RecSys Weekly 2026-W35 | Recsys Frontier

type

Post

status

Published

date

Aug 29, 2026 05:33

slug

rec-weekly-en-2026-W35

summary

This week's recommendation systems research clusters around three technical threads: Semantic ID engineering is moving into deep water, retrieval systems are shifting from "static pipelines" to "query-adaptive" architectures, and the role of LLMs/Agents in the recommendation stack is evolving from "enhancement components" to "closed-loop decision makers." Thread 1: Semantic IDs move from "generatable" to "usable, maintainable, drift-resistant": Two Kuaishou papers tackle codebook structure and dynamic updates respectively — a single-level large codebook replaces multi-level residual quantization, compressing three-level SIDs into two levels, cutting autoregressive decoding FLOPs by 47.93%-48.70% with +0.792% online consumption metrics; TAGR designs dynamic semantic IDs (LSID) for live-stream advertising, achieving +8.5% room entry rate and +16.1% revenue. Tencent's Tlow replaces RQ-VAE with flow-based transforms to solve codebook dependency, lifting CTR by 10.32% in WeChat multimodal retrieval. All three point to the same conclusion — SID engineering stability (not generation accuracy) is now the deployment bottleneck. Thread 2: Retrieval efficiency shifts from "one-size-fits-all" to "query- and context-adaptive": Alibaba's TransRetrieval uses weighted average aggregation to resolve token-norm divergence from feature heterogeneity, validating log-linear scaling on a 4-billion-interaction dataset with +2.53% online revenue. AdaWidth goes further — dynamically allocating different embedding widths per query, matching SOTA NDCG@10 with 55%-84% fewer dimensions across 6 tasks. VK's multi-hash user embeddings shrink the ID embedding table by 98%, cutting single-node temporal neighbor sampling cost from O(deg(v)+k) to O(log(deg(v))+k). The granularity of efficiency optimization is moving from "system-level" down to "query-level." Thread 3: LLMs/Agents move from "recommendation engines" to "drivers of system self-evolution": Alibaba's Astar hands the "propose evolution dir

Weekly Overview

Thread 1: Semantic IDs move from "generatable" to "usable, maintainable, drift-resistant": Two Kuaishou papers tackle codebook structure and dynamic updates respectively — a single-level large codebook replaces multi-level residual quantization, compressing three-level SIDs into two levels, cutting autoregressive decoding FLOPs by 47.93%-48.70% with +0.792% online consumption metrics; TAGR designs dynamic semantic IDs (LSID) for live-stream advertising, achieving +8.5% room entry rate and +16.1% revenue. Tencent's Tlow replaces RQ-VAE with flow-based transforms to solve codebook dependency, lifting CTR by 10.32% in WeChat multimodal retrieval. All three point to the same conclusion — SID engineering stability (not generation accuracy) is now the deployment bottleneck.

Thread 2: Retrieval efficiency shifts from "one-size-fits-all" to "query- and context-adaptive": Alibaba's TransRetrieval uses weighted average aggregation to resolve token-norm divergence from feature heterogeneity, validating log-linear scaling on a 4-billion-interaction dataset with +2.53% online revenue. AdaWidth goes further — dynamically allocating different embedding widths per query, matching SOTA NDCG@10 with 55%-84% fewer dimensions across 6 tasks. VK's multi-hash user embeddings shrink the ID embedding table by 98%, cutting single-node temporal neighbor sampling cost from O(deg(v)+k) to O(log(deg(v))+k). The granularity of efficiency optimization is moving from "system-level" down to "query-level."

Thread 3: LLMs/Agents move from "recommendation engines" to "drivers of system self-evolution": Alibaba's Astar hands the "propose evolution directions" step — previously dependent on senior experts — to the model: Astar-8B achieves a single-proposal success rate of 0.6786 in real execution evaluation, double that of human experts (0.3229). PayPal's SCOUT reframes MCP tool exposure as a context selection problem, cutting tool token consumption from 140.2k to 1.3k (a 99% reduction). Both point to the same trend: LLMs are no longer just part of the recommendation model — they're beginning to take over the iteration process of recommendation systems themselves.

Generative Recommendation & Semantic Encoding

Semantic ID work was dense this week, and nearly all of it comes from industry. The core debate has shifted from "how to generate better SIDs" to "how SID systems handle dynamic traffic, control decoding costs, and transfer across architectures."

AMBER (Meta)** — Raises the perspective from "item" to "event." Each interaction isn't a token but a complete temporal snapshot (user, item, context, outcome signals); AMBER compresses this snapshot into a single Event Token. This introduces a new scaling dimension — snapshot resolution, i.e., how much information each event encodes. Event Tokens are precomputed and cached, decoupling snapshot resolution from serving-time computation. On industrial-scale ranking and retrieval benchmarks, AMBER advances the compute-quality Pareto frontier. A notable result: at sufficient capacity, a single unified tokenizer outperforms tokenizers designed per entity type, suggesting positive transfer across entity types. Event Tokens can also be injected as historical features into non-LLM rankers with statistically significant gains — implying semantic encoding benefits aren't limited to generative architectures. Unlike TIGER's autoregressive semantic ID generation or LlamaRec's sequence modeling, AMBER changes the answer to "what does each position encode" from "item" to "event."

TAGR (Kuaishou)** — A three-layer time-adaptive framework for live-stream advertising. Static semantic IDs can't track rapidly changing live content; TAGR proposes Live Semantic-Collaborative IDs (LSID): periodically refreshing each active ad's SID based on current live context and promoted products, while retaining a stable hierarchical token vocabulary for autoregressive generation. At the intent layer, multi-time-granularity room-entry history serves as the primary intent sequence, auxiliary behaviors are input separately, and next-token prediction is weighted by post-request intent evidence and commercial value. At the alignment layer, Intermittent On-Policy Preference Optimization (IOPO) periodically samples fresh candidate groups from the current policy for preference updates, interleaved with supervised NTP maintenance. Online results: +8.5% room entry rate, +7.4% add-to-cart rate, +16.1% revenue. Compared to DualGR's long/short-term interest separation, TAGR's differentiator is building "time-adaptivity" into the tokenizer itself, not just behavior modeling.

Tlow (Tencent)** — Solves RQ-VAE's codebook dependency from a distributional perspective. RQ-VAE codebooks have inherent inter-codebook dependencies and inefficient decoding; efficient independent tokenizers like OPQ can't handle the dimensional correlations and distributional complexity of semantic embeddings. Tlow uses flow-based transforms to map raw semantic embeddings into a standard normal latent space, gaining both dimensional independence and distributional simplification, then performs independent tokenization on the transformed space. It also introduces codebook guidance to align the codebook space with the token embedding space. In WeChat multimodal retrieval, the token-ID retrieval model lifts user CTR by 10.32% (global) and 11.64% (new items). Compared to PIT's personalized dynamic tokenizer, Tlow's entry point is the distribution shape itself — simplify the embedding space first, then quantize.

Single-Level Large Semantic Codebook (Kuaishou)** — Directly challenges the structural assumptions of multi-level residual quantization. Three-level SIDs compress to two levels: one semantic token plus one collaborative disambiguation token. It also introduces an exposure-aware dynamic update mechanism — temporal weight decay, EMA center updates, exposure-weighted penalties — to address static codebook mismatch with traffic drift. On OneRec-V1/V2, Recall@10 improves 5.0%-8.8%, NDCG@10 improves 4.1%-8.5%. Across three serving architectures, shorter SIDs reduce autoregressive decoding FLOPs by 47.93%-48.70%, with single-card QPS up 28.57%-47.0%. A five-day online A/B (2.5% traffic) shows +0.792% consumption metrics. This work complements GRank's generate-rank framework — GRank handles post-generation ranking, this handles codebook efficiency in generation itself.

PRQ-KMeans — Treats residual quantization as a "progressive commonality removal" process and identifies three limitations: corpus-global shared components consume first-level capacity, hard assignment ignores gradual similarity of neighboring codewords, and full-codebook subtraction retains variation along the selected codeword direction in residuals. Solutions: removing the global mean component, top-k similarity-weighted centroid updates, and projected residuals replacing full-codebook subtraction. On industrial search datasets, HitRate improves 7.4%, MRR 11.8%. Unlike RQ-MoE's MoE-style quantization, PRQ-KMeans preserves RQ's progressive structure but corrects the geometric errors in residual computation.

SST (Semantic Subword Tokenization) — Identifies an overlooked problem with fixed-length SIDs: intra-item attention overload — too much attention spent on low-level intra-item dependencies rather than high-level inter-item behavior transitions. SST uses Item-level Subword Tokenization (IST) to merge stable adjacent atomic tokens into compact semantic subwords, then Behavior-induced Co-occurrence Augmentation (BCA) injects coarse-grained semantic prefix transfer signals. Across three public datasets and three generative recommendation backbones, SST outperforms both fixed-length and variable-length SID baselines. This is essentially migrating BPE from NLP into SID space — though NLP subwords have clear linguistic motivation; the "merging" criteria here need more theoretical grounding.

The Disconnect Between Better Descriptive Reasoning Trace Quality and Recommendation Effectiveness (Spotify)** — A counterintuitive empirical study. A 2x2 factorial design (SID vs. natural language titles × minimal vs. extensive SID alignment) finds that introducing explicit descriptive reasoning traces actually degrades traditional offline recommendation effectiveness — even when natural language titles produce more grounded traces. Extensive SID alignment improves trace quality but not recommendation effectiveness. This demonstrates a disconnect between reasoning trace quality and recommendation effectiveness, a warning for CoT applications in generative recommendation. This contrasts with SCoTER, which distills CoT patterns into the backbone and achieves +2.14% GMV on Tencent's ad platform — while this paper finds direct supervised training on reasoning traces is actively harmful.

Retrieval & Embedding Efficiency

This week's retrieval efficiency work has two clear directions: pushing system-level efficiency optimization down to the query level (AdaWidth, RetrievalRouter), and solving graph model and Transformer scaling bottlenecks at industrial scale (Multi-Hash, TransRetrieval).

Multi-Hash Graph Neural Networks (VK)** — Deploys an end-to-end GNN ranking system on a social graph with 194 million users and 28 billion edges. Two key designs: multi-hash ID embeddings (reducing ID embedding table size by 98%) and timestamp-sorted CSR storage with binary search (cutting temporal neighbor sampling cost from O(deg(v)+k) to O(log(deg(v))+k)). Online A/B: +16% friend additions, +11.5% unique adding users. The value here is the systematic engineering decisions — multi-hash isn't new, but using it as the primary node representation in an industrial GNN rather than an auxiliary feature requires validating the quality gap against full embedding tables; the paper confirms losslessness via ablation.

TransRetrieval (Alibaba)** — Directly addresses "why Transformer scaling laws don't hold for recommendation retrieval." Root cause: heterogeneous fields cause token-norm divergence, breaking the homogeneous token assumption Transformers rely on. The solution: weighted average aggregation (restoring homogeneity), target token compression (reducing per-candidate FLOPs by 85%), and position-style domain embeddings (unifying multiple domains, turning cross-domain data into a scaling asset). On a 4-billion-interaction industrial dataset, scaling from 0.1 to 2 MFLOPs per target yields Recall@2000 improvements of +19.3/+22.2pt, confirming log-linear scaling. Online revenue +2.53%. Compared to EGR's shared LLM embedding approach, TransRetrieval focuses on feature-engineering-level scalability — not changing the backbone, but fixing the input distribution.

AdaWidth — Query-adaptive embedding dimension width. An orthogonal prefix adapter applies the same learned rotation to queries and documents, concentrating discriminative signal into leading coordinates (preserving full-width inner products), and a lightweight router reads order statistics from the query's existing ranking, evaluating more dimensions only for queries where top results would change. Theoretical analysis bounds prefix sufficiency: required dimensions grow logarithmically with corpus size, decrease logarithmically with retrieval depth, and follow a heavy-tailed distribution across queries. Across 6 retrieval tasks and 5 frozen encoders, it matches SOTA dimensionality reduction methods' NDCG@10 with 55%-84% fewer dimensions. Unlike Matryoshka Representation Learning's fixed truncation and JPQ's query-agnostic quantization, AdaWidth makes "how many dimensions are enough" a query-level dynamic decision for the first time.

RetrievalRouter — Demonstrates that no single retrieval pipeline is uniformly optimal across queries. A lightweight query-aware router predicts the optimal pipeline (modality + architecture) from query text, with a single tunable parameter exposing the full accuracy-latency frontier. Against every static baseline, RetrievalRouter has an operating point that is both more accurate and faster: 2.5% more accurate than the best static baseline and 12.4x faster. This is fundamentally turning "system-level selection" in traditional pipeline design into "query-level selection."

PUMA — Post-hoc sparsification of frozen general-purpose multimodal embeddings. Two stages: pretraining preserves dense dot-product geometry, then fine-tunes a sparse encoder for retrieval. On Qwen3-VL-Embedding-2B, four of five benchmarks show statistically indistinguishable or better performance than dense retrieval. Storage reduces 8-16x, and inference on large candidate pools is 25x faster than exact dense scoring. Two failure modes identified: insufficient pre-TopK support and retrieval-misaligned active support. The value of sparsification: storage and inference gains without retraining the backbone.

RetrievalFormer (Amazon)** — Systematic evaluation of two-tower encoders in a shared search-recommendation index. Core finding: under a strict zero-leakage cold-start protocol, the content tower achieves 0.172 Recall@20, 1.4x the strongest dedicated cold-start method (0.124) — without cold-start-specific training. But it reveals the tradeoff between sampled InfoNCE and full softmax training: full softmax improves Recall@20 by 54% on MIND-small but exhausts accelerator memory at 240k items. Approximate nearest neighbor search can't explain the remaining gap. The value isn't methodological novelty but honest scale-quality boundary characterization — the infeasibility of full-quality training at catalog scale is a shared problem for generative recommendation and two-tower retrieval alike.

Retrieval Needs Multivectors — Theoretical work. First explicit construction of query-document sets where single-vector embeddings require exponential size for correct ranking, while polynomial-size multivector embeddings suffice. The ANDOR benchmark instantiates these hard cases — SOTA single-vector models perform poorly zero-shot, with only marginal gains after fine-tuning; multivector models consistently outperform with significant fine-tuning gains.

LLM/Agent-Enhanced Recommendation & Reasoning

The most notable shift in LLM/Agent work this week: Agents are beginning to intervene in the iteration process of recommendation systems themselves (Astar), and retrieval orchestration is moving from fixed pipelines to LLM-generated executable programs (ProRetrieval).

Astar (Alibaba/Lazada)** — One of the most aggressive industrial works this week. The "propose evolution directions" step in the AI system iteration loop previously depended on senior experts. Astar learns evolution direction generation from industrial system commit history: on the data side, pairwise sample expansion and noise filtering build a large-scale evolution corpus; on the model side, mid-training, SFT, and RL train with hierarchical prompting guiding evolution direction generation, with an RL reward model serving as a fast surrogate evaluator. Astar-8B achieves a single-proposal success rate of 0.6786, far exceeding human experts (0.3229) and GPT-5.5 (0.3071). Deployed on Lazada's ad system, it guided 20 consecutive iterations (two weeks), with offline Hitrate@200 +23.6%, online GMV +4.86%, ad revenue +1.82%. This work elevates Aligning LLMs for Controllable Recommendations's RL alignment approach from "recommendation behavior alignment" to "system evolution direction alignment."

ProRetrieval — Trains an LLM as a retrieval orchestrator. Given a natural language query, the model synthesizes an executable program — a hybrid DSL interleaving SQL operators (structured fields) with vector retrieval primitives (text and images), where SQL itself provides the logical algebra for fusing heterogeneous candidate sets. Qwen3-4B is trained with GRPO and DAPO using hierarchical four-part rewards. The 4B model surpasses GPT-5.5 (e-commerce Hit@1 0.81 vs. 0.69; email 0.91 vs. 0.86). Unlike Search-R1's RL retrieval reasoning, ProRetrieval's action space isn't "what to retrieve next" but "the compositional structure of the entire retrieval program."

CoVeMem — Replaces an Agent's text memory with vector memory. Existing agentic recommendation systems use textual narrative memory maintained by LLM calls — rewritten once per interaction, unable to leverage full interaction history; collaborative evidence (gradient similarity across the entire catalog) is lost when translated into sentences. CoVeMem uses frozen LightGCN user and item states as the memory bank; the candidate set itself retrieves the most relevant historical states as soft tokens entering the LLM context. List-wise collaborative training aligned to item semantic anchors with masked candidates teaches the model to read these states and rank through them. 19/20 metric units meet or exceed the strongest text-memory agent, with zero additional LLM memory maintenance calls. The core contribution: memory becomes gradient-updatable — full interaction history finally serves as training data, not just context.

DuELRec (SK Telecom)** — Domain-gated dual-expert framework addressing negative transfer in cross-domain sequential recommendation. Single-domain experts restrict attention to in-domain items, cross-domain experts attend across all domains, and a gating mechanism uses single-domain signals to reduce cross-domain noise. Dual-sampling token-to-item contrastive learning lets the LLM capture item-level collaborative signals. Wins against 26 SOTA methods, with online CTR +47.6%. Compared to LLMRec's pure token-level modeling, DuELRec's key move is explicitly injecting item-level collaborative signals into the LLM — insufficient token-granularity alignment is the root cause of LLMRec's negative transfer.

SCOUT (PayPal)** — The context selection problem of tool exposure on an MCP gateway. Full tool schemas saturate the context window (140.2k tokens in PayPal's deployment, 70.1% of context); SCOUT reframes tool exposure as a context selection problem — injecting only tools relevant to the current step. Hybrid retrieval (BM25 + dense vectors + RRF) exposes two MCP meta-tools (tool_search and execute_tool), supported by a zero-downtime catalog update pipeline. Production token consumption drops from 140.2k to 1.3k (0.8%), a 99% reduction. Exposed as standard MCP tools, model-agnostic.

SWIM — A list-level evaluator modeling user behavior as a finite-horizon prefix session-level survival process. SWIM decomposes the current list's contribution to session-level goals into recursive survival distributions and arrival-position-conditioned rewards, estimated in parallel with a causal-masked Transformer under industrial latency constraints. More accurate than point-wise aggregate evaluators — the latter implicitly assume exposure independence, ignoring context dependence, user continuation, and diminishing marginal utility of repeated content.

The Laws of Context Allocation — Exposes a "diagnostic hallucination" in RAG: standard relevance proxy metrics fail catastrophically on hard negatives. Replacing them with causal leave-one-out probes accurately isolates generation dependencies and formally calibrates the structural dilution of LLM attention. Key finding: single-block context widening is an "architectural trap" (penalized by relevance decay); instead, iteratively allocating computation across multiple sequential generations yields 16.7-20.5 absolute percentage points of combined recall improvement, robust when scaled to 32B models.

AGR — LLM Agent for group recommendation. A token-based hash table memory module dynamically manages group/user historical interactions, supporting insertion, update, retrieval, forgetting, and summarization. The reasoning module executes four steps: group interest collection, consensus refinement, multi-dimensional evaluation, and explainable recommendation generation. SFT teaches the model to invoke memory/reasoning modules; GRPO enhances autonomous coordination. Significantly surpasses SOTA on LastFM and Douban. Unlike STAR's teacher trajectory distillation, AGR preserves explicit multi-step reasoning rather than internalizing it into a single step.

Multimodal & Industrial System Architecture

DCEO (Alibaba)** — Replaces hand-designed multi-objective fusion weights with actor-critic. Existing e-commerce search combines item-level objectives (click, add-to-cart, purchase, GMV) via hand-crafted fusion weights that can't personalize and align poorly with user-level long-term goals. DCEO first assembles item-level proxy scores into user-level proxy metrics, quantifies their alignment with the ultimate goal via relative causal effects, then an actor dynamically generates context-dependent fusion weights directly optimizing relative causal effects. 41-day online A/B: GMV +0.36%. Compared to ESMM and MMoE's static multi-objective structures, DCEO makes "how to fuse" itself a learnable problem.

CRRN (Alibaba)** — Trigger-Introduced Recommendation (TIR) scenario: users click an item of interest (trigger), and subsequent pages recommend related target items. Triggers carry strong immediate interest but are more ambiguous and implicit than search queries. CRRN has three components: a Trigger-Target Interaction layer (personalized gating extracting interaction features), a Cascading Interest Fusion module (explicitly estimating user trigger intent, adaptively fusing immediate and personalized interests), and Category-assisted Pairwise Loss (leveraging category associations between trigger and target to enhance relevance). Outperforms SOTA on both industrial and public datasets, validated by online A/B. Continues DIN's attention-based interest modeling but makes "trigger-target relevance" an explicit modeling target.

Native Multimodal Representation (Alibaba)** — Finds that end-to-end joint training of multimodal encoders and CTR models doesn't yield improvements. Reason: user behavior in raw CTR data is driven by both multimodal semantics and non-multimodal factors, causing supervision ambiguity and inconsistent encoder updates. Solution: Mine-Then-Train — mining multimodal-interpretable high-quality training samples from CTR data and fine-tuning the encoder on them to align with user click preferences. The generalizable insight: there's a systematic misalignment between multimodal pretraining objectives and CTR task objectives; direct end-to-end training isn't the answer.

HEGM (VK)** — A replication study exposing three failure modes in EGMN (Exponential-Gaussian Mixture Network): variance collapse, component redundancy, and inactive components. HEGM fixes these via hierarchical skip-watch decomposition, KL variance regularization, and structured initialization. A 1.5-month production A/B confirms statistically significant engagement gains. This is a rare systematic failure-mode analysis of an existing SOTA method — the takeaway for practitioners: "don't treat SOTA as a black box."

Directions to Watch

Semantic ID "lifecycle management" is becoming an independent research problem. Four industrial papers this week (TAGR's LSID dynamic refresh, Kuaishou's exposure-aware codebook updates, Tlow's distributional transforms, PRQ-KMeans's residual correction) tackle SID system drift under dynamic traffic from different angles. Static codebook mismatch under new item launches, traffic shifts, and seasonal fluctuations has been independently identified as a bottleneck by multiple teams. Worth watching: incremental update protocols for SIDs, conflict resolution for multi-scenario shared codebooks, and whether a unified evaluation benchmark emerges for SID efficiency comparisons against RQ-VAE-style methods.

Query-level adaptivity moves from "whether" to "how." AdaWidth proves different queries need different dimension widths, RetrievalRouter proves different queries need different pipelines, CoVeMem proves different candidates need different memory read patterns. The common signal across all three: retrieval resource allocation granularity is descending from system-level and request-level to finer units. The practical value is inference cost optimization — when embedding tables exceed 200GB and candidate pools reach billions, query-dynamic compute allocation translates directly into latency and cost savings.

The closed loop of Agent-mediated system self-iteration is preliminarily validated. Astar's 20 consecutive iterations at Lazada demonstrate the feasibility of "AI proposes evolution directions + automated execution + automated evaluation." While the single-proposal success rate of 0.6786 isn't perfect, it's already double that of human experts. If this direction holds, it will change the R&D paradigm of recommendation systems — from "humans propose directions, systems validate" to "systems propose directions, humans review." Astar currently focuses on ad systems; generalization to the full recommendation stack and diversity of evolution directions (avoiding local optima) are the questions to watch next.

Paper Roundup

Generative Recommendation & Semantic Encoding

AMBER — Meta proposes an event-centric paradigm compressing each temporal snapshot into a single Event Token; unified tokenizer outperforms specialized ones at sufficient capacity.

TAGR — Kuaishou's three-layer time-adaptive generative recommendation for live-stream ads; +8.5% room entry rate, +16.1% revenue.

Tlow — Tencent uses flow-based transforms to solve RQ-VAE codebook dependency; WeChat multimodal retrieval CTR +10.32%, new items +11.64%.

Single-Level Large Semantic Codebook — Kuaishou replaces multi-level residual quantization with a single-level large codebook, exposure-aware dynamic updates; +0.792% consumption metrics, 48% decoding FLOPs reduction.

PRQ-KMeans — Projected residual quantization improves semantic ID generation; industrial search HitRate +7.4%, MRR +11.8%.

SST — Semantic subword tokenization solves intra-item attention overload; outperforms fixed-length SID baselines on three datasets.

The Disconnect Between Reasoning Trace Quality and Recommendation Effectiveness — Spotify's 2x2 factorial experiment finds reasoning trace quality is disconnected from recommendation effectiveness.

PrismRec — Spectral decomposition + flow matching framework for micro-video recommendation; outperforms SOTA by up to 22.65% with lowest inference cost.

ANR-DiffRec — Item co-occurrence matrix as collaborative prior in discrete diffusion training, item-level adaptive noise rescheduling.

CodeHID — Hierarchical semantic indexing + two-stage generation guidance for generative code retrieval.

GRAFT — Graph-distilled generative retrieval for scientific literature exploration; recovers 91% of graph teacher's Recall@20, aspect label precision 0.922.

Adaptive Item-based Collaborative Structures — Item-level adaptive noise rescheduling + collaborative prior-guided discrete diffusion recommendation.

Retrieval & Embedding Efficiency

Multi-Hash GNN — VK deploys GNN ranking on a 194M-user, 28B-edge social graph; multi-hash embeddings cut table size by 98%, friend additions +16%.

TransRetrieval — Alibaba's weighted average aggregation + target token compression + position-style domain embeddings; log-linear scaling validated on 4B interactions, online revenue +2.53%.

RetrievalRouter — Query-aware router jointly selects modality and architecture; 2.5% more accurate and 12.4x faster than best static baseline.

AdaWidth — Query-adaptive embedding width; matches SOTA NDCG@10 with 55%-84% fewer dimensions.

AnchorQE — Training-free query expansion ensemble, encoding original and expanded queries separately with interpolation; gains up to 12.89%.

PUMA — Post-hoc sparsification of frozen multimodal embeddings; 8-16x storage reduction, 25x faster inference.

RetrievalFormer — Amazon's two-tower shared search-recommendation index; strict cold-start Recall@20 of 0.172 (1.4x dedicated methods).

Retrieval Needs Multivectors — First proof that single-vector embeddings require exponential size for ranking tasks; ANDOR benchmark validates.

LLM/Agent-Enhanced Recommendation & Reasoning

Astar — Alibaba/Lazada learns evolution direction generation from iteration history; Astar-8B proposal success rate 0.6786 (human 0.3229), online GMV +4.86%.

ProRetrieval — Qwen3-4B performs retrieval orchestration via program synthesis; Hit@1 surpasses GPT-5.5 and Claude Opus 4.7.

CoVeMem — Collaborative vector memory replaces text memory, memory becomes gradient-updatable; 19/20 metrics meet or exceed strongest text-memory agent.

DuELRec — SK Telecom's domain-gated dual experts solve cross-domain negative transfer; beats 26 SOTA methods, online CTR +47.6%.

SCOUT — PayPal reframes MCP tool exposure as context selection; tool token consumption down 99%.

SWIM — List-level evaluator modeling finite-horizon prefix session-level survival processes.

MoPLEx — Mixed Plackett-Luce model for multi-objective alignment; clustering accuracy +43.7%, ranking accuracy +15.2%.

CRAMER — Request-aware masking controls frozen backbone parameters for instant request adaptation without retraining.

The Laws of Context Allocation — Causal leave-one-out probes reveal standard relevance metric failures; iterative context allocation yields 16.7-20.5pt recall gains.

AGR — Memory-augmented reasoning LLM Agent for group recommendation; SFT + GRPO training.

Risk-Aware Reranking — Models query-conditional relevance and tool-level exposure risk in tool retrieval, explicitly trading off safety and utility.

FashionKG-RAG — Domain-level fashion knowledge graph + two-layer path reranking RAG framework.

CAIRO — User-context-aware item profiling framework, dynamically selecting user-item pair relevant information.

Clarify-Then-Search — Baidu builds a deep search clarification benchmark (518 instances); clarification substantially improves end-to-end utility.

WARP — Wasserstein-1 distance calibrates RAG opinion summarization; distribution error reduced by at least 43%.

Storage-Retrieval Gap in Parametric KG Memory — Compiling knowledge graphs into LoRA adapter libraries is feasible, but similarity retrieval can't recover stored knowledge.

Multimodal & Industrial System Architecture

DCEO — Alibaba uses actor-critic to dynamically generate multi-objective fusion weights, directly optimizing relative causal effects; 41-day A/B, GMV +0.36%.

CRRN — Alibaba's cascading relevance-driven model for trigger-introduced recommendation; outperforms SOTA on industrial and public datasets, validated online.

Native Multimodal Representation — Alibaba finds end-to-end multimodal CTR training fails; proposes Mine-Then-Train for encoder fine-tuning.

HEGM — VK fixes EGMN's variance collapse, component redundancy, and inactive components; 1.5-month A/B confirms significant gains.

D3ER — Gradient boosting introduced to multimodal recommendation for the first time, decoupling homogeneous/heterogeneous discriminative information.

HSR — Models sequential recommendation as a dissipative Hamiltonian system, frequency-domain closed-form solutions + learnable dissipation.

MOSAIC — Meta-review aggregation of neighbor evidence solves UGC sparsity and incompleteness.

From GBT to Deep Recommenders — Intuit migrates production customer-service recommendation from gradient boosted trees to a pairwise-binary deep recommender, surpassing baseline in later conversation stages.

One Hierarchy, Two Systems — DoorDash uses a single hierarchical semantic ID for both discovery ranking and query rewriting, validated online.

AdaptedKG — Structure-matched KG evidence for behavior denoising; offline computation, no KG needed at inference.

Career Path Recommendation — TCS uses LSTM unsupervised learning of career event embeddings with Levenshtein alignment for personalized career path recommendation.