From Hidden Profiles to Governable Personalization: Recommender Systems in the Age of LLM Agents

Personalization has traditionally depended on platform-specific user models that are optimized for prediction but remain largely inaccessible to the people they describe. As LLM-based assistants increasingly mediate search, shopping, travel, and content access, this arrangement may be giving way to a new personalization stack in which user representation is no longer confined to isolated platforms. In this paper, we argue that the key issue is not simply that large language models can enhance recommendation quality, but that they reconfigure where and how user representations are produced, exposed, and acted upon. We propose a shift from hidden platform profiling toward governable personalization, where user representations may become more inspectable, revisable, portable, and consequential across services. Building on this view, we identify five research fronts for recommender systems: transparent yet privacy-preserving user modeling, intent translation and alignment, cross-domain representation and memory design, trustworthy commercialization in assistant-mediated environments, and operational mechanisms for ownership, access, and accountability. We position these not as isolated technical challenges, but as interconnected design problems created by the emergence of LLM agents as intermediaries between users and digital platforms. We argue that the future of recommender systems will depend not only on better inference, but on building personalization systems that users can meaningfully understand, shape, and govern.

Transparent and Controllable Recommendation Filtering via Multimodal Multi-Agent Collaboration

While personalized recommender systems excel at content discovery, they frequently expose users to undesirable or discomforting information, highlighting the critical need for user-centric filtering tools. Current methods leveraging Large Language Models (LLMs) struggle with two major bottlenecks: they lack multimodal awareness to identify visually inappropriate content, and they are highly prone to "over-association" -- incorrectly generalizing a user's specific dislike (e.g., anxiety-inducing marketing) to block benign, educational materials. These unconstrained hallucinations lead to a high volume of false positives, ultimately undermining user agency. To overcome these challenges, we introduce a novel framework that integrates end-to-cloud collaboration, multimodal perception, and multi-agent orchestration. Our system employs a fact-grounded adjudication pipeline to eliminate inferential hallucinations. Furthermore, it constructs a dynamic, two-tier preference graph that allows for explicit, human-in-the-loop modifications (via Delta-adjustments), explicitly preventing the algorithm from catastrophically forgetting fine-grained user intents. Evaluated on an adversarial dataset comprising 473 highly confusing samples, the proposed architecture effectively curbed over-association, decreasing the false positive rate by 74.3% and achieving nearly twice the F1-Score of traditional text-only baselines. Additionally, a 7-day longitudinal field study with 19 participants demonstrated robust intent alignment and enhanced governance efficiency. User feedback confirmed that the framework drastically improves algorithmic transparency, rebuilds user control, and alleviates the fear of missing out (FOMO), paving the way for transparent human-AI co-governance in personalized feeds.

Drift-Aware Continual Tokenization for Generative Recommendation

Generative recommendation commonly adopts a two-stage pipeline in which a learnable tokenizer maps items to discrete token sequences (i.e. identifiers) and an autoregressive generative recommender model (GRM) performs prediction based on these identifiers. Recent tokenizers further incorporate collaborative signals so that items with similar user-behavior patterns receive similar codes, substantially improving recommendation quality. However, real-world environments evolve continuously: new items cause identifier collision and shifts, while new interactions induce collaborative drift in existing items (e.g., changing co-occurrence patterns and popularity). Fully retraining both tokenizer and GRM is often prohibitively expensive, yet naively fine-tuning the tokenizer can alter token sequences for the majority of existing items, undermining the GRM's learned token-embedding alignment. To balance plasticity and stability for collaborative tokenizers, we propose DACT, a Drift-Aware Continual Tokenization framework with two stages: (i) tokenizer fine-tuning, augmented with a jointly trained Collaborative Drift Identification Module (CDIM) that outputs item-level drift confidence and enables differentiated optimization for drifting and stationary items; and (ii) hierarchical code reassignment using a relaxed-to-strict strategy to update token sequences while limiting unnecessary changes. Experiments on three real-world datasets with two representative GRMs show that DACT consistently achieves better performance than baselines, demonstrating effective adaptation to collaborative evolution with reduced disruption to prior knowledge. Our implementation is publicly available at https://github.com/HomesAmaranta/DACT for reproducibility.

The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training

Large language models trained on natural language exhibit pronounced anisotropy: a small number of directions concentrate disproportionate energy, while the remaining dimensions form a broad semantic tail. In low-bit training regimes, this geometry becomes numerically unstable. Because blockwise quantization scales are determined by extreme elementwise magnitudes, dominant directions stretch the dynamic range, compressing long-tail semantic variation into narrow numerical bins. We show that this instability is primarily driven by a coherent rank-one mean bias, which constitutes the dominant component of spectral anisotropy in LLM representations. This mean component emerges systematically across layers and training stages and accounts for the majority of extreme activation magnitudes, making it the principal driver of dynamic-range inflation under low precision. Crucially, because the dominant instability is rank-one, it can be eliminated through a simple source-level mean-subtraction operation. This bias-centric conditioning recovers most of the stability benefits of SVD-based spectral methods while requiring only reduction operations and standard quantization kernels. Empirical results on FP4 (W4A4G4) training show that mean removal substantially narrows the loss gap to BF16 and restores downstream performance, providing a hardware-efficient path to stable low-bit LLM training.

Power Echoes: Investigating Moderation Biases in Online Power-Asymmetric Conflicts

Online power-asymmetric conflicts are prevalent, and most platforms rely on human moderators to conduct moderation currently. Previous studies have been continuously focusing on investigating human moderation biases in different scenarios, while moderation biases under power-asymmetric conflicts remain unexplored. Therefore, we aim to investigate the types of power-related biases human moderators exhibit in power-asymmetric conflict moderation (RQ1) and further explore the influence of AI's suggestions on these biases (RQ2). For this goal, we conducted a mixed design experiment with 50 participants by leveraging the real conflicts between consumers and merchants as a scenario. Results suggest several biases towards supporting the powerful party within these two moderation modes. AI assistance alleviates most biases of human moderation, but also amplifies a few. Based on these results, we propose several insights into future research on human moderation and human-AI collaborative moderation systems for power-asymmetric conflicts.

Multi-Head Attention as a Source of Catastrophic Forgetting in MoE Transformers

Mixture-of-Experts (MoE) architectures are often considered a natural fit for continual learning because sparse routing should localize updates and reduce interference, yet MoE Transformers still forget substantially even with sparse, well-balanced expert utilization. We attribute this gap to a pre-routing bottleneck: multi-head attention concatenates head-specific signals into a single post-attention router input, forcing routing to act on co-occurring feature compositions rather than separable head channels. We show that this router input simultaneously encodes multiple separately decodable semantic and structural factors with uneven head support, and that different feature compositions induce weakly aligned parameter-gradient directions; as a result, routing maps many distinct compositions to the same route. We quantify this collision effect via a route-wise effective composition number $N_{eff}$ and find that higher $N_{eff}$ is associated with larger old-task loss increases after continual training. Motivated by these findings, we propose MH-MoE, which performs head-wise routing over sub-representations to increase routing granularity and reduce composition collisions. On TRACE with Qwen3-0.6B/8B, MH-MoE effectively mitigates forgetting, reducing BWT on Qwen3-0.6B from 11.2% (LoRAMoE) to 4.5%.

RQ-GMM: Residual Quantized Gaussian Mixture Model for Multimodal Semantic Discretization in CTR Prediction

Multimodal content is crucial for click-through rate (CTR) prediction. However, directly incorporating continuous embeddings from pre-trained models into CTR models yields suboptimal results due to misaligned optimization objectives and convergence speed inconsistency during joint training. Discretizing embeddings into semantic IDs before feeding them into CTR models offers a more effective solution, yet existing methods suffer from limited codebook utilization, reconstruction accuracy, and semantic discriminability. We propose RQ-GMM (Residual Quantized Gaussian Mixture Model), which introduces probabilistic modeling to better capture the statistical structure of multimodal embedding spaces. Through Gaussian Mixture Models combined with residual quantization, RQ-GMM achieves superior codebook utilization and reconstruction accuracy. Experiments on public datasets and online A/B tests on a large-scale short-video platform serving hundreds of millions of users demonstrate substantial improvements: RQ-GMM yields a 1.502% gain in Advertiser Value over strong baselines. The method has been fully deployed, serving daily recommendations for hundreds of millions of users.

SD-MoE: Spectral Decomposition for Effective Expert Specialization

Mixture-of-Experts (MoE) architectures scale Large Language Models via expert specialization induced by conditional computation. In practice, however, expert specialization often fails: some experts become functionally similar, while others functioning as de facto shared experts, limiting the effective capacity and model performance. In this work, we analysis from a spectral perspective on parameter and gradient spaces, uncover that (1) experts share highly overlapping dominant spectral components in their parameters, (2) dominant gradient subspaces are strongly aligned across experts, driven by ubiquitous low-rank structure in human corpus, and (3) gating mechanisms preferentially route inputs along these dominant directions, further limiting specialization. To address this, we propose Spectral-Decoupled MoE (SD-MoE), which decomposes both parameter and gradient in the spectral space. SD-MoE improves performance across downstream tasks, enables effective expert specialization, incurring minimal additional computation, and can be seamlessly integrated into a wide range of existing MoE architectures, including Qwen and DeepSeek.

Distribution-Aware End-to-End Embedding for Streaming Numerical Features in Click-Through Rate Prediction

This paper explores effective numerical feature embedding for Click-Through Rate prediction in streaming environments. Conventional static binning methods rely on offline statistics of numerical distributions; however, this inherently two-stage process often triggers semantic drift during bin boundary updates. While neural embedding methods enable end-to-end learning, they often discard explicit distributional information. Integrating such information end-to-end is challenging because streaming features often violate the i.i.d. assumption, precluding unbiased estimation of the population distribution via the expectation of order statistics. Furthermore, the critical context dependency of numerical distributions is often neglected. To this end, we propose DAES, an end-to-end framework designed to tackle numerical feature embedding in streaming training scenarios by integrating distributional information with an adaptive modulation mechanism. Specifically, we introduce an efficient reservoir-sampling-based distribution estimation method and two field-aware distribution modulation strategies to capture streaming distributions and field-dependent semantics. DAES significantly outperforms existing approaches as demonstrated by extensive offline and online experiments and has been fully deployed on a leading short-video platform with hundreds of millions of daily active users.

Feature-Indexed Federated Recommendation with Residual-Quantized Codebooks

Federated recommendation provides a privacy-preserving solution for training recommender systems without centralizing user interactions. However, existing methods follow an ID-indexed communication paradigm that transmit whole item embeddings between clients and the server, which has three major limitations: 1) consumes uncontrollable communication resources, 2) the uploaded item information cannot generalize to related non-interacted items, and 3) is sensitive to client noisy feedback. To solve these problems, it is necessary to fundamentally change the existing ID-indexed communication paradigm. Therefore, we propose a feature-indexed communication paradigm that transmits feature code embeddings as codebooks rather than raw item embeddings. Building on this paradigm, we present RQFedRec, which assigns each item a list of discrete code IDs via Residual Quantization (RQ)-Kmeans. Each client generates and trains code embeddings as codebooks based on discrete code IDs provided by the server, and the server collects and aggregates these codebooks rather than item embeddings. This design makes communication controllable since the codebooks could cover all items, enabling updates to propagate across related items in same code ID. In addition, since code embedding represents many items, which is more robust to a single noisy item. To jointly capture semantic and collaborative information, RQFedRec further adopts a collaborative-semantic dual-channel aggregation with a curriculum strategy that emphasizes semantic codes early and gradually increases the contribution of collaborative codes over training. Extensive experiments on real-world datasets demonstrate that RQFedRec consistently outperforms state-of-the-art federated recommendation baselines while significantly reducing communication overhead.