type
status
date
slug
summary
tags
category
icon
password
priority
Weekly Overview
This week's recommendation systems research runs along three technical threads. First, Semantic ID-driven generative retrieval keeps gaining momentum. Spotify released two papers simultaneously — one deploys a SID system in production with A/B test results (new show discovery rate +14.3%), the other treats SID as a standalone modality unifying search, recommendation, and reasoning. Industrial SID systems have moved past "can this work?" into "how do we make it work better." Second, multimodal retrieval and representation compression: Apple delivered a production-grade unified retrieval architecture for text, images, and video; Aalto University distilled a 2B-parameter VLM into a 69M text encoder (50x latency reduction); POSTECH identified and fixed a modality collapse problem in VLM embedders for recommendation.
Third, information flow control in industrial ranking. Three papers from Alibaba and Meta attack the same problem from different angles — stop feeding all features and signals indiscriminately, and instead control when features enter interaction layers (deferred masking), select which behaviors deserve fine-grained interaction (core behavior selection), and standardize the distributional semantics of behavior signals (conditional debiasing). Meta's MBD framework stands out. It's deployed on two billion-user short-video platforms, reducing the correlation between watch time and video duration from 0.514 to 0.003.
Semantic ID and LLM-Driven Generative Retrieval
Spotify released two generative recommendation papers this week — one already live with A/B test results, the other treating SID as a standalone modality to unify search, recommendation, and reasoning. Add Amazon's work on retrieval model training efficiency, and the takeaway is clear: industrial SID systems are shifting from feasibility to optimization.
Deploying Semantic ID-based Generative Retrieval for Large-Scale Podcast Discovery at Spotify (2603.17540) — Spotify
Existing podcast recommendation relies heavily on long-term interaction patterns and struggles to capture short-term intent shifts. GLIDE reframes podcast recommendation as an instruction-following generation task built on Semantic IDs. The model uses a decoder-only Transformer backbone. It concatenates recent listening history with lightweight user context on the input side and injects long-term user embeddings via soft prompts — capturing both stable preferences and immediate intent. SIDs are obtained by semantic discretization of the podcast catalog, so every autoregressively generated token sequence maps to a real catalog entry. The online A/B test covered millions of users: non-habitual podcast stream plays rose 5.4%, new show discovery rate rose 14.3%, all within production latency and cost constraints.
Worth noting: GLIDE's dual-path design — soft prompt for long-term preferences, explicit input for short-term context — shares conceptual ground with Kuaishou's DAS (2508.10584) multi-view contrastive alignment and Meituan's DOS (2602.04460) user-item dual-stream framework. All three explicitly decompose different signal sources within a SID system. GLIDE's differentiator is the instruction-following paradigm: text prompts control generation behavior (recommend vs. explore), adding a layer of steerability beyond pure ID sequence modeling. Pinterest's PinRec (2504.10507) uses outcome-conditioned generation to balance saves and clicks; GLIDE achieves similar multi-objective control through instruction templates — different paths, same destination.
A Unified Language Model for Large Scale Search, Recommendation, and Reasoning — NEO (2603.17533) — Spotify
Current approaches to LLM-based recommendation either rely on external tool calls or remain confined to a single task. NEO's central idea: treat SID as an independent modality, interleaving SID tokens with natural language tokens in the same generation sequence. The implementation starts with staged alignment on a pretrained decoder-only LLM — teaching the model SID as a new "language" — followed by instruction tuning for multi-task support. Constrained decoding ensures generated item IDs always land within the catalog, while free-text output remains unrestricted. The catalog exceeds 10 million items across multiple media types. In offline experiments, NEO outperforms task-specific baselines on recommendation, search, and user understanding tasks, and demonstrates cross-task transfer.
NEO's "language-steerability" concept — using natural language prompts to control whether output is IDs, text, or a hybrid — extends IDGenRec's (2403.19021) approach of building item IDs from human-language tokens. IDGenRec focused on making IDs semantically meaningful; NEO goes further by enabling free switching between SID and natural language within one generation space. However, NEO only has offline validation — no online A/B test data — so production deployment remains unproven. For reference, Alibaba's URM (2502.03041) has already validated LLM-based generative retrieval in online ad serving, with a 3% lift on core metrics and latency in the tens of milliseconds. Whether NEO can hold latency at comparable scale remains unproven.
OPERA: Online Data Pruning for Efficient Retrieval Model Adaptation (2603.17205) — Amazon
Not all training pairs contribute equally in domain fine-tuning of dense retrievers. OPERA proposes two data pruning strategies. Static pruning (SP) keeps only high-similarity query-document pairs, lifting NDCG@10 by 0.5% but hurting recall — query diversity drops. Dynamic pruning (DP) resolves this trade-off: a two-stage adaptive mechanism adjusts sampling probabilities at both query and document granularity, prioritizing high-quality samples while retaining access to the full training set. Results across 8 datasets spanning 6 domains: DP lifts NDCG@10 by 1.9% and Recall@20 by 0.7%, with an average rank of 1.38. The more practical number: DP reaches comparable performance in under 50% of standard fine-tuning time. The method is architecture-agnostic — it works on LLM-based retrievers like Qwen3-Embedding as well.
These three papers reveal two diverging routes for industrial SID systems. One is the GLIDE approach — "SID + instruction following" — using LLM language capabilities to enhance recommendation controllability. The other is the NEO approach — "SID as modality" — pursuing full unification of search, recommendation, and reasoning. Whether these converge depends on whether unified models can match specialized systems on latency and online performance.
Multimodal Retrieval and Representation Compression
Three papers this week target the same core tension: multimodal understanding capability versus inference efficiency. From Apple's production-grade multimodal retrieval architecture, to Aalto University's extreme VLM compression, to POSTECH's fix for modality collapse in VLM embedders — the direction is clear: get the multimodal capabilities of large models into online systems at minimal cost.
AMES: Approximate Multi-modal Enterprise Search via Late Interaction Retrieval (2603.13537) — Apple
The core challenge in multimodal enterprise search: text, images, and video each require separate retrieval pipelines, and architectural complexity grows linearly with the number of modalities. AMES maps all three modalities into a shared multi-vector representation space. Text tokens, image patches, and video frames pass through the same encoder to produce vector sequences — cross-modal retrieval no longer needs modality-specific logic.
Retrieval runs in two stages. Stage one executes parallel ANN search per query token (Solr native KNN, numCandidates=250), then aggregates by document to get Top-M (M=12) approximate MaxSim scores. Stage two uses an accelerator for exact MaxSim re-ranking. The system integrates directly into Apache Solr, using parent-child document structure to store embeddings, with PyTorch batch-computing MaxSim on the client side. On the ViDoRe V3 English industrial subset, it achieves NDCG@10 of 58.1 with ColQwen3.0 as the encoder. The paper acknowledges it does not provide system-level latency and throughput benchmarks — a gap that weakens the production deployment case. But the engineering value of "zero-modification Solr integration" is real.
NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder (2603.12824) — Aalto University
The status quo in visual document retrieval (VDR): query and document sides share the same multi-billion-parameter VLM encoder. But queries are short text; only documents need visual understanding. NanoVDR exploits this asymmetry. The 2B Qwen3-VL teacher handles offline document indexing; the query side gets a distilled 69M DistilBERT student.
The choice of distillation objective matters substantially. The paper systematically compares six objective functions: pointwise cosine alignment, ranking loss, various weighted combinations of both, plus InfoNCE. The conclusion is definitive — pure cosine alignment consistently wins across three ViDoRe benchmark versions, reaching NDCG@5 of 82.2/61.4/44.1 on v1/v2/v3. InfoNCE performs worst, with only 30.0 on v3. This is counterintuitive: ranking losses underperform simple pointwise alignment in distillation.
The final NanoVDR-S-Multi model (69M parameters) achieves NDCG@5 of 61.9 on ViDoRe v2 and 46.5 on v3, retaining 95.1% of teacher quality. Compared to DSE-Qwen2 (2B), it has 32x fewer parameters and reduces CPU query latency from 2,539ms to 51ms — a 50x speedup. Total training cost: under 13 GPU hours on H200. Cross-lingual transfer is the main bottleneck: English retention hits 94.3%, but Portuguese drops to 75.6%. Adding machine-translated training data boosts Portuguese NDCG@5 by 9.3 points, narrowing the retention gap from 18.6pp to 2.7pp.
VLM2Rec: Resolving Modality Collapse in VLM Embedders for Multimodal Sequential Recommendation (2603.17450) — POSTECH
A hidden pitfall in using VLMs for recommendation embeddings: standard contrastive learning fine-tuning worsens modality collapse. The optimization process gets dominated by one modality, and representation quality for the other degrades. VLM2Rec builds on Qwen2.5-VL-3B with LoRA (rank=16, alpha=32) fine-tuning and introduces two targeted fixes.
Weak-modality Penalized Contrastive Learning (WPCL) is the key mechanism. It detects which modality contributes weaker gradients in the current batch and applies penalty weights to the weak modality, forcing the optimizer to attend to both modalities equally. The ablation data is definitive: removing WPCL drops NDCG@20 on the Beauty dataset from 0.4121 to 0.2592 — a 37% decline. Across four Amazon domains, VLM2Rec lifts Hit@10 by 12%–22% and NDCG@10 by 9%–32% over the strongest baselines. All experiments use public datasets; online validation is absent.
From NanoVDR's asymmetric distillation to AMES's unified late interaction architecture to VLM2Rec's modality-balanced training, these three papers trace the same path: start with a large VLM for high-quality multimodal representations, then compress, unify, or stabilize them into a deployable form.
Industrial Ranking and Feature Interaction Modeling
Three papers this week focus on two pain points in the ranking stage: granularity control of feature interactions and bias removal from behavior signals. Alibaba contributes two feature interaction modeling papers; Meta proposes a cross-dimensional behavior signal debiasing framework. All have online deployment validation.
Deferred is Better: A Framework for Multi-Granularity Deferred Interaction of Heterogeneous Features (2603.12586) — Alibaba
Ranking models typically feed all features into interaction layers at once. The problem: sparse features (like item ID) and dense features (like price) differ vastly in information density. Low-information features entering interaction too early inject noise and can even cause model collapse. MGDIN's core idea is "deferred introduction" — let high-information-density features build robust representations first, then unlock low-information-density features layer by layer.
Two steps. First, multi-granularity feature grouping: K groups with different window sizes partition raw features into subsets of more uniform information density. Window granularities are set at {32, 64, 96, 128}, with each group processing feature interactions in parallel. Second, hierarchical masking: across 3 layers, layer 1 activates only 33% of feature groups, layer 2 activates 66%, layer 3 activates 100%. Spatial complexity drops from standard attention's O(n²) to sum((n/g_h)²). On an industrial dataset with 7 billion interaction records, AUC reaches 0.6994 — a +0.54% improvement over the best baseline. Online A/B test: CTR +1.2%, with zero additional inference latency.
Bridging Sequential and Contextual Features with a Dual-View of Fine-grained Core-Behaviors and Global Interest-Distribution (2603.12578) — Alibaba
Traditional CTR models aggregate user behavior sequences into a single vector before interacting with contextual features. This aggregation discards behavioral detail. Letting each behavior interact directly with contextual features preserves information but scales quadratically with sequence length L, and irrelevant behaviors introduce noise that drowns out useful signals.
CDNet resolves this with a dual-view approach. Fine-grained view: cosine similarity selects top-k core behaviors (default k=16), and only these interact with contextual features — reducing complexity from O((L+N_f)²) to O((k+1+N_f)²). Coarse-grained view: the similarity range is split into 5 equal-width buckets, counting behaviors per bucket to construct a global interest distribution vector as compensation. On a Taobao dataset with 89 million records, AUC reaches 0.6388 — a +0.58% improvement over the best baseline. Online A/B test extended the behavior sequence length to 1,600 with 100 core behaviors: CTR +2.24% relative improvement, zero latency increase.
MBD: A Model-Based Debiasing Framework Across User, Content, and Model Dimensions (2603.14422) — Meta
Value models in recommendation ranking typically aggregate multiple behavior signals — watch time, loop rate, like rate, comment rate — into a single score. These signals carry inherent biases: watch time favors long videos, loop rate favors short videos, comment probability favors video over images. MBD asks: can biased signals be systematically converted to unbiased ones while preserving personalization?
MBD assumes behavior signals follow a Gaussian distribution. It conditions on feature subsets (e.g., video duration, user region) and estimates contextual mean μ and variance σ² directly within the MTML ranking model. Raw prediction p is standardized to RPS = (p - μ) / σ, interpretable as a percentile. The debiasing module plugs into the existing MTML model as a lightweight branch with gradient isolation to prevent interference with the main model. Extra compute overhead stays below 5%.
In offline validation, the correlation between watch time and video duration drops from 0.514 to 0.003 after MBD processing — duration bias is nearly eliminated. Online A/B tests span three scenarios: media duration debiasing yields watch time +0.198% and shares +0.44%; content format debiasing yields likes +0.421%; cold-start debiasing yields content ramp-up rate +0.190%. The traffic reallocation data is informative: 5–10 minute videos see only +0.13% more impressions but +0.73% more watch time — a 562% efficiency ratio — demonstrating that debiased systems allocate traffic more precisely to high-quality content. The framework is deployed on two billion-user short-video platforms.
These three papers point in the same direction: ranking model improvements are shifting from "stack a bigger model" to "control information flow more precisely." Whether it's controlling when features enter interaction (MGDIN's deferred masking), selecting which behaviors deserve fine-grained interaction (CDNet's core behavior selection), or standardizing the distributional semantics of behavior signals (MBD's conditional debiasing) — the underlying goal is the same: process the right information at the right granularity.
Directions to Watch
Instruction-following paradigm for Semantic ID. Spotify's GLIDE validates a path: upgrade SID systems from "generate item ID sequences" to "generate item IDs following natural language instructions." Recommendation controllability no longer depends on post-processing rules — it's internalized in the generation process. Spotify (GLIDE + NEO) has the deepest investment in this direction; Kuaishou (DAS, OneMall) and Meituan (DOS) are also pushing forward. For recommendation scenarios requiring multi-objective balancing — exploration vs. exploitation, diversity vs. relevance — the practical value is clear.
Asymmetric deployment of VLMs. NanoVDR's core insight is simple: queries and documents differ in complexity, so encoders should differ too. The 2B VLM handles offline document indexing only; the online query side runs a 69M text encoder at 50x lower latency. This asymmetric distillation approach applies to any recommendation system involving visual content — product image understanding, short-video thumbnail understanding, ad creative retrieval. Training cost is just 13 GPU hours. The engineering bar is low.
Distributional standardization of behavior signals. Meta's MBD framework offers an approach more general than traditional debiasing methods: instead of designing a bespoke correction for each bias type, it uses conditional distribution modeling to convert all behavior signals to percentiles. This lets biases across different dimensions — content duration, content format, user activity level — be handled with a single framework. Validated on two billion-user platforms. For any recommendation system using multi-signal fusion scoring, this direction is worth tracking.
Paper Roundup
Semantic ID and Generative Retrieval
GLIDE — Spotify reframes podcast recommendation as SID-based instruction-following generation; online A/B test: new show discovery rate +14.3%, non-habitual podcast stream plays +5.4%.
NEO — Spotify treats SID as a standalone modality unifying search, recommendation, and reasoning; offline validation on a 10M+ item catalog outperforms task-specific baselines.
OPERA — Amazon proposes an online data pruning framework for retrieval model fine-tuning; NDCG@10 +1.9% across 8 datasets, training time cut in half.
Multimodal Retrieval and Representation Compression
AMES — Apple proposes a unified multimodal late interaction retrieval architecture; zero-modification Solr integration, ViDoRe V3 NDCG@10 of 58.1.
NanoVDR — Aalto University distills a 2B VLM into a 69M text encoder; retains 95.1% quality, 50x CPU latency reduction, training cost just 13 GPU hours.
VLM2Rec — POSTECH resolves modality collapse in VLM embedders; Hit@10 up 12%–22% across four Amazon domains.
Industrial Ranking and Feature Interaction
MGDIN — Alibaba proposes a multi-granularity deferred interaction network; AUC +0.54% on 7 billion records, online CTR +1.2%.
CDNet — Alibaba combines core behavior selection with interest distribution compensation for CTR prediction; online CTR +2.24%, zero latency increase.
MBD — Meta proposes a cross-dimensional behavior signal debiasing framework; watch time–video duration correlation drops from 0.514 to 0.003, deployed on two billion-user platforms.
Other
Location Aware Embedding — Industry team proposes a location-aware embedding framework, jointly embedding queries and locations into a low-dimensional space for improved geographic targeting in search ads.
Shopping Companion — Memory-augmented LLM shopping agent combining long-term memory retrieval with shopping assistance; a lightweight model outperforms GPT-5 on a 1.2M product benchmark.
EASP — JD.com proposes an environment-aware search planning paradigm — lightweight retrieval probes capture environment snapshots before an LLM generates search plans; online A/B test lifts UCVR and GMV (specific percentages not disclosed in the paper).