec takes a different path — instead of replacing the cascade, it uses an LLM ranker as the final ranking layer, achieving statistically significant gains over the production ranker in A/B tests. Two routes validated in parallel within the same week is the strongest signal in this week's papers. Kuaishou's PushDualGen tackles explainability in generative recommendation: after generating SIDs, it attaches a skippable copy as an explanation — effective play rate +8.50%, dissatisfaction rate -37.70%.

Thread 2 (Causal inference moves from ideas to deployment): LinkedIn's decision-centric causal optimization framework delivers +7.20% long-term value on Feed marketing traffic, unifying causal effect estimation, Bayesian bandits, and linear programming allocation under a single objective. Meta's MARCO operates at a finer grain — using click types as free behavioral labels to decompose click intent, conversion per click +2.80%. The shared takeaway: causal recommendation is no longer just a debiasing technique in papers — it's a deployable source of revenue in production systems.

Thread 3 (Systems engineering for multi-task and full-funnel optimization): Alibaba's IntHQ deploys on Amap, addressing three collapse problems in multi-task learning for generative recommendation — UVCTR +1.60%. Alibaba's DREAM stacks an agent-based meta-control layer atop the existing pipeline on Taobao's homepage — IPV +2.06%. ByteDance's TM20K extends e-commerce behavior sequences to 20K length — ADSS +1.036% with only 5.6% added serving latency.

Generative Recommendation & LLM Recommendation

The discussion focus in generative recommendation shifted this week from "can it run" to "how to make it stable, affordable, and explainable." Three industrial deployment papers offer three different answers.

Sona (Yandex) — Sona's core design is "one model does everything." An encoder maps user behavior sequences into hidden states; an autoregressive decoder and a ranking module share this user representation, jointly updated via next-token-prediction and distillation objectives. This design differs from HiGR's hierarchical planning approach — Sona has no explicit list-level planning; instead, generation and ranking couple through the same user state. The result: a single deployed model fully replaces the multi-stage cascade previously powered by large Transformer models like Argus. Online metrics: Active Users +4.53%, Total Listening Time +6.30%, Likes +11.42%. Worth noting: the Active Users gain is 2.35x that of Argus, the previous strongest model — the payoff from full-funnel generative replacement isn't incremental; it's a difference in magnitude.

GenRec (Netflix) — Unlike Sona, GenRec keeps the existing pipeline and only replaces the final ranking layer. Two-phase framework: Phase 1 adapts an open-source LLM to Netflix's content domain; Phase 2 post-trains on recommendation ranking data. The paper's most important contribution is in serving design — prefill-only inference. Traditional LLM generative ranking requires decode steps; GenRec avoids autoregressive decoding latency by doing prefill only. This follows the staged computation idea from xGR, but GenRec goes further by confining "generation" to the context engineering level — input is verbalized user history, output is a ranking score rather than text. The paper doesn't disclose specific gains, but states clearly that "statistically significant gains were achieved with substantially fewer Phase-2 labeled samples." That data efficiency point connects to HyMiRec's design of compressing user history with a lightweight recommender to reduce LLM input costs.

The two Alibaba papers are best read together. MetaStrategy (Taobao homepage "Guess You Like") takes an unconventional angle in generative ranking — instead of generating item sequences, it generates an executable JSON ranking strategy. The strategy controls objective weights, content preferences, experience constraints, and position policies; after passing a validator, it instantiates an isolated Generator that competes with existing models. A 4B teacher distills to a 0.8B student, with LLM inference kept outside synchronous ranking. Seven-day online A/B: click PV +2.11%, IPV +3.12%, GMV +2.83%. This contrasts with IntHQ (Amap) — IntHQ solves the internal multi-task structure problem in generative recommendation (three collapses: source collapse, relation collapse, hierarchy collapse), while MetaStrategy bypasses internal structure entirely, having the LLM output strategy parameters to steer existing models. The former achieves UVCTR +1.60%, the latter GMV +2.83% — both work in their respective scenarios, but the philosophies are completely different.

PushDualGen (Kuaishou) addresses another pain point in generative recommendation deployment: unexplainability. OneRec-Thinking uses CoT to make SID generation traceable, but inference cost is prohibitive. PushDualGen's approach: generate the SID first, then generate a copy as an explanation — the copy is skippable, so no inference cost is sacrificed. Online: effective play rate +8.50%, dissatisfaction rate -37.70%. This "skippable explanation" design differs from EviSnap's facet cards — EviSnap treats explanations as one of the core recommendation signals; PushDualGen treats explanations as an auxiliary communication layer. For industrial systems, the latter is more pragmatic.

Several academic advances are also worth noting. HCGRec addresses a very specific problem in RL post-training for semantic-ID generative recommendation: when an early semantic token enters the wrong branch, subsequent rollouts can almost never reach the ground-truth item, leaving group-relative optimization with all-zero rewards and no valid gradients. The method injects a minimal target-prefix hint for hard samples, letting the model generate the remainder under the correct semantic branch. Zero-advantage samples drop from 70%+ to below 20%. The root problem relates to the structural risk of SID hierarchy that has existed since the TIGER era — HCGRec is the first to systematically address it. FSGR exposes a neglected fairness issue in SID — high-frequency tokens are systematically over-predicted. Gini fairness improves by over 20%, but the paper's more important contribution is identifying the problem: codebook imbalance in SID propagates to downstream recommendation fairness. Centroid Initialization is a minimal intervention — initializing SID token embeddings with semantic-space centroids instead of random Gaussians yields Recall@5 +16% under pure SFT and +60% for cold start. These academic works point in the same direction: SID infrastructure issues (initialization, fairness, sparse RL signals) are becoming core topics in generative recommendation research.

Cross-domain generative recommendation also saw a new industrial framework this week. HD-Rec (Kuaishou with City University of Hong Kong) uses hierarchical domain-aware quantizers to resolve semantic granularity mismatch in unified cross-domain models: coarse-grained codebooks are globally shared, fine-grained codebooks route adaptively, paired with domain-adaptive sparse MoE and cross-granularity routing consistency constraints. This extends the SID idea to cross-domain settings — prior methods mostly used globally shared representations or lightweight domain adaptation, which lack modeling capacity for heterogeneous patterns. Kuaishou's investment in generative recommendation runs deeper than outsiders typically recognize.

Industrial Ranking & Full-Funnel Optimization

This week's industrial ranking papers share a common trait: no major architectural changes — the gains come from objective functions, training procedures, and data utilization.

MARCO (Meta) starts from a simple observation that the industry has long overlooked: different click types on the same ad have up to 4x differences in actual conversion rate. Standard CVR models treat all clicks as the same event, producing seemingly perfect overall calibration but biased per-intent calibration. MARCO's approach: use click types as free behavioral labels, train per-intent CVR heads, and combine intent-specific CVR estimates at serving time according to the predicted intent distribution. The paper theoretically proves that decomposition doesn't increase overall risk. Online results: per-intent calibration reaches ~100%, conversion per click +2.80%, topline metrics +0.98%. The paper's value for ad ranking practice lies in its theoretical rigor — it's not simply "multi-task learning plus a trick," but derives the necessity of decomposition from the source of calibration bias. This follows the spirit of ESCM2's approach to ESMM's inherent bias, but MARCO uses observed behavioral labels rather than counterfactual regularization.

From Prediction to Incrementality (LinkedIn) tackles a more fundamental problem: when the business objective is incremental impact, prediction-score-based allocation systems waste resources on users who "would have acted anyway." The solution is a decision-centric causal optimization framework — a Transformer-backbone causal network for individual treatment effect estimation, a Bayesian neural bandit layer for uncertainty-aware exploration, and a dual-based linear programming layer for allocation under global constraints. The end-to-end policy achieves +7.20% long-term value on LinkedIn Feed marketing traffic. The paper's value is unifying three technology stacks that normally belong to different fields (causal inference, bandits, LP) under a single deployable objective. Compared to uplift modeling's similar constraint handling for causal retrieval optimization at Pinterest, LinkedIn's approach is more explicit at the allocation layer.

The two Taobao homepage papers can be read as a whole. DREAM stacks an agent-based meta-control layer atop the existing pipeline — a three-layer Intent Engine fuses on-device signals into L0/L1/L2 intent representations; a Meta Engine performs M1-M2-M3 hierarchical reasoning (intent summarization, strategy planning, parameter translation). A Reward Dual Loop explores the policy space via offline simulation and calibrates results with online feedback. Reranking control alone delivers IPV +2.06%; extended to full ranking, IPV +2.71% and GMV +1.31%. MetaStrategy was discussed earlier — generating executable strategies rather than item sequences. The common thread: neither replaces the core ranking model; both introduce a "strategy generation" agent layer above it. For a system at Taobao's scale, replacing the core model is too costly — meta-control is the more realistic path.

TM20K (ByteDance) solves the efficiency problem of ultra-long behavior sequence modeling. Prior methods like SIM and ETA use search or clustering for sequence compression, sacrificing fine-grained information. TM20K's approach: the teacher trains on the full 20K tokens; the student compresses sequences via token merging; distillation compensates for information loss. Online: ADSS +1.036%, latency +5.6% only. The "teacher keeps all tokens, student merges efficiently" idea sidesteps the inherent trade-offs of compression methods — distillation covers the compression loss.

DrEM addresses a hidden problem in ensemble ranking: pxtrs output by upstream multi-task models serve simultaneously as features and supervision signals, so noise propagates at both ends. A risk-denoising robust loss corrects empirical risk; a preference-preserving ranking consistency regularizer improves output stability. The paper also provides theoretical analysis — approximate distributions of prediction noise and the superiority of the robust loss under flip-probability estimation error. This contrasts with AdaTT's inter-task fusion approach — AdaTT handles relationships between tasks; DrEM handles propagation of upstream prediction noise.

NAVER WEBTOON's Progressive Alignment proposes a three-stage LP-FFT-RFT post-training framework: first freeze pretrained representations and do linear probing to stabilize randomly initialized downstream heads, then full fine-tuning for task adaptation, and finally RL fine-tuning with a learned reward model aligned to business metrics. The key design: the policy trains on dense implicit feedback while the reward model is supervised by business metrics — this avoids the variance problem of using sparse business objectives directly as training signals. Online A/B shows effective gains, though the paper doesn't disclose specific numbers.

MISO offers a non-model methodological perspective — using model internal states (parameters, activations, gradients, normalization statistics) to guide decisions about "which component to scale, replace, or retire," reducing trial-and-error cost. In an ad ranking case, this workflow requires fewer validation runs than both expert-driven and black-box search approaches.

Sequence & Multimodal Representation Learning

This week's sequence and multimodal work skews academic, but several papers raise new questions worth attention.

PRISM identifies a systemic problem in Transformer-based sequential recommendation: dot-product attention is inherently biased toward similar items, causing heterogeneous relations to be systematically ignored. The method uses multi-view attention calibration — an Affinity View refines homogeneous relations; a Contrast View exposes heterogeneous relations suppressed by similarity bias. Average Hit@10 and NDCG@10 improvements of ~5-10% across 7 datasets. This problem has existed since the SASRec era, but no one has systematically analyzed and solved it until now.

NTCF revisits the propagation depth problem in graph collaborative filtering: using a uniform number of propagation layers for all nodes ignores the fact that peripheral nodes over-smooth quickly while hub nodes under-explore beyond their immediate neighborhoods. A local degree imbalance score serves as a discrete Ricci curvature proxy, assigning adaptive propagation depth per node. The paper theoretically proves NTCF strictly generalizes NGCF, and curvature-aware scheduling preserves more discriminative information at depth.

TimeRoute handles temporal dynamics in multimodal fusion: the utility of item modalities drifts over time — chocolate consumption shifts from text-guided to vision-guided around Valentine's Day. Time-aware modality routing maps a personalized modality distribution per user; a diffusion graph reconstructor uses FiLM conditioning to suppress stale modality edges. Up to 9.8% Recall@K improvement on TikTok data. This "modality temporality" problem has practical relevance in e-commerce scenarios with seasonal and holiday effects.

SMD handles modality missingness in multimodal sequential recommendation: during training, modality streams of the entire behavior history are independently erased with probability p, forcing the model not to rely on any single modality. At 95% missing rate, HR@10 retention improves from 22% to 61%. It's a four-line code change, but highly practical for production catalogs where images or text are frequently missing.

DualSpectralCF incorporates explicit negative feedback (1-star reviews, dislike buttons) into training-free spectral collaborative filtering. Signed input signals encode explicit dislikes; a signed item-item operator mixes "liked together" and "disliked together" similarities. Recall@20 improves up to +32.6% (with tuning); +1.9% to +16.0% at default parameters. Cold-start users see the largest gains (+29.2%). The idea is simple but practical — most recommendation systems have negative feedback signals that go underutilized.

VisGate (IKEA with KTH) uses a gating mechanism for adaptive multimodal fusion, treating visual utility as a latent variable rather than a fixed attribute. The learned gate doubles as a measurement tool: visual utility rises with interaction sparsity and correlates with visual distinctiveness.

Low-Interaction-Rank is a theoretical framework: defining a low interaction-rank function class to unify the theoretical foundations of two-encoder architectures. Sample complexity is the sum (not product) of encoder complexities; whitening can fix interaction patterns up to permutation and sign. This is useful reference for understanding the expressiveness boundaries of two-tower models.

PreGress proposes a ranking-native graph pretraining framework — multi-task pretraining with degree centrality prediction and attribute reconstruction, plus lightweight prompt modules for downstream ranking task adaptation. This is one of the few graph pretraining works directly targeting ranking, validated on Yelp2018 and MovieLens-100K.

Causal Inference & Offline Evaluation

This week's causal inference work is dominated by diagnostic studies, all answering the question: "when can't we trust offline results?"

Fast A/B/n Testing presents an elegant theoretical result: tree-coupled feedback sharing reduces expected reward queries from JT to T+o(T) while preserving exact per-policy trajectory distributions. The path identity N(T)=T+ΣD_{e,t} attributes the extra queries to accumulated mismatch on tree edges. This means comparing multiple adaptive policies no longer requires each policy to run full trajectories independently. The theoretical contribution is clear; experiments cover reward model evaluation and language model evaluation.

When Offline Evaluation Misleads comes from a deployed push notification system (Thumbtack), and the conclusion is striking for practitioners: detecting a mean effect and learning to act on it are two different things. In a deployment with five arms and one split, learned routing at every granularity performs no better than random; a matched-moment noise placebo reproduces 100% of the oracle's surface-level gains — the apparent "learnable structure" is just order statistics of noise. The authors propose an actionable diagnostic protocol: screening along two dimensions — alignment (does optimizing the reward move the north star?) and learnability (can a bandit identify the reward-optimal policy?). The protocol's value is decomposing "why offline evaluation lies" into verifiable steps.

TTT-Embed handles test-time optimization for dense retrieval: distilling ranking rewards into a lightweight vector in the embedding space of a frozen model. No weight access, no index modification, only scalar ranking scores needed. Up to +8.36 nDCG@10 across 15 MTEB retrieval tasks, and it learns a reusable test-time state — the trade-off among global/task/query sharing scopes is controlled by a single scope parameter. As the reward budget expands, the optimal sharing scope dynamically shifts from global toward query.

DR-CVR revisits CVR causal effect estimation from semiparametric theory, proposing a doubly robust estimator for chained structures (click → conversion). The theoretical contribution outweighs the method itself: numerical stability is validated on synthetic and real data, with an important observation — naively combining loss debiasing with standard causal estimators fails; CVR-style objectives require specially designed estimators.

Directions to Watch

Full-funnel validation of generative recommendation is accelerating. Five industrial deployment papers this week come from different companies with different routes: Yandex fully replaces the cascade, Netflix replaces the final ranking layer, Kuaishou adds explainable copy generation, Alibaba (Amap) optimizes multi-task structure, Alibaba (Taobao) does strategy generation. The common thread: all have clear online metrics and engineering solutions for QPS/cost constraints. This density indicates generative recommendation has passed the proof-of-concept stage and entered an engineering race. Worth tracking: the capacity-vs-cost trade-off curve — Sona doesn't rely on handcrafted features but needs large-model distillation; PushDualGen trades skippable copy for explainability; MetaStrategy distills a 4B teacher to a 0.8B student.

The "prediction vs. incrementality" paradigm shift. LinkedIn moves from prediction scores to causal effect estimation; Meta moves from unified CVR to intent decomposition. The driver is the business objective itself — when the goal is incrementality rather than accuracy, prediction models are systematically misaligned. Worth watching: whether this thinking propagates to content recommendation — should Douyin/Kuaishou watch-time optimization shift from "predicting watch time" to "incremental watch time"? The technical challenge is variance control in incrementality estimation; LinkedIn's Bayesian bandit layer experience may be reusable.

Evaluation methodology enters the "debunking illusions" phase. This week's group recommendation tie-breaking audit, LLM recommendation confidence calibration audit, offline evaluation diagnostic protocol, and reward-SNR lower bound proof — these works independently question the validity of existing evaluation pipelines. The practical implication for industry: many reported "improvements" may be artifacts of the evaluation protocol. Recommend running a similar audit on your own offline evaluation pipeline, particularly watching for fake gains created by the combination of score compression during training and deterministic tie-breaking during evaluation.

Paper Roundup

Generative Recommendation

GenRec — Netflix proposes an LLM-backed ranker: Phase 1 adaptation + Phase 2 post-training, prefill-only inference cuts cost; statistically significant A/B gains over the production ranker.

Sona — Yandex Music replaces a cascade of 15+ candidate generators plus ranking with a single generative model; Active Users +4.53%, listening time +6.30%, Likes +11.42%.

PushDualGen — Kuaishou proposes a lightweight SID generator with skippable explanatory copy; effective play rate +8.50%, dissatisfaction rate -37.70%.

MetaStrategy — Alibaba proposes executable JSON ranking strategy generation for Taobao homepage "Guess You Like"; click PV +2.11%, IPV +3.12%, GMV +2.83%.

IntHQ — Alibaba proposes dual-stream decoupling + task interaction + hierarchical querying for multi-task generative recommendation on Amap; UVCTR +1.60%.

HD-Rec — Kuaishou with City University of Hong Kong proposes hierarchical domain-aware quantizers + domain-adaptive sparse MoE for cross-domain generative recommendation.

HCGRec — SJTU and Meituan propose hint-conditioned generative recommendation to recover zero-reward sample signals; zero-advantage samples drop from 70%+ to below 20%.

FSGR — First to reveal token frequency bias in SID generative recommendation; Gini fairness improves by over 20%.

Centroid Initialization — Centroid initialization for SID token embeddings; pure SFT Recall@5 +16%, cold start +60%.

DrIG — Proposes dual-role identifiers to unify multimodal generative retrieval; outperforms SOTA generative baselines across M-BEIR.

TLCAO — Baidu proposes token-level reward allocation for generative document retrieval, substantially outperforming sequence-level rewards.

REAM — First model-merging framework to compress LLM recommendation inference, attention-head-level fine-grained merging; inference length reduced 24.3%.

Industrial Ranking & Full-Funnel

MARCO — Meta decomposes CVR prediction by click intent; conversion per click +2.80%, topline +0.98%.

From Prediction to Incrementality — LinkedIn proposes decision-centric causal optimization (Transformer causal network + Bayesian bandit + LP allocation); long-term value +7.20%.

DREAM — Alibaba proposes an agent-based meta-control architecture for Taobao homepage; reranking control IPV +2.06%, extended to full ranking IPV +2.71%, GMV +1.31%.

TM20K — ByteDance proposes two-stage distillation with teacher on full tokens + student token merging, extending sequences to 20K; ADSS +1.036%, latency +5.6% only.

DrEM — Alibaba proposes a dual-end robust ensemble ranking framework handling upstream pxtr noise propagation at both supervision and feature ends.

Progressive Alignment — NAVER WEBTOON proposes a three-stage LP-FFT-RFT post-training framework separating downstream adaptation from business metric alignment; validated effective in online A/B.

STAR — Combines structured tokenization with target-aware interest representation in the Tencent UniRec Challenge; ablations show temporal context yields the largest gains.

Sequence & Multimodal

PRISM — Proposes multi-view attention calibration (Affinity + Contrast Views) to mitigate similarity bias; average Hit@10 and NDCG@10 +5-10% across 7 datasets.

NTCF — Reinterprets graph collaborative filtering with curvature-aware propagation depth; strictly generalizes NGCF, outperforms mainstream GCF baselines on three public datasets.

TimeRoute — Time-aware modality routing + diffusion graph reconstruction; Recall@K up to +9.8% on TikTok and Amazon datasets.

SMD — Proposes sequential modality dropout; HR@10 retention improves from 22% to 61% at 95% missing rate.

DualSpectralCF — Training-free signed spectral collaborative filtering; Recall@20 up to +32.6%, cold start +29.2%.

VisGate — IKEA/KTH propose gated multimodal fusion with the gate serving as a visual utility measurement tool.

Low-Interaction-Rank — Proposes a low interaction-rank theoretical framework unifying two-encoder architectures; whitening recovers interpretable concept axes.

PreGress — First ranking-native graph pretraining + prompting framework; validated on six public graphs and two recommendation benchmarks.

Causal Inference & Evaluation

Fast A/B/n Testing — Tree-coupled feedback sharing enables exact multi-policy comparison; expected queries reduced from JT to T+o(T).

When Offline Evaluation Misleads — Thumbtack proposes a diagnostic protocol for delayed-feedback CMAB; distinguishes alignment from learnability.

TTT-Embed — Test-time optimization of embedding vectors; up to +8.36 nDCG@10 across 15 MTEB tasks, resolves catastrophic forgetting.

DR-CVR — Semiparametric theory yields a doubly robust estimator for chained CVR structures; faster convergence rate + target regularization improves stability.

Tie-Aware Evaluation — Reveals the tie-breaking illusion in group recommendation evaluation; multiple previously reported gains shrink substantially under tie-aware protocols.

SPACE — Generates synthetic users via community inference + optimal transport + constrained diffusion; long-tail POI exposure substantially improved.

Confidence Calibration Audit — First joint audit of hallucination rate and confidence calibration in LLM recommendation; four LLMs systematically under-confident.

Reward-SNR Floor — Proves the reward-SNR detectability lower bound ρ*(N)≈2.8/√N; detecting effects ≠ learning per-instance policies.

ReliableNet — Chance-constrained programming directly controls joint confident error probability; lowest JCW under both in-distribution and distribution shift.

DrEM — See industrial ranking section above.

Other

BOUND — Brief-guided corrective preference distillation; Bamboogle +5.6 EM, BrowseComp-Plus +4.8.

LoongReflect — Formalizes reflection as a memory control policy, combining global distillation + GRPO dual-channel coordination.

SAGE — SLO-aware adaptive retrieval; P95 latency -36%, retrieval cost -51%, EM only -2pp.

EAHR — Exact adaptive hybrid retrieval with dynamic depth; exactly reproduces Top-20 across 150 query-snapshot combinations, latency ratio 23-30.

Prof-K — Distribution-free single-pass filtering top-k; 1.5-10x speedup over PyTorch topk.

BC-ICL — Amazon uses a pretrained tabular foundation model for bootstrap conditional action selection; early regret advantage on standard bandit suites.

Accurate Ensembles, Fragile Narratives — Multi-scale stacked ensemble achieves ROC-AUC 0.9539, but LLM explanation fidelity audit reveals prompt engineering shortcomings.

ConnectionMind — Meta applies social graph + LLM graph reasoning to recommendation; video watch time +0.43%.

TSPORec — Preference optimization for token selection; performance up to +31.25%, efficiency up to +63.4%.

DistilVDR — 524M dual-student distillation for visual document retrieval; reaches 86.9% of an 8B teacher with 15.6x smaller index.

CVR Decomposition — See above.

Generative Universal Multimodal Retrieval — See above.

MIJSR — Structural + semantic multi-interest mining for search-recommendation mixed sequences; validated on two open datasets.

Invisible to the Machine — AI recommendation audit based on full market census; 85.6% of venues never recommended.

RCCP — Retrieval-corrected conformal prediction; achieves target coverage in all settings with the lowest Winkler score.

Batch Size or Negatives — Under fixed memory budgets, n~B, k~1 is optimal; faster convergence validated on four real datasets.

AnnoIndex — Structured materialized index + SQL extension execution plan; average F1 0.87.