Papers for

streaming service engineers

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Improving recommendations by adding user relationships to shallow autoencoders

Alternatives to Shallow Autoencoders for Collaborative Filtering

Abstract: Collaborative filtering (CF) is a cornerstone of recommender systems and plays a relevant role in many modern applications. CF uses user-item interaction data to discover future preferences. Although deep learning models have shown promise in CF, Embarrassingly Shallow Autoencoders for Sparse Data (EASE) has gained attention for its outstanding ranking accuracy provided by its closed-form solution. EASE relies primarily on relationships among items to fit a full-rank high-dimensional linear mapping. We hypothesize that this design limits its capacity to capture similar but not equivalent fine-grained user relationships, consequently limiting its recommendation accuracy. This paper introduces an alternative formulation based on EASE that takes advantage of user-user information. Furthermore, we propose a hybrid model that combines both user-user and item-item adjacency distributions. Our experiments reveal that the proposed models outperform EASE on well-known recommendation benchmarks, highlighting the significance of including the user alternative in shallow autoencoder studies for CF.

Wed 4 NovMachine LearningInformation Retrieval
The gist
Recommender systems suggest items like movies or products based on what users have liked before. A popular method, called EASE, uses item relationships but might miss subtle similarities between users. The authors propose new versions that also consider how users relate to each other, which helps make better recommendations. Their tests show these new approaches perform better than the original EASE method on common datasets.
Open → 2611.06610v1

LLM embeddings adapt progressively to improve item recommendations

PACERec: Progressive Adaptation of LLM Item Embeddings with Collaborative Evidence for Recommendation

Abstract: Large language model (LLM) embeddings have been increasingly adopted in sequential recommendation to encode static item attributes and alleviate the limitations of item ID embeddings under sparse interactions. However, these embeddings typically capture only static content similarity while neglecting sequential transitions, and since their adaptation quality relies on sparse per-item updates or global geometric calibration, existing methods still struggle to explicitly transfer preceding interaction contexts from frequent to long-tail items under unevenly distributed behavioral evidence, limiting recommendation effectiveness across the catalog. To address this challenge, we propose PACERec, an ID-free progressive adaptation framework that incorporates preceding interaction contexts into semantic representation adaptation by learning a shared velocity field from items with sufficient interactions via frequency-weighted flow matching. In this way, PACERec progressively guides static LLM embeddings toward collaborative behavioral targets while preserving original item semantics, overcoming sparse per-item updates to effectively generalize sequential transition patterns to long-tail items for both user modeling and candidate scoring. Extensive experiments on three real-world datasets across three sequential backbones demonstrate the superiority of our method over competitive baselines in both overall and long-tail recommendation.

Mon 2 NovInformation RetrievalArtificial Intelligence
The gist
Recommendation systems sometimes struggle to suggest items that people like less often interact with, partly because they rely on limited data. The authors propose a new method called PACERec that adjusts large language model embeddings by learning how users typically move from one item to another. This approach helps the system better understand and recommend less popular items while keeping the original item meanings. Tests show it improves recommendations overall and especially for items with fewer interactions.
Open → 2611.03068v1

Recommendation system improves personalization using multi-view temporal thinking

Think Globally, Generate Personally: Multi-View Thinking with Adaptive Generation for Generative Recommendation

Abstract: Semantic-ID-based (SID) generative recommendation (GR) represents each item as a sequence of discrete tokens and predicts the next item through autoregressive SID generation. However, existing methods face two challenges. First, insufficient shared context modeling for user sequences: full-history and recent behaviors provide overlapping yet complementary evidence, but modeling them within a single sequence may dilute current cues, whereas processing them through separate streams may fail to coordinate them. Second, insufficient personalized adaptation in progressive SID decisions: although decoder states evolve with generated prefixes, existing models typically use a shared output-side transformation across recommendation instances, without explicitly selecting a transformation suited to the current user need. To address these challenges, we propose Multi-View Thinking with Adaptive Generation (MV-Thinker), a framework that Thinks Globally and Generates Personally. First, Multi-View Temporal Thinking uses view-specific thinking queries and readers over overlapping full-history and recent-history views, coordinating both sources into complementary user-conditioned contexts before SID generation. Second, Hierarchical Residual Adaptation organizes lightweight adapters into nested residual paths before a shared vocabulary projection, adaptively transforming context-bearing decoder states for SID decisions. Extensive experiments on four public datasets demonstrate that MV-Thinker consistently outperforms strong discriminative and generative recommendation baselines across multiple ranking metrics. The implementation code and datasets are available on: https://anonymous.4open.science/r/MV-Thinker

Mon 2 NovArtificial Intelligence
The gist
Recommendation systems try to guess what you might like next, but they often struggle to use all your past actions properly. The authors found that looking at both your full history and your recent activities separately and then combining them helps make better guesses. They also designed a way to tailor the recommendation process to each user more precisely. Their approach, called MV-Thinker, uses these ideas to pick items more accurately across multiple tests.
Open → 2611.02463v1

Fair algorithms improve recommendations for users in overlapping groups

Fair Recommender Systems for Intersecting Groups

Abstract: Recommendation systems that optimize for the average user can fail to capture the preferences of diverse demographic groups. We consider a dynamic recommendation system, where arriving users can belong to different demographic groups. Importantly, these groups can intersect, and a user may belong to several groups. We first show that over time, any group-blind policy (which excludes all current-user group information) recommending $K \geq 1$ items incurs linear total regret under mild conditions. In turn, we introduce three group-aware algorithms based on the classical Upper Confidence Bound algorithm (UCB): i) Intersection-Fitting, which runs UCB on each intersection, but has a sublinear group- and intersection-level regret guarantee with an exponential dependence in the number of groups; and ii) two group-level heuristics, Inner-Outer UCB and Group-Item UCB which allocate different arms to different groups, and use a sleeping expert meta-learner to choose which active group's recommendation an intersection should inherit at any given time step. We provide experimental results on two datasets, Goodreads and BoardGameGeek, and highlight that our heuristics can be preferable to the intersection-fitting algorithm when intersection-level preferences are closely related to group-level preferences, but that intersection-fitting performs better when different intersections within a same group prefer different items.

Thu 29 OctMachine LearningComputers and SocietyComputer Science and Game Theory
The gist
Recommendation systems that show what users might like often focus on average preferences, which can miss the tastes of smaller or overlapping groups of people. The authors study systems where users belong to multiple intersecting groups and propose methods that consider group information to make fairer recommendations. They show that ignoring group membership leads to worse results over time and introduce new algorithms that adapt recommendations based on overlapping groups. Experiments on book and game recommendation data show benefits depending on how much individual groups differ in their preferences.
Open → 2610.40936v1

Improving item recommendations by modeling semantic and collaborative relations

REGER: Modeling inter-item Relations with Semantic IDs for Generative Recommendation

Abstract: Personalized recommendation depends on learning semantic and collaborative relations among items from online content and user interactions. Generative recommendation (GR) offers a promising paradigm by autoregressively generating discrete semantic IDs (SIDs) of items. However, randomly initialized SID token embeddings and GR models trained from scratch do not directly inherit the semantic relations inherent in items, while inter-item collaborative relations are learned largely implicitly through token-level attention and next-token supervision. To address this limitation, we propose REGER, a unified framework that explicitly incorporates inter-item relations into semantic pretraining, history encoding, and SID generation. We pretrain the GR model through generative semantic retrieval, enabling it to inherit inter-item semantic similarity by learning to generate semantic neighbors' SIDs. We combine global attention with local attention aligned with SID boundaries to capture historical context and directed transitions between adjacent items. We further develop a theoretical framework for hierarchical preference decomposition over the SID prefix tree, proving an exact factorization of collaborative preference distributions and a weighted decomposition of item-level KL divergence into prefix-conditional KL divergences. Building on this analysis, we distill a discriminative teacher's full conditional distributions over all legal next tokens at ground-truth prefixes along the target SID path, jointly optimizing this objective with cross-entropy. Extensive experiments on real-world datasets demonstrate consistent improvements over existing GR baselines and support the effectiveness of the proposed preference distillation.

Tue 27 OctInformation RetrievalArtificial Intelligence
The gist
Personalized recommendations often rely on understanding how items relate based on their meaning and user behavior. The authors found that current generative recommendation methods start without capturing these relationships well. They created a new approach called REGER that teaches the system to recognize and use connections between items, both in their meaning and in how users interact with them. Their method uses smarter ways to look at item history and includes a way to break down user preferences more precisely, helping the system suggest better next items. Experiments show this approach works better than earlier methods.
Open → 2610.37584v1

Semantic id based recommendation improves by focusing on one token level

LESSRec: Level-Specific History Modeling for Semantic ID-Based Recommendation

Abstract: Semantic ID-based recommendation represents each item as a sequence of semantic tokens obtained through residual quantization. This paradigm generates the next item's tokens autoregressively from coarse to fine, entailing each prediction attending to historical tokens at all semantic levels. However, our inference-time analysis reveals that historical tokens at the target-token's semantic level tend to be more informative, whereas other levels can introduce interference. Motivated by these findings, we propose LESSRec that separately encodes the historical token sequence at each semantic level and restricts attention over historical tokens to the sequence at the target token's semantic level during prediction. To complement these level-specific histories, LESSRec explicitly provides an item-level semantic sequence that represents each historical item's overall semantics in chronological order. The decoder accesses the level-aligned token sequence and the item-level semantic sequence through two separate cross-attention branches and fuses their contexts. We further introduce LESSRec+, which uses separate encoders for different semantic-level histories with the same total parameter count to reduce encoder computation. Extensive experiments across three public datasets and three tokenizers demonstrate the recommendation effectiveness of LESSRec and LESSRec+, with efficiency analyses highlighting the computational benefits of LESSRec+.

Tue 27 OctInformation Retrieval
The gist
Recommending items in apps often means predicting which item a person will want next, by analyzing the history of items they interacted with. This paper finds that looking at the history of tokens (word-like pieces representing item features) only at the same level as the item being predicted is more helpful than mixing all levels together. The authors propose a method called LESSRec that separately processes token histories by level and uses an additional summary of each item's overall meaning. LESSRec and its variant LESSRec+ show better recommendation quality and faster computation on multiple datasets.
Open → 2610.37286v1

Progressive image compression improves quality and efficiency with bound constraints

Bound-Conditioned Latent Inference for Progressive Image Compression

Abstract: Progressive image compression reconstructs images from prefixes of a single bitstream and therefore involves sequential latent inference under partial observations. Existing trit-plane codecs mainly treat decoded trits as discrete symbols for probability prediction, without fully exploiting their role as interval constraints on the underlying continuous latents. We propose bound-conditioned latent inference (BLI), a framework that reformulates progressive trit-plane coding as sequential latent inference under shrinking interval constraints. At each coding step, the trits decoded so far define lower and upper bounds for each latent element. BLI uses the center and width of these bounds, together with hyperprior information shared by the encoder and decoder, to jointly refine the Gaussian parameters $(μ,σ)$ for entropy modeling and the latent estimate $\hat{y}$ for reconstruction. As additional trits are decoded, the feasible intervals shrink monotonically, yielding tighter constraints for the next inference step and enabling flexible in-plane refinement schedules. Experiments on Kodak, CLIC, and JPEG-AI show that BLI achieves state-of-the-art rate-distortion performance among progressive image codecs, reducing BD-rate by $17.78\%$ over DPICT and by $7.96\%$ over CTC on Kodak, while replacing CTC's plane-specific predictors with a shared model that uses about $5.3\times$ fewer parameters and achieves lower latency than CTC.

Thu 22 OctComputer Vision and Pattern Recognition
The gist
Progressive image compression lets you see a low-quality version of an image that gradually improves as more data loads. The authors noticed current methods treated parts of the compressed data as simple codes, ignoring their ability to narrow down possible image details. They developed a new method called bound-conditioned latent inference (BLI), which uses these data parts as interval limits to better guess the image details step by step. This approach leads to clearer images with less data and speeds up decoding compared to existing techniques.
Open → 2610.30966v1

Popularity bias reduction improves relevance in recommendation systems

Mitigating Popularity Bias in Sequential Recommendations with Sparse Prior-Regularized Objective

Abstract: Popularity bias is a well-known challenge in recommender systems, often leading to the over-exposure of popular items and reduced catalog coverage. In this work, we propose PR-entmax, a sparse prior-regularized objective that makes popularity correction context-dependent. We derive our method from a generalized variational principle in which an item popularity prior is incorporated directly into a sparse probability mapping, yielding a popularity-aware active support: the set of items receiving non-zero probabilities and hence non-zero gradients is determined jointly by contextual relevance and item popularity. As a result, the prior affects not only item probabilities, but also which items participate in gradient updates during training. The method requires no architectural changes and can be integrated into existing training pipelines by modifying only the loss function. At inference time, the prior term is removed, and items are ranked by the learned contextual relevance alone. Experiments across multiple benchmark datasets and sequential recommendation architectures show that PR-entmax consistently reduces popularity bias while preserving or improving recommendation relevance, advancing the relevance-debiasing trade-off.

Tue 13 OctInformation Retrieval
The gist
Recommender systems often show popular items too much, making it hard for users to find less-known things. The authors created a new method called PR-entmax that uses a popularity-aware rule during training to balance popularity and relevance. This method changes which items get updated in the system, reducing the focus on popular items without making recommendations worse. Their tests show it works well on various datasets and models.
Open → 2610.17823v1

Perceptual audio coding improves sound quality by smarter bit allocation

Co-Adaptive Entropy Modeling and Rate-Distortion-Optimal Bit Allocation for Perceptual Audio Coding

Abstract: We present a perceptual audio coder in which the entropy-coding model and the bit allocation co-adapt across blocks without model side information. Mantissas are coded by a hybrid of bit packing and arithmetic coding with per-band probability models learned online. Encoder and decoder accumulate identical histograms from decoded blocks and advance the same model state. The model enters bit allocation directly: per-band rate--distortion frontiers combine measured perceptual distortion with entropy-coded rate estimates under the current model, and a dynamic program attains the exact minimum of total objective perceptual distortion for each block's budget under the model's rate estimates, with reservoir carry-over. Under a strict per-block rate, entropy-coding-aware bit allocation lowers this distortion (band noise-to-mask ratio in linear scale) by 15-24% relative to the same solver without it, and co-adaptation removes a further 3-4% (up to 15% on oboe) that a backward-adaptive coder alone cannot realize in one pass; a MUSHRA test at 80 kbps per channel confirms the ranking.

Mon 12 OctInformation Theory
The gist
Compressing audio often means choosing which parts to keep and which to simplify so it still sounds good. This paper presents a method where the models used to compress sound and the way bits are shared adapt together for each chunk of audio without extra info. The authors show that this approach reduces distortion and improves perceived sound quality compared to previous methods. They tested it with listening tests and found meaningful improvements at common streaming quality levels.
Open → 2610.17459v1

MPASCNet improves compressed music restoration with adaptive spectral compression

MPASCNet: Magnitude-Phase-Aware Adaptive Spectral Compression for Compressed Music Restoration

Abstract: Compressed music restoration aims to recover high-fidelity audio from lossy codecs such as MP3 by restoring missing spectral content and reducing coding artifacts. Recently, generative adversarial networks and band-split models have made progress in compressed music restoration. However, existing methods either operate without explicitly modeling phase or rely on predefined frequency partitions, which may limit restoration performance and the use of input-dependent frequency information. To address these issues, we propose MPASCNet, a Magnitude-Phase-Aware Adaptive Spectral Compression network. Instead of using fixed bands, MPASCNet combines magnitude and phase features to guide cross-attention, adaptively compressing full-resolution spectral features into compact tokens. These tokens are then processed by a spectro-temporal backbone to model spectral-temporal dependencies. Finally, separate decoder branches reconstruct the full-resolution magnitude and wrapped phase spectra. Experiments on MP3 restoration show that MPASCNet outperforms the Apollo and AEROMamba baselines across multiple restoration metrics while using only 6.49 M parameters. The code is available at https://github.com/xinghour/MPASCNet.

Mon 12 OctSound
The gist
Compressed music files like MP3s often lose sound quality because some audio details are removed. The authors developed MPASCNet, a computer model that looks at both the loudness and timing patterns of music to better guess and restore those missing parts. Unlike earlier methods, MPASCNet adapts to the input music instead of using fixed frequency groups, helping it recover cleaner, more natural sound. Tests showed it worked better than some past models while using less computer power.
Open → 2610.16416v1

Large language models outperform multi-token methods in personalized recommendations

Personalized LLM Recommendation with Semantic IDs: Autoregressive Beam Search vs. Multi-Token Prediction

Abstract: Large-scale personalization is shifting from embedding-table retrieval to LLM-based recommenders that generate the Semantic ID (SID) of the item to recommend, where each SID is a short sequence of discrete hierarchical content codes. A single instruction-tuned model can then unify retrieval, content understanding, and language, and recommend brand-new titles from their content alone using SID; then, how should the model generate the SID? Traditionally, models rely on a single language modeling head that generates tokens autoregressively, typically using constrained beam search. In contrast, an emerging body of work explores multi-token prediction (MTP), where dedicated heads predict all tokens in parallel, substantially reducing the decoding latency associated with beam search. We report an industry-scale, apples-to-apples comparison of these two generation paradigms: single-head autoregressive decoding vs. MTP. We introduce a new auxiliary ranking loss, defined over the joint SID score, to train the MTP objective. We evaluate the SID training objectives on both quality and training/inference cost, inside a multi-task LLM that jointly does personalized retrieval, content question-answering (QA), and instruction following over a large streaming-media catalog. Our experiments show single-head beam search consistently beats MTP on most product-critical metrics: ranking and content QA, and trains at roughly 2.6x lower dollar cost on Qwen3-8B scale; MTP wins only on raw inference throughput. Around this finding, we detail an efficient training and inference stack and report end-to-end gains over the previous non-SID baseline equipped with an auxiliary ranking head: +77% item cold-start ranking, +22% content-QA accuracy, +2.8x faster training, with language metric regression traced to SID vocabulary expansion. We address the language regression by reinforcement learning post training.

Sun 11 OctInformation RetrievalArtificial IntelligenceComputation and Language
The gist
Personalizing recommendations using large language models (LLMs) involves generating short codes that represent items. The authors compared two ways to produce these codes: making predictions one token at a time using beam search, or predicting all tokens simultaneously with a multi-token method. They found that the single-token beam search method performed better on recommendation accuracy and answering content questions, though generating predictions simultaneously was faster. They also improved training speed and addressed some language prediction setbacks with extra learning steps.
Open → 2610.15864v1

Pathwise alignment improves recommendations with sparse feedback

Pathwise Alignment for Collaborative Filtering: Bilateral Conditional Flow Regularization

Abstract: Collaborative filtering predicts user-item interactions from implicit feedback. We focus on matrix factorization (MF) trained with the Bayesian Personalized Ranking loss. This loss updates each user embedding only through that user's observed items and sampled negatives, so sparse feedback gives MF few constraints. To constrain each user through other users, we propose Pathwise Alignment for Collaborative Filtering (PACF), which trains MF with one network shared by all training pairs. At points between user and item embeddings, a flow matching term trains the network to predict the user-to-item step, reversed and scaled for sampled negatives. A trajectory consistency term pulls each user's predicted path to end at the observed item. Both terms update the embeddings, so the network is discarded after training, thereby maintaining the efficiency of MF at inference. We prove that a network with bounded sensitivity to its input point accurately fits users who share an item only if their embeddings are close, so these users constrain one another. Against eleven discriminative and generative baselines, PACF achieves the highest performance on all seven datasets. Its relative gains over MF are largest where feedback is sparsest. The shared network thus supplies the missing constraints through other users.

Sun 11 OctInformation Retrieval
The gist
Recommender systems predict what users might like based on their past choices, but this is hard when they have given very little feedback. The authors propose a new way to train recommendation models called Pathwise Alignment for Collaborative Filtering, which connects users who share some interests by learning paths through their preferences. This approach adds useful constraints from other users’ data without slowing down predictions. They show this method improves recommendation accuracy, especially when user feedback is sparse.
Open → 2610.15506v1

Neural audio coder reconstructs high quality sound at low bitrates

HE-NAC: High-Efficiency Neural Audio Coding via Hierarchical and Residual Multi-Band Quantization

Abstract: High-fidelity audio coding requires faithful reconstruction of perceptually important high-frequency components, which are often lost due to the aggressive temporal decimation and quantization of neural codecs. State-of-the-art neural audio coding methods rely on adversarial learning to hallucinate these missing components, yielding perceptually plausible but imprecise reconstructions; a limitation for high-fidelity music applications. As a remedy, we propose high-efficiency neural audio coding (HE-NAC), which provides a multi-band extension of the popular residual vector quantization (RVQ) scheme. HE-NAC introduces an effective training mechanism that harmonizes a bitrate-free blind upsampling process, residual learning to compensate it with minimal bitrate, and a multi-head decoder architecture for band-specific reconstruction. As a result, HE-NAC achieves a balanced trade-off, reliably reconstructing the core-band ($\leq$ 3 kHz) component while preserving overall sound fidelity. Subjective listening tests and objective evaluations demonstrate that HE-NAC outperforms both the DAC baseline and MPEG HE-AAC at equal or lower bitrates.

Sun 11 OctSound
The gist
Encoding audio into small digital files often loses important high-frequency sounds, making music less clear. The authors developed a new neural audio coder that splits sounds into different frequency bands and fixes missing details with smart learning techniques. This method keeps important low-frequency sounds while better preserving the overall quality of the music. Tests show it works better than some common audio compression standards at similar or lower file sizes.
Open → 2610.15243v1

Shared update scheme improves personalized predictions and content generation

Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

Abstract: A user's movie, news, and dialogue histories differ in their native actions and outputs, yet each interaction supplies evidence that can update user memory. We study whether these histories can train one reusable update mechanism. An action-on-item schema pairs a mapped interaction role with a content embedding, allowing shared update parameters to operate on separate user states. We establish invariance to native relabeling, bounded state changes under item-embedding perturbations, and a pooled-training bound under explicit compatibility conditions. The Multi-Timescale State Hypothesis (MTSH) specifies how this evidence enters, persists, and is consumed; PerTIDE implements it with action gating, three state-space traces, fusion, and command-conditioned readout. On PENS, the same history encoder supports both next-news prediction and personalized headline generation. In a controlled PENS-to-MovieLens experiment, a frozen source-trained core exceeds an identically structured random core by 15.23 MRR points after fitting the same target consumer. On MIND, PerTIDE retains a 4.12-point MRR advantage over a same-input three-branch state-space control. Action, readout, and trace interventions identify complementary contributions to these gains. Together, the theory and experiments support learning history updates across compatible sources and reusing them through predictive and generative consumers.

Thu 1 OctMachine Learning
The gist
People interact with many kinds of content like movies, news, and chats, each having different actions and results. The authors studied if a single method can update a user’s memory from all these different interactions. They created a system that treats actions linked to content in a common way and showed it works well across different tasks like predicting news or creating personalized headlines. Their method can also transfer what it learns from one type of content to another, improving recommendations even when trained elsewhere.
Open → 2610.01375v1

Selective routing improves novelty in personalized recommendations

Routing Between Generative and Collaborative User Profiles: A Serving-Time Gate for Controllable Novelty

Abstract: Large language models (LLMs) enable rich semantic user profiles for recommendation, but such profiles are more expensive to generate and are not necessarily desirable to deploy uniformly. We study whether LLM-generated profiles can instead be invoked selectively within a production recommendation pipeline. Using a real-world streaming dataset covering movies, TV shows, and sports content, we train a serving-time routing gate that assigns each user to either a collaborative sequential recommendation model or a recommendation model driven by an LLM-generated profile. The gate uses only serving-time features and learns to identify users for whom profile-based routing can increase Novelty@10 while preserving ranking relevance. A routing threshold controls how aggressively users are sent to the generative model, exposing a tunable novelty--relevance trade-off. At an overall NDCG-loss budget of 5\%, the learned gate increases Novelty@10 by 6.5\% while routing 12.5\% of users, outperforming simple heuristic and random routing policies at comparable relevance cost. These results show that LLM-generated user profiles can serve as a controllable complement to collaborative recommendation, while results with non-generative semantic profiles indicate that the benefit stems from selective routing rather than LLM generation alone.

Wed 30 SeptInformation Retrieval
The gist
Making recommendations more interesting without losing accuracy is tricky. The authors studied how to use smart user profiles made by large language models (LLMs) only when it really helps. They built a system that decides in real time if a user should see recommendations based on an LLM profile or a usual collaborative method. This system boosts the chance of showing new content without making recommendations worse overall.
Open → 2609.39043v1

Sol-Refiner speeds up high-resolution video refinement with one step

SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video

Abstract: High-resolution video generation is expensive, as its cost grows rapidly with the number of spatiotemporal tokens. A practical alternative first generates a lower-resolution video and then applies a refiner, but conventional multi-step refinement introduces a second sampling bottleneck. We present SoL-Refiner, a one-step video refiner that transforms low-resolution model outputs into 4K videos with a single denoising step. Our three-stage recipe combines high-resolution continual training, reinforcement learning (RL) post-training, and a final one-step distillation. We introduce Refiner-Bench, a video refinement benchmark constructed from the outputs of different video generators, and use a shared-input protocol to compare refiners at approximately 2K output resolution. At 2K, the one-step SoL-Refiner outperforms all external refiners on the VBench and UniPercept averages, while at $3840\!\times\!2176$ it improves both metrics over the three-step LTX-2.3 Refiner. With the complete acceleration stack, SoL-Refiner achieves an $8.91\times$ speedup in refinement latency over the same baseline in our 2K latency setting.

Tue 29 SeptComputer Vision and Pattern Recognition
The gist
Making high-resolution videos takes a lot of computer power. The authors found a way to first make a lower-quality video and quickly improve it to very high quality in just one step. They trained their method carefully and tested it against others, showing it is both faster and better at refining videos. This approach can make producing detailed videos more efficient and quicker.
Open → 2609.37969v1

OmniRoute improves efficiency in multimodal audio visual models

OmniRoute: Mapping Temporal Semantic Evidence to Audio-Visual Token Budgets for Efficient Omnimodal Large Language Models

Abstract: Omnimodal large language models (Omni-LLMs) encode audio and visual streams into temporally interleaved token sequences for multimodal reasoning. However, processing long audio-visual token sequences incurs substantial prefill costs. Existing compression methods have made progress, but often overlook temporal changes in audio-visual semantic relevance. Motivated by temporal variation and local continuity, we propose OmniRoute, a training-free, two-stage compression framework. First, Temporal Evidence-Guided Budgeting (TEGB) derives chunk-wise modality preferences and initial leading-modality budgets from semantic relevance and local content variation. Second, Budget-Constrained Semantic Compression (BCSC) compresses the leading modality and then calibrates the follower's retention target using the actual retained fraction. For video, it combines spatiotemporal grouping with query-guided selection; for audio, it selects tokens based on encoder attention and query relevance, then merges residual tokens into context anchors under visual guidance. Experiments on four representative benchmarks demonstrate a better trade-off between inference efficiency and performance than competitive baselines. The code and interface will be released to facilitate further research.

Tue 29 SeptComputer Vision and Pattern Recognition
The gist
Processing both audio and video together in language models takes a lot of computing power, especially when handling long clips. The authors designed OmniRoute to reduce this workload by smartly choosing what parts of the audio and video to keep based on how important they are over time. It works in two steps: first, it figures out which modality (audio or video) matters more at different times, then it compresses that data accordingly while adjusting the other. This results in faster processing without losing much understanding.
Open → 2609.37052v1

SPADE metric measures truly surprising recommendations beyond popularity and similarity

SPADE: Escaping the Popularity-Similarity Frontier to Measure Serendipitous Recommendations

Abstract: Recommender systems engineer serendipity to foster active exploration and break predictable consumption cycles. The problem with existing offline beyond-accuracy metrics is that they often either isolate historical similarity or global popularity. We aim to design an evaluation metric that examines similarity, popularity, and actual user relevance. To achieve this, we introduce SPADE (Serendipitous Pareto Distance Evaluation). SPADE maps all items into a two-dimensional space to directly calculate a user-specific Pareto frontier of maximally popular and historically similar items. The final serendipity score is then computed by averaging the minimum Euclidean distance from this boundary strictly for the correctly recommended test-set items. Evaluating SPADE across five datasets and five baseline algorithms confirms its effectiveness; our results show that the metric successfully prevents algorithms from exploiting beyond-accuracy measures with irrelevant or non-personalized recommendations, reliably isolating serendipitous discoveries.

Fri 25 SeptInformation RetrievalArtificial IntelligenceMachine Learning
The gist
Recommender systems often suggest items based on popularity or similarity, which can make recommendations predictable. The authors created SPADE, a new way to evaluate how surprising or serendipitous recommendations are by considering popularity, similarity, and what the user actually likes. SPADE works by mapping items on a chart and measuring how far recommended items are from popular or similar items that a user already knows. Testing this across several datasets showed SPADE reliably identifies recommendations that are both relevant and genuinely unexpected.
Open → 2609.31164v1

Latent semantic fusion improves sequential recommendation accuracy

LSF-SR: Latent Semantic Fusion for Sequential Recommendation via Flow-based Conditional Variational Autoencoders

Abstract: Sequential recommendation aims to predict users' future interests from their historical interactions. Although Large Language Models (LLMs) capture rich item semantics, existing methods often struggle to align collaborative signals with textual semantic knowledge. As a result, the learned item representations fail to capture the complementary strengths of both signals, leading to suboptimal recommendation quality. To address this limitation, we propose Latent Semantic Fusion for Sequential Recommendation via Flow-based Conditional Variational Autoencoders (LSF-SR), a novel framework that uses a Conditional Variational Autoencoder (CVAE) with Normalizing Flows to fuse item ID embeddings and LLM-generated semantic signals. At the core of LSF-SR is a conditional fusion module augmented with planar or radial flows. This module learns a flexible latent space that encourages items with similar semantic profiles to cluster together within the latent manifold. Through extensive experiments on five public benchmark datasets, we demonstrate that LSF-SR consistently outperforms state-of-the-art baselines, achieving gains of up to 12.98% and 14.13% in Recall@20 and NDCG@20, respectively.

Thu 24 SeptInformation Retrieval
The gist
Sequential recommendation systems try to guess what items you might like next based on what you liked before. The authors found that combining different types of information about items, like their unique IDs and the meanings captured by text-based language models, can be tricky. They developed a new method called LSF-SR that uses a special kind of machine learning model to blend these data types more effectively. This approach helps group similar items together in a better way, which makes the recommendations more accurate.
Open → 2609.29815v1

Semantic user profiling methods compared for streaming recommendations

When LLM-Based User Profiling Adds Value in Production Streaming Recommendation

Abstract: Personalized recommendation depends critically on how user representations are constructed from historical behavior. Two paradigms have emerged for constructing semantic user profiles in content-based recommendation. First, aggregate methods derive user representations as numerical aggregates of semantic item embeddings. Second, LLM-based methods generate natural-language summaries of user preferences and encode them through a text encoder. Each paradigm can be combined with temporal disentanglement of recent versus historical behavior. LLM-based profile generation is significantly more expensive than aggregate approaches, raising the question of when this additional cost is justified. We present a systematic comparison of four semantic user-profiling strategies, factorially crossed across representation type and temporal handling, evaluated on a real-world production dataset. The comparison reveals how these strategies differ across user behavior types, across both accuracy and beyond-accuracy dimensions of recommendation quality, and across the temporal-window setting that governs the disentanglement.

Wed 23 SeptInformation Retrieval
The gist
Online services often suggest content based on what users liked before, but creating user profiles from past actions varies in complexity and cost. The authors compare two main ways to build these profiles: one that averages numerical representations of items, and another that uses large language models (LLMs) to write natural descriptions of user interests. They tested these methods on a real streaming service dataset, looking at how well each works depending on the user’s behavior and the timing of actions. Their findings help decide when the expensive LLM-based profiling is worth using compared to simpler methods.
Open → 2609.27183v1

Physics based system forecasts LEO satellite internet quality anywhere

Reading the Sky to Forecast the Ground: Physics-Informed Link-State Forecasting for LEO Networks at Any Location

Abstract: In this paper, we introduce Gnomon, a physics-informed system that forecasts user-perceived low-Earth-orbit (LEO) downlink throughput, uplink throughput, and round-trip time (RTT) under different levels of trace availability. Gnomon's physics layer reconstructs the serving geometry and four-leg bent-pipe attenuation from public weather, orbital, routing, and licensing data. Based on what is available, Gnomon conditions on the target terminal's own history (Mode 1), measurements from nearby publicly reachable dishes (Mode 2), or the physical covariates alone (Mode 3) to predict the link state: Modes 1 and 2 share a fine-tuned time-series foundation model, while Mode 3 uses a compact boosted-tree estimator. All three modes expose a common output interface and can be selected without retraining. We evaluate Gnomon using 8,260 minutes of 1 Hz measurements collected at nine sites across five states in the U.S. We train on three sites and hold out the remaining six sites and their serving beams. On these unseen sites, the own-trace mode reduces downlink-throughput and RTT prediction error by 17% and 11% relative to the strongest published baseline and, to our knowledge, provides the first LEO uplink forecasts. The neighbor-trace mode requires no on-site hardware, while the covariate-only mode reduces downlink-throughput and RTT error by 24.6% and 78.8% relative to the only prior covariate-only forecaster. Moreover, experiments show that Gnomon provides calibrated quantile bands and improves adaptive-bitrate streaming driven over real TCP flows on replayed Starlink links.

Tue 22 SeptNetworking and Internet Architecture
The gist
Low-Earth-orbit (LEO) satellites provide internet, but their connection quality can vary with weather and location, making it hard to predict. This paper presents Gnomon, a system that uses physics and public data sources to forecast download speed, upload speed, and latency for satellite internet users, even with limited local measurements. Gnomon offers three different ways to predict network quality based on what data is available, improving accuracy on new or unseen locations. It also helps streaming video perform better by adjusting quality based on these forecasts.
Open → 2609.26696v1

Feature decorrelation improves balance in sequential item recommendations

A Redundancy Reduction Approach for Controllable Sequential Recommendations

Abstract: Sequential recommendation must operate under long-tailed item distributions and popularity-driven concentration, often forcing practitioners to trade short-list accuracy against long-tail exposure. In this work, we study feature decorrelation as a mechanism for shaping representation geometry in dot-product sequential recommenders, and analyze how this, in turn, affects popularity-driven concentration. We propose a decorrelation-regularized training framework that augments next-item prediction with an auxiliary redundancy-reduction term, and instantiate it with BT-SR, which uses the Barlow Twins objective. To form label-consistent positive pairs without synthetic corruptions, we pair user histories that share the same next-item target. Beyond accuracy, we provide a geometric analysis showing how decorrelation suppresses shared low-rank directions in the user representation space that can give popular items a global scoring advantage, and we introduce a bucket-based alignment concentration metric to quantify this effect. Experiments on five public benchmarks show that BT-SR consistently improves next-item ranking quality, while the decorrelation strength acts as a simple control knob that reallocates accuracy across head and tail items, enabling accuracy-exposure trade-offs. Our analysis also reveals that the impact on head-vs-tail exposure differs across datasets, reflecting interactions between decorrelation and data temporal structure.

Sun 20 SeptInformation Retrieval
The gist
Items people see recommended often favor popular choices, making it hard for less-known items to appear. The authors studied a new way to reduce overlap in recommendation features, which helps spread attention more evenly across popular and less popular items. They created a training method that encourages this feature variety by comparing user histories with similar next-item interests. This method improved recommendation rankings and allows control over how much focus is given to popular versus less popular items.
Open → 2609.23849v1