type
Post
status
Published
date
Aug 22, 2026 05:32
slug
rec-weekly-en-2026-W34
summary
This week (2026-08-16 ~ 2026-08-22), recommendation system research centers on three main threads: industrial systems moving from "scenario-specific models" to "unified architectures," generative recommendation shifting from "works offline" to "deployable online," and agent-based recommendation entering the engineering and evaluation standardization phase. Thread 1: Unified architectures accelerate consolidation of multiple business streams. Xiaohongshu's OneModel uses a single model to unify organic recommendation, advertising, and merchant traffic lines, with online ad-side CTR up 8.18%. Meta's UniDot unifies sequence modeling and feature interaction from an FM dot-product perspective. The shared implication of both works: as storage and compute dividends fade, the engineering cost of fragmented multi-scenario deployments becomes the primary bottleneck. Thread 2: Generative recommendation accelerates toward deployment. Kuaishou's OGR delivers end-to-end generation of ordered slates, with online Effective Views up 1.120% and NDCG@5 up 48.2% on industrial data. EchoRec extends multi-token prediction from an efficiency tool to a dense supervision signal. The keyword for generative recommendation this week is "output artifacts" — not just generating a single item, but generating an ordered list. Thread 3: Engineering of agent-based recommendation systems. Alibaba's PILOT uses LLM agents for experiment management and policy search, improving search efficiency from 53.3% to 93.3%. Microsoft's AdsWorldEngine enables agents and tools to co-evolve in conversational advertising, with online RPM up 22%. Agent-based recommendation is moving from single-point inference to process governance.
tags
Recommendation Systems
Weekly
Papers
category
Rec Tech Report
icon
📚
password
priority
1
Weekly Overview
This week (2026-08-16 ~ 2026-08-22), recommendation system research centers on three main threads: industrial systems moving from "scenario-specific models" to "unified architectures," generative recommendation shifting from "works offline" to "deployable online," and agent-based recommendation entering the engineering and evaluation standardization phase.
Thread 1: Unified architectures accelerate consolidation of multiple business streams. Xiaohongshu's OneModel uses a single model to unify organic recommendation, advertising, and merchant traffic lines, with online ad-side CTR up 8.18%. Meta's UniDot unifies sequence modeling and feature interaction from an FM dot-product perspective. The shared implication of both works: as storage and compute dividends fade, the engineering cost of fragmented multi-scenario deployments becomes the primary bottleneck.
Thread 2: Generative recommendation accelerates toward deployment. Kuaishou's OGR delivers end-to-end generation of ordered slates, with online Effective Views up 1.120% and NDCG@5 up 48.2% on industrial data. EchoRec extends multi-token prediction from an efficiency tool to a dense supervision signal. The keyword for generative recommendation this week is "output artifacts" — not just generating a single item, but generating an ordered list.
Thread 3: Engineering of agent-based recommendation systems. Alibaba's PILOT uses LLM agents for experiment management and policy search, improving search efficiency from 53.3% to 93.3%. Microsoft's AdsWorldEngine enables agents and tools to co-evolve in conversational advertising, with online RPM up 22%. Agent-based recommendation is moving from single-point inference to process governance.
Generative Recommendation and Semantic IDs
Two clear progress lines emerge in generative recommendation this week: semantic ID construction quality is being systematically scrutinized, and generation targets are evolving from single items to slate-level alignment.
Kuaishou's OGR (OGR) proposes an end-to-end "Once Generated, Ranked" framework. Its core is TUSID — a hierarchical semantic ID that fuses item-level semantic information with local collaborative signals. Compared to TIGER's pure semantic IDs, TUSID injects collaborative filtering signals into the ID construction process, addressing the ID collision problem where "semantically similar but behaviorally distinct" items get conflated. The generation side uses list-wise preference planning plus pipeline position-level decoding to directly generate ordered slates rather than single items. Training employs SPA (reward-guided conservative policy optimization) for preference alignment, avoiding the mediocre recommendations that pure likelihood imitation produces. Offline experiments show a 48.2% relative NDCG@5 improvement on industrial datasets, and online A/B tests show Effective Views up 1.120%. The framework extends ideas from DualGR — long/short-term interest routing and exposure-aware loss — to slate generation.
A diagnostic view of semantic IDs. Huawei's SIDScope (SIDScope) is a noteworthy diagnostic resource. It systematically evaluates the interface health of semantic IDs derived from nine tokenizers (covering seven families including TIGER, E4SRec, M2TRec). The core finding is "mechanism conditionality" — when retrieval consumes SID prefixes, prefix alignment strongly correlates with candidate exposure; when scoring becomes prefix-agnostic, this correlation disappears. It also reveals a hidden gap: the probability difference between "valid target path exists" and "target item uniquely retrieved" is 1.2–3.0 percentage points. This means many generative recommendation models' "correct generation" in offline evaluation does not translate to "correct retrieval."
EchoRec turns multi-token prediction into a supervision signal. (EchoRec) observes that future behavior carries a "semantic echo" of current behavior — similarity far above random pairs, but decaying with intent shifts. Based on this, EchoRec designs a two-stage module: Horizon-aware Preference Generation (HPG) appends a lightweight auxiliary branch to the base recommender, modeling preference evolution conditionally step by step; Verifiable Holistic-Preference Alignment (VHA) uses a cycle-consistent projector to suppress spurious alignments, theoretically ruling out rank collapse under invertible transformations. The auxiliary components are discarded at inference, adding zero online overhead. It outperforms TIGER, EAGER, and other baselines on Amazon Beauty/Sports/Toys, and naturally acquires multi-item generation capability.
Decoupled temporal encoding. (Decoupled Temporal Encoding) comes from a real food delivery and instant retail recommendation system, observing multi-level temporal regularities in user behavior: recency effects, meal-time peaks, weekday/weekend switches, and promotional traffic surges. Existing methods (timestamp features, interval embeddings, decay functions, attention biases) typically compress heterogeneous temporal signals into a unified representation, failing to distinguish macro temporal dynamics from local sequential cues. DTE decouples them with two complementary modules — a personalized macro-temporal module injects compact temporal primitives into item embeddings, and a time-gated micro-sequence module introduces relative order bias only when interactions are temporally dense. Lightweight, parameter-efficient, and easy to integrate into existing systems.
Outfit generation. (USCM) formalizes outfit generation as a Constrained Ensemble Generation (CEG) problem, modeled as a finite-horizon deterministic MDP. USCM jointly models ensemble-level compatibility and latent combination intent, using Latent Expansion Monte Carlo Tree Search (LE-MCTS) for item retrieval during composition, balancing local aesthetic synergy with global structural balance. It achieves SOTA on Polyvore Outfits, with zero-shot validation on iFashion and PolyvoreU.
Common trend: generative recommendation is no longer only about single-item generation quality — slate-level utility, end-to-end retrieval resolvability, and temporal structure are now part of the modeling objective. SID construction quality is emerging as a new evaluation dimension.
LLM Reasoning and Agent-Enhanced Recommendation
This week's agent recommendation highlights aren't in reasoning chain design but in "governance" — how to make agents decide within safety boundaries, how to make agents and their tools co-evolve, and how to keep inference costs affordable.
Alibaba PILOT: agents for experiment governance. (PILOT) is an LLM-agent framework deployed at Taobao. The core idea isn't reasoning over recommendation models — it's handing the "optimize the recommendation system" task itself to agents. Three roles collaborate: the Experiment Manager drives the full experiment lifecycle (task intake, observation governance, anomaly recovery, post-hoc review), only selecting actions from a rule-generated legal command package; the Search Planner proposes decision-tree candidates for user segments when the Manager requests them; the Memory Curator asynchronously distills experiment results into policy-level domain knowledge, fault-isolated from the main loop. Deployed across 5 Taobao experiment buckets, it improves IPV by 1.40% (vs. ROAM's 1.00%) and GMV by 1.50% (vs. ROAM's 1.13%), with search efficiency up from 53.3% to 93.3% — all without human intervention. PILOT's value: it turns recommendation system optimization experience into reusable methodology rather than starting from scratch each time.
Microsoft AdsWorldEngine: agents and tools co-evolve. (AdsWorldEngine) is an agentic framework for conversational advertising. The Opportunity Gate decides whether to show ads, the Orchestrator generates business intent, calls ad tools, and builds a Top-3 ad list, and the Evaluator scores served ads for offline optimization. The core innovation is an iterative actor-tool training loop — first SFT + agentic RL trains the Orchestrator, then high/low reward rollouts construct preference data to train the tools themselves. This creates a self-improvement cycle: the system not only learns to use ad tools but learns to improve them from reward behavior. A companion label-grounded judgment modeling approach enriches labels with thought traces, filters inconsistent rationales via reflection, and uses a cost-sensitive GRPO variant to preserve asymmetric reward gaps. Offline diversity improves 60%, relevance 80%; online A/B tests show RPM up 22% and ad coverage up 74%. Compared to RecThinker's fixed tool calls, AdsWorldEngine's tools themselves become learnable objects.
Prune, don't distill. (rEDMRec) argues LLM reasoning doesn't need to be regenerated every time. If reasoning can be compressed into structured memory, a lightweight model only needs to retrieve. rEDMRec distills teacher LLM reasoning into four editable experience channels: long-term preference, short-term context, item awareness, and counterfactual hard-negative comparisons. An LLM memory controller handles Add/Delete/Modify/Keep operations, optimizing memory entries via K-agent debate. Lightweight student models (3B–20B) rank purely by retrieving memory, never calling the teacher. Validated on ML-1M, Amazon Beauty, and Steam with ten student backbones, HR@1 improves up to 13.3% over the second-best baseline. Ablations show: short-term context is the only channel consistent across all capacity tiers; long-term, item-aware, and counterfactual contributions are capacity-dependent — they can even reverse on the strongest student models.
Cognition-inspired two-stage recommendation. (CARA) models user decisions as two mechanisms: intuitive affective preference and deliberative rational evaluation. CARA first narrows the search space with coarse-grained preference constraints in candidate filtering, then makes final recommendation decisions via dual-perspective decision modeling (affective + rational). The introduced boundary-aware KTO strategy prioritizes training on samples the model occasionally gets right but is unstable on, increasing the density of informative preference signals. Relative improvements up to 10.15% across three Amazon Reviews domains.
The continual recommendation trilemma. (TRACER) identifies the Stability-Plasticity-Cognitivity (SPC) Trilemma: LLMs' general semantic priors (Cognitivity) conflict with preserving personalized historical preferences (Stability) and adapting to individual interest shifts (Plasticity). TRACER uses three specialized modules working together to prevent any single factor from dominating. Improvements up to 14.38% on five real datasets, extending the multi-task fusion approach of AdaTT and DTRN.
Layer-wise early-exit fusion for compact LLMs. (FLEXRec) inserts prediction heads at every Transformer layer of Qwen 3 1.7B and Llama 3.2 3B, using an AC-Router to dynamically select the number and identity of exit layers per user sequence, with a target-k hinge loss constraining routing sparsity. It achieves SOTA accuracy among compact backbone methods on three real datasets while maintaining full-corpus ranking efficiency.
Reliable conversational recommendation under fixed catalogs. (MACS) in fixed-catalog scenarios assigns LLMs language tasks (parsing user requests, preference elicitation, generating responses) while deterministic modules handle correctness-critical operations (item retrieval, hard constraint filtering, brand exclusion, progressive relaxation). A session-persistent preference layer tracks constraints across turns. On a 140-query single-turn benchmark: Pass rate 87.1%, brand compliance 1.000. On a 10-scenario multi-turn benchmark: Pass@5 reaches 72% (vs. GPT+Catalog's 56%) with zero constraint drift.
Training-free LLM recommendation refinement. (CoRRe) adopts a post-LLM paradigm — injecting collaborative signals after the LLM generates item representations. It refines embedding directions with an item-item co-purchase graph and adjusts magnitudes with item popularity, matching or exceeding some trained methods without any training.
Truncation improvement for BoN distillation. (TUP) decomposes reweighting in Best-of-N distillation into two independent dimensions: truncating low-ranked samples plus sharpening high-ranked weights. Theoretical proof: for any unknown oracle reward, optimal monotone reweighting can be matched by low-rank truncation rules — supporting "deleting" rather than "downweighting" low-ranked samples. Closed-form normalization, purely offline training, competitive with DPO, KTO, ORPO, SimPO on AlpacaEval and MT-Bench.
Component-level knowledge distillation. (GOD) notes that traditional distillation trains teacher and student independently then matches outputs, confounding the contribution of student components. GOD uses "grafting" to build hybrid models — replacing frozen teacher components with trainable student counterparts, separately evaluating student embeddings + teacher encoder and teacher embeddings + student encoder, providing component-level feedback. Inference uses only the student, zero extra cost. Improvements up to 13.92% on three real datasets.
SCoRD's continual distillation. (SCoRD) targets the high cost of continual distillation for LLM re-rankers. A semantic reasoning assistant distills the LLM's user-intent inference capability into reusable intent-level guidance; selective distillation (only on low-confidence sequences) avoids repeated LLM inference. Intent drift signals feed back to the re-ranker, enabling efficient retriever-reranker co-adaptation.
Retrieval, Recall, and Index Compression Optimization
This week's retrieval theme is "compression and refinement" — extracting more value from existing resources within limited storage, compute, and context budgets.
Meituan's reasoning-driven personalized retrieval. (TTP) addresses the intent gap problem in local life e-commerce: sparse ambiguous queries plus rich user history are needed to complete intent. TTP uses two-stage training to unify reasoning and retrieval: an SFT stage builds cold-start capability, and a GRPO stage aligns the reasoning process with retrieval utility. It explicitly reasons over user historical purchase sequences, generates intent-augmented queries, then encodes them into unified dense embeddings. Online A/B tests show order volume up 0.46%. It extends HyMiRec's multi-interest extraction but turns reasoning from implicit embedding interaction into an explicit reasoning process.
Structure-aware quantization. (Variable Bit Allocation for Quantization) observes that embeddings from modern models exhibit significant geometric structure (Matryoshka properties), while existing quantization schemes allocate bits uniformly across dimensions. It proposes a simple variable bit allocation framework: partition embeddings into contiguous buckets and use a greedy strategy for non-uniform storage allocation. At the same storage budget, non-uniform allocation consistently outperforms uniform baselines; the low-bit regime gains the most — PQ recall up 8%, SQ up 18%. Beyond DRQ and UniNote, this points to a new direction in structure-aware compression.
Task-aligned quantization for late interaction. (CrossQ) targets the large index problem of ColBERT-style late interaction retrieval. Standard compression minimizes token reconstruction error, but ranking quality depends on preserving sparse "winner token" scores. CrossQ's insight: compute lightweight document context at indexing time (not stored), and use it to condition token codebook selection. The training objective aligns with ranking — preserving candidate score distributions and protecting hard-negative margins. At 2 B/token, MRR@10 improves 0.010 over the strongest same-size baseline; at 4 B/token, it achieves 64× compression of original token storage; at 8 B/token, it retains 98% of full-precision ColBERT's MRR@10.
Relational recall under a fixed budget. (DBRR) answers a direct question: under a fixed retrieval budget, how do you recover more evidence beyond top-k? DBRR allocates the budget as relevance-selected seeds plus constrained graph neighborhoods. On HotpotQA FullWiki, full supporting-evidence recovery improves 23.8 percentage points (paired risk difference 0.2377, bootstrap 95% CI 0.2269–0.2489). Bridge-type questions drive the main gains (28.7 points); comparison-type questions gain less (4.2 points). Real relations outperform random-neighbor and degree-preserving rewired graph controls. Conclusion: evidence completeness depends not only on which items rank highest but on how context is allocated around them.
Unified query expansion and retrieval. (DEPT) solves the moving-target problem of a single model doing both query expansion and dense retrieval: retrieval supervision should improve query-side expansion, but the same update also moves the document embeddings that are the retrieval target. DEPT keeps fine-tuned document embeddings close to cached initial embeddings while allowing retrieval gradients to flow into the generator through straight-through decoding. Validated on five BEIR datasets with Qwen3-4B and LLaMA-3.2-3B, average nDCG@10 surpasses training-free, independently-trained, and staged-unified baselines.
Cross-engine portability of static pruning. (Static Pruning Across Sparse Retrieval Regimes) is the first systematic comparison of static pruning across three engines (exact inverted index, BMP, SEISMIC), covering 1,140 configurations. Findings: index-side pruning (documents and posting lists) is portable — consistently reducing latency 1.2–6.6× and index size 18–82%, because sparse retrieval is memory-bound; query-side pruning is already internalized by modern engines (BMP's β, SEISMIC's query_cut); static and dynamic pruning are complementary — combining document and query pruning on BMP yields 2.5× speedup with NDCG@10 within 0.003 of the exact baseline. NDCG@10 saturates while Recall@10 remains in the 85–95% range — this inflection point can serve as a portable stopping criterion.
Unsupervised sparse semantic retrieval. (Sparse Coverage) addresses the problem in patent retrieval where single-vector embeddings may compress multiple technical components. It maps local span embeddings to a sparse vocabulary of centers in embedding space, with centers selected by a coverage-oriented k-center objective, and span activations over nearby centers producing sparse representations compatible with inverted indexes. On CLEF-IP 2013, it matches or exceeds document-level recall of strong dense patent encoders.
Multi-aspect retrieval benchmark. (MAPLE) notes that existing paper retrieval benchmarks only evaluate single-query-paper relevance, ignoring the multiple searchable aspects of the same paper. MAPLE contains 2,095 queries covering three aspects: motivation, method, and experimental results. The strongest model reaches 98.1% AnyAspect@20 but only 15.7% AllAspect@20 — a massive gap. Experiment/result queries and table-reference queries are hard for most retrievers.
RAG diversity beats redundancy. (Redundancy and Diversity in RAG) uses controlled FictionalQA experiments, excluding confounds from parametric priors and answer relevance, and finds: duplicate documents and LLM rewrites don't significantly improve answer correctness, while diverse documents (news, blogs, different genres) improve it 17%–47%. The gains are driven by genre diversity, not by more relevant answers being available.
Source-style collapse in agent tool retrieval. (ToolScout) discovers a collapse phenomenon in fine-tuned retrievers on source-specific slices — FT-1100, despite higher lexical overlap with gold tools, collapses on another source slice. Query-side TF-IDF fingerprints identify failing source styles better than semantic or length proxies. ToolScout uses this signal for routing, improving coverage from 22.3% to 86.1% on a mixed 4,996-query stream.
Multi-Scenario Unified Ranking and Industrial Evaluation
The industry's main theme this week is clear: moving from scenario-specific models to unified architectures. Meanwhile, evaluation methodology shows two directions — offline task design (impression share prediction) and regression testing of benchmarks.
Xiaohongshu OneModel: unified multi-stream ranking. (OneModel) uses a single model covering three traffic lines: organic recommendation, advertising, and merchant services. Method: map heterogeneous behaviors to a shared event sequence, learn long-context user representations with an action-oriented backbone, and introduce Scenario-aware Information Modulation to balance cross-stream transfer with scenario specialization. Deployment uses hierarchical user representations, multi-objective training, feature decomposition, user feature prefetching, shared user tower computation, and graph-level inference optimization. Online A/B numbers: Explore Feed Time Spent +0.33%, Engagement +1.25%; Feed Advertising ad value +3.43%, CTR +8.18%; Merchant Recommendation DGMV +1.1867%, GPM +2.1585%. Compared to HSTU-style architectures, OneModel emphasizes multi-scenario conditioning — shared long context plus scenario modulation.
Meta UniDot: a unified view of FM dot products. (UniDot) makes an elegant observation: FM's embedding inner product and attention's query-key dot product are the same primitive. Based on this, UniDot tokenizes non-sequential fields and multi-domain behavior sequences into a shared token space, runs a token-mixing bus (feature interaction) and a sequence-retrieval bus (sequence modeling) in parallel, fuses states via MLP-Mixer at each layer, and carries explicit per-layer dot-product interactions to the classifier through an FM Highway. Runner-up in the TAAC KDD Cup 2026 industrial track.
Google SDF: dual-mechanism decomposition of content staleness. (SDF) deployed on Google Discover (hundreds of millions of DAU), decomposes content staleness into two mechanisms: supersession (new updates making old coverage stale) and relevance decay (information value naturally decaying over its lifecycle). Two complementary filters: a relational staleness model detects supersession between item pairs, and a predicted traffic ratio (PTR) model predicts relevance decay from content. Stale candidates are pruned disjunctively before ranking. After two years of production deployment, user-submitted staleness feedback dropped 54.9% relative to pre-deployment. This is more precise than traditional age truncation and behavioral heuristics — age truncation can't reflect true relevance loss, and behavioral heuristics rely on lagging signals.
A new offline evaluation task. (Impression Share Prediction) Meta proposes impression share prediction as an offline evaluation task: given a candidate ranking model, predict its impression distribution across target buckets (grouped by optimization objectives like clicks, video views). This is a counterfactual task — the candidate has never been served online. A structural causal model identifies counterfactual effects, and a statistical learning framework predicts from the candidate's early interaction confidence signals and current system state. Random Forest reduces L1 error 49% over a constant baseline on models seen during training; for unseen models, the first hour is hardest — capacity state still reflects the previous model, and an encoder-conditioned architecture recovers +22% L1 after simulating a 2-hour rollout.
Sequential decision-making for e-commerce media ranking. (SMEO) Amazon's media ranking problem: images, videos, and 3D renders on product detail pages are an ordered set, consumed sequentially by customers. SMEO has two stages: first, learn a trajectory utility model that estimates how ordered evidence from consumed media prefixes helps customers reach purchase decisions (mitigating position bias and depth imbalance); then, train an autoregressive ranking policy using survival-weighted reward-to-go to prioritize the most decision-relevant information. Offline evaluation (doubly robust off-policy estimation) shows conversion up 5.5% and 15% fewer swipes needed for customers to reach purchase decisions.
Benchmark regression testing. (Do Sequential Recommendation Benchmarks Really Require Higher-Order Sequence Modelling?) uses two simple recency-weighted pairwise probes (SeqRules and PCTM) to test whether existing benchmarks actually require higher-order sequence modeling. Results: on three Amazon datasets, at least one probe exceeds eSASRec reproductions by 15–38%; on MovieLens-1M by 4.4%; on MovieLens-20M it lags by 27.3%. On four other datasets, at least one probe exceeds sampled-softmax SASRec reproductions by 9–28%. The conclusion: existing benchmarks may not be suitable for measuring gains from higher-order sequence modeling. This directly informs how teams should evaluate sequence recommenders like DualGR and HiGR.
Financial heterogeneous graph benchmark. (FinFraudBench) provides two large-scale heterogeneous graph datasets (CreditCard-Fraud and BankTrans-Fraud) with 8.99M nodes and 89.23M directed edges respectively, preserving six entity types and fourteen edge types, with natural fraud rates mirroring deployment constraints. It includes standardized evaluation protocols covering ranking and imbalance-sensitive classification metrics, benchmarking GCN, GAT, GraphSAGE, HGT, RGCN baselines.
Privacy-aware cloud-device collaborative recommendation. (PriCoRec) addresses the challenge of privacy regulations restricting cloud-side processing of sensitive user data (age, gender, etc.). It designs cloud-side pre-ranking (available features) plus device-side fine-ranking (local personalized features). Cloud-side pre-ranking introduces a diversity regularizer to improve candidate quality; a cloud-guided training mechanism uses knowledge distillation to keep the device-side model lightweight while maintaining performance. Traditional knowledge distillation on the device side typically only does label imitation — PriCoRec instead uses cloud-side rich features to guide device-side model training.
Ranking collapse in time-series foundation models. (Forecast Collapse) uncovers a cautionary phenomenon: when TSFMs predict hourly returns for 1,000 US stocks, predictions become nearly flat with poor ranking ability. The cause is twofold: low predictability limits calibrated point forecast magnitude, and per-sequence objectives fail to identify cross-sequence structure. This exposes a calibration-ranking tradeoff — optimizing squared error produces flat forecasts, while directly optimizing cross-correlation improves ranking but inflates forecast magnitude. The CalibRank objective nearly doubles cross-correlation on Finance1K while keeping magnitude close to target.
Directions to Watch
Multimodal foundation models become standard infrastructure for industrial asset personalization. Netflix's practice (Multimedia Asset Personalization) demonstrates how CLIP and their in-house trimodal model MediaFM unify five canvas models, solve cold start, and make video preview personalization outperform unimodal baselines. A key lesson: a simple offline proxy task (predicting popularity winners) can accelerate model screening as a gate before end-to-end A/B testing. This corroborates BiListing at Airbnb and Spotify's 2T-HGNN for audiobook recommendation — multimodal foundation model embeddings will become standard infrastructure in industrial recommendation systems.
Innovation in offline evaluation methods. Several works this week point in the same direction: offline evaluation shouldn't just be a proxy for prediction accuracy — it should predict some verifiable online outcome. Meta's impression share prediction forecasts candidate models' impression distributions, Netflix's proxy task predicts popularity winners, and impression share prediction uses structural causal models to identify counterfactual effects. This hints at the next step for offline evaluation — moving from "does this model have high AUC" to "how will impressions distribute if this model goes live."
Unified architectures expanding across business streams. OneModel's cross-stream unification (organic + ads + merchant) and OneMall's multi-scenario e-commerce unification confirm a judgment: user behavior within a platform is a continuous cross-stream trajectory, and modeling it separately fragments user representations. Worth watching next: how unified models achieve fine-grained scenario conditioning — OneModel's Scenario-aware Information Modulation is one direction, UniDot's FM perspective is another.
Paper Roundup
Generative Recommendation and Semantic IDs
OGR — Kuaishou's end-to-end ordered slate generation framework; TUSID fuses semantic and collaborative IDs; online Effective Views +1.120%, NDCG@5 +48.2%.
SIDScope — Diagnostic resource for semantic ID interfaces, covering nine tokenizers across seven families; reveals path-item inconsistency hidden gap of 1.2–3.0 percentage points.
EchoRec — Cycle-consistent preference alignment extends multi-token prediction from efficiency tool to dense supervision; outperforms TIGER, EAGER on three Amazon datasets.
Decoupled Temporal Encoding — Decouples macro temporal dynamics from micro sequential information; lightweight, deployment-friendly, applicable to generative recommendation.
USCM — Formalizes outfit generation as CEG problem; USCM + LE-MCTS achieves SOTA on Polyvore Outfits, zero-shot validated on iFashion and PolyvoreU.
LLM Reasoning and Agent-Enhanced Recommendation
PILOT — Taobao-deployed LLM-agent experiment optimization framework; three-role controlled loop; search efficiency 53.3%→93.3%, GMV +1.50%.
AdsWorldEngine — Microsoft's self-evolving conversational advertising agent; Orchestrator and tools co-evolve; online RPM +22%, ad coverage +74%.
rEDMRec — Distills LLM reasoning into four editable experience channels; student models (3B–20B) rank purely by retrieval; HR@1 up +13.3%.
RecPFN — Prior-fitting network for sequential recommendation ICL; synthetic data pretraining; zero-shot SOTA on eight benchmarks.
SCoRD — Semantic reasoning assistant enables retriever-reranker continual co-adaptation; selective distillation cuts LLM inference cost.
CoRRe — Post-LLM paradigm injects collaborative signals to refine item embeddings; matches some trained methods without training.
CARA — Cognition-inspired two-stage recommendation framework; boundary-aware KTO strategy; relative improvements up to 10.15%.
FLEXRec — Layer-wise early-exit fusion with adaptive routing for compact LLMs; SOTA accuracy on Qwen 3 1.7B / Llama 3.2 3B while maintaining full-corpus ranking efficiency.
MACS — Hybrid multi-agent conversational e-commerce recommendation; deterministic constraint enforcement; multi-turn Pass@5 72% (vs. GPT+Catalog 56%).
TRACER — Balances stability-plasticity-cognitivity trilemma; improvements up to 14.38% on five real datasets.
TUP — Truncation-sharpening decomposition for BoN distillation; theoretical proof that low-rank truncation matches optimal monotone reweighting.
GOD — Component-level grafting distillation; improvements up to 13.92% on three real datasets.
Retrieval, Recall, and Index Compression Optimization
TTP — Meituan's reasoning-driven personalized dense retrieval; SFT + GRPO explicit user intent reasoning; online order volume +0.46%.
SSR-GRPO — Alibaba combines semantic IDs with RL for e-commerce retrieval; dual-perspective relevance evaluation + hard negative mining; deployed.
Quantization Beyond Uniform Bit Allocation — Variable bit allocation quantization exploiting Matryoshka structure; PQ recall up +8%, SQ +18%.
CrossQ — Cross-token conditioned quantization; MRR@10 +0.010 at 2 B/token; 64× compression at 4 B/token.
DBRR — Seed + graph neighborhood allocation under fixed budget; HotpotQA supporting-evidence recovery +23.8 percentage points.
DEPT — Single decoder-only LLM jointly optimizes query expansion and retrieval end-to-end; average nDCG@10 gains on five BEIR datasets.
Static Pruning Across Sparse Retrieval Regimes — Cross-engine pruning portability study across 1,140 configurations; index-side pruning portable; static and dynamic pruning complementary.
Sparse Coverage — Unsupervised sparse semantic retrieval; k-center coverage center selection; matches or exceeds strong dense encoders on CLEF-IP 2013 document-level recall.
MAPLE — Multi-aspect paper retrieval benchmark; strongest model only 15.7% AllAspect@20.
How retriever redundancy and diversity impact RAG — Controlled experiments prove document diversity improves correctness 17–47%; redundancy and rewriting show no significant gains.
ToolScout — Source-style collapse detection and routing; TF-IDF fingerprints lift coverage from 22.3% to 86.1%.
Multi-Scenario Unified Ranking and Industrial Evaluation
OneModel — Xiaohongshu's unified multi-stream ranking (organic + ads + merchant); ad CTR +8.18%, DGMV +1.1867%.
SDF — Staleness filtering system deployed on Google Discover; decomposes supersession and decay dual mechanisms; user staleness complaints down 54.9%.
UniDot — FM dot-product perspective unifies feature interaction and sequence modeling; runner-up in KDD Cup 2026 industrial track.
Impression Share Prediction — Meta's new offline evaluation task; Random Forest L1 error −49%; encoder-conditioned simulation recovers +22%.
SMEO — Amazon's two-stage sequential decision-making for e-commerce media ranking; offline conversion +5.5%, swipes −15%.
SAGA — Generative action embeddings for financial multi-surface behavior sequences; per-field tokenization; downstream click and conversion gains.
Do Sequential Benchmarks Need Higher-Order Modeling? — Simple probes beat Transformer reproductions on multiple benchmarks, questioning existing benchmarks' ability to measure higher-order sequence modeling.
FinFraudBench — Large-scale heterogeneous graph financial fraud benchmark; 8.99M nodes, 89.23M edges, six entity types.
PriCoRec — Cloud-device collaborative privacy-preserving advertising recommendation; diversity regularizer + cloud-guided training; sensitive features stay on device.
GEO-Flag — GEO optimization webpage detection benchmark; IPT lifts F1 from 0.862 to 0.944 on ModernBERT.
Forecast Collapse — Forecast collapse phenomenon in time-series foundation models; CalibRank nearly triples cross-correlation.
RelArena-α — Unified relational learning benchmark framework + TabPFN-Rel + model-agnostic interface; three open-source software systems.
Other
Lipschitz Bandits with Arbitrary Feedback Delays — Lipschitz bandits under arbitrary feedback delays; elimination algorithm (stochastic) + EXP3 (adversarial); regret bounds match the no-delay case plus a √D delay cost.