arXiv subjects

Lichan Hong

Publications and source records attributed to Lichan Hong.

At least 19 recordsLinked to original sources

arXiv2026

Scaffolding Minds: Optimizing Latent Visual Target Representations for Multimodal Reasoning

Latent reasoning has advanced multimodal reasoning through a two-stage training paradigm: (1) a helper image is encoded into latent tokens to teach visual chain-of-thought during a supervised fine-tuning (SFT) stage, and (2) these latent tokens are further refined with reward feedback during a reinforcement learning (RL) stage. In this paper, we identify two key limitations of this framework, one in each stage. First, the SFT stage typically relies on an off-the-shelf vision encoder to encode the helper image, yielding suboptimal latent representations that may not be well aligned with the downstream reasoning task. Second, existing RL methods treat the latent component only through deterministic regularization, which constrains policy drift but does not create alternative latent trajectories for exploration. To address these limitations, we propose Scaffolding Minds. Our approach learns a dedicated scaffolding encoder that provides an optimized target in latent space, and learns both the mean and variance of the RL sampler. We further show that these two improvements are complementary, together yielding substantial gains over strong baselines. Empirically, our method improves over the strongest latent-reasoning baseline by +9.5% on FrozenLake spatial planning, with the gain widening to +19% at 32x32 grid map, and by +5.2% on average across nine visual-centric reasoning benchmarks.

cs.CV↗

arXiv2026

Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in recommendation systems

Large-scale recommendation systems face "Memory Wall" bottlenecks due to massive, dense embedding tables. While generative retrieval uses discrete tokens for IDs, high-dimensional context still relies on inefficient dense formats. Inspired by computer vision data compression, we propose Dual-purpose Semantic IDs to achieve LLM-level I/O efficiency. Our methodology uses hierarchical quantization to condense continuous embeddings into discrete Semantic IDs performing two concurrent roles: (1) Collaborative Identity: modeling user-item interactions via learnable embedding table; and (2) Content Reconstruction: using a lightweight Semantic Decoder for on-the-fly embedding approximation. This approach replaces massive vector storage with on-demand reconstruction, reducing system overhead and data footprints. We demonstrate the efficacy of our framework through offline evaluations and successful online deployment in production-scale ranking and retrieval systems at a major video sharing platform, showing that discrete tokens are indeed all you need for highly efficient, content-rich recommendation.

cs.IR↗

arXiv2026

TokenMinds: Pretrained User Tokens and Embeddings for User Understanding in Large Recommender Systems

User modeling in industrial recommender systems typically produces dense embeddings, which suffer from representational constraints inherent to fixed-dimensional vectors. An emerging alternative for discrete user representation -- using LLMs to generate text-based user tokens -- captures topical co-occurrences rather than deep sequential behavior dynamics and produces outputs that are difficult to ground to item attributes. Meanwhile, Semantic ID (SID) based item tokenization has proven effective for improving generalization in generative recommendation, yet discrete SID-based representations for users remain largely unexplored. We propose TokenMinds, an industrial-scale system that extends the PLUM framework from item retrieval to user modeling, generating both discrete SID-based user tokens and dense user embeddings via an encoder-decoder architecture adapted from pre-trained LLMs. This dual-output design provides the complementary benefits of discrete, semantically grounded user representations while maintaining compatibility with existing downstream models that rely on dense embeddings. Additionally, the shared SID vocabulary naturally extends to cross-scenario modeling: by unifying long-form and short-form video behaviors into a single model, we substantially reduce training and serving costs. We validate TokenMinds through extensive offline experiments and live launches on multiple YouTube surfaces, served on full user traffic (billions of users) via an asynchronous infrastructure that decouples representation generation from downstream scoring. Focusing on ranking as the primary downstream use case, our results confirm the practical viability of SID-based user tokens at industrial scale and demonstrate that tokens and dense embeddings provide complementary value across different production ranking systems.

cs.IR↗

arXiv2026

Token Factory: Efficiently Integrating Diverse Signals into Large Recommendation Models

Large Recommendation Models (LRMs) have demonstrated promising capabilities in industry-scale recommendation tasks. However, holistically integrating traditional signals into these transformer-based architectures effectively and efficiently remains a major challenge. Conventional approaches that "textualize" these signals directly or create discrete item representations often lead to excessively long prompts, substantial memory footprints, and high computational overhead. To overcome these limitations, we propose "Token Factory", a framework designed to transform traditional signals into "soft tokens" that can be directly processed by LRMs. This approach enables efficient integration and compression of heterogeneous input features, preventing prompt length explosion while enhancing model performance. We detail the architecture of Token Factory and present experimental results validating its effectiveness in a production-scale recommendation environment.

cs.IR↗

arXiv2026

LLM-Based User Personas for Recommendations at Scale

Large Language Models (LLMs) offer unprecedented potential for enhancing recommendation systems through their world knowledge and reasoning capabilities. However, existing approaches often rely on structured IDs or offline processing, limiting semantic richness, real-time adaptability, and user-facing interpretability. In this paper, we introduce a novel framework that enables real-time generation of LLM-based user interest personas for a large-scale commercial video recommendation platform. Our method generates natural-language user interest personas that address the exploitation-exploration trade-off by combining the summarization of existing interests with novel topics, directly during serving. To overcome the computational challenges of online LLM inference at a billion-user scale, we design a cost-efficient architecture leveraging knowledge distillation, asynchronous inference, and input optimization via semantically clustered video representations. Extensive offline evaluations, user studies, and live A/B tests demonstrate significant improvements in viewer value. This work bridges the gap between high-level semantic understanding and industrial-scale recommendation, paving the way for more dynamic, explainable, and satisfying personalized experiences.

cs.IR↗

arXiv2026

ORBIT: Preserving Foundational Language Capabilities in GenRetrieval via Origin-Regulated Merging

Despite the rapid advancements in large language model (LLM) development, fine-tuning them for specific tasks often results in the catastrophic forgetting of their general, language-based reasoning abilities. This work investigates and addresses this challenge in the context of the Generative Retrieval (GenRetrieval) task. During GenRetrieval fine-tuning, we find this forgetting occurs rapidly and correlates with the distance between the fine-tuned and original model parameters. Given these observations, we propose ORBIT, a novel approach that actively tracks the distance between fine-tuned and initial model weights, and uses a weight averaging strategy to constrain model drift during GenRetrieval fine-tuning when this inter-model distance exceeds a maximum threshold. Our results show that ORBIT retains substantial text and retrieval performance by outperforming both common continual learning baselines and related regularization methods that also employ weight averaging.

cs.CL↗

arXiv2026

Vectorizing the Trie: Efficient Constrained Decoding for LLM-based Generative Retrieval on Accelerators

Generative retrieval has emerged as a powerful paradigm for LLM-based recommendation. However, industrial recommender systems often benefit from restricting the output space to a constrained subset of items based on business logic (e.g. enforcing content freshness or product category), which standard autoregressive decoding cannot natively support. Moreover, existing constrained decoding methods that make use of prefix trees (Tries) incur severe latency penalties on hardware accelerators (TPUs/GPUs). In this work, we introduce STATIC (Sparse Transition Matrix-Accelerated Trie Index for Constrained Decoding), an efficient and scalable constrained decoding technique designed specifically for high-throughput LLM-based generative retrieval on TPUs/GPUs. By flattening the prefix tree into a static Compressed Sparse Row (CSR) matrix, we transform irregular tree traversals into fully vectorized sparse matrix operations, unlocking massive efficiency gains on hardware accelerators. We deploy STATIC on a large-scale industrial video recommendation platform serving billions of users. STATIC produces significant product metric impact with minimal latency overhead (0.033 ms per step and 0.25% of inference time), achieving a 948x speedup over a CPU trie implementation and a 47-1033x speedup over a hardware-accelerated binary-search baseline. Furthermore, the runtime overhead of STATIC remains extremely low across a wide range of practical configurations. To the best of our knowledge, STATIC enables the first production-scale deployment of strictly constrained generative retrieval. In addition, evaluation on academic benchmarks demonstrates that STATIC can considerably improve cold-start performance for generative retrieval. Our code is available at https://github.com/youtube/static-constraint-decoding.

cs.IR↗

arXiv2025

Selecting User Histories to Generate LLM Users for Cold-Start Item Recommendation

Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning, generalization, and simulating human-like behavior across a wide range of tasks. These strengths present new opportunities to enhance traditional recommendation systems (RS), especially in the cold-start item scenario where newly introduced items lack interactions. Existing works have used LLMs to address cold-start issues in traditional RS through data augmentation, but they have limitations. One recent work directly addresses this issue by prompting LLMs to generate augmented interaction data between randomly sampled users and cold-start items. Then, they train the traditional RS with augmented data, incorporating collaborative signals for cold-start items. Although they use LLMs to provide cold-start items with feedback, they use partial user histories, which does not allow the LLM to fully emulate the user. Furthermore, randomly selecting users is not optimal for augmentation. To address these challenges, we leverage the LLM as a user and develop a reinforcement learning (RL) framework that trains a policy to select users for augmentation, optimizing for cold-start item performance after augmented training. The policy model learns to select users for cold-start item data augmentation based on their behavioral features and histories. To optimize user selection for cold-start item performance, we employ a policy gradient method that updates the policy in the direction of actions that lead to high rewards. Experiments on Amazon Product Review datasets show substantial gains in cold-start item recall, demonstrating the effectiveness of our method as a scalable, serving-efficient augmentation strategy for modern RS.

cs.IR↗

arXiv2025

Balancing Fine-tuning and RAG: A Hybrid Strategy for Dynamic LLM Recommendation Updates

Large Language Models (LLMs) empower recommendation systems through their advanced reasoning and planning capabilities. However, the dynamic nature of user interests and content poses a significant challenge: While initial fine-tuning aligns LLMs with domain knowledge and user preferences, it fails to capture such real-time changes, necessitating robust update mechanisms. This paper investigates strategies for updating LLM-powered recommenders, focusing on the trade-offs between ongoing fine-tuning and Retrieval-Augmented Generation (RAG). Using an LLM-powered user interest exploration system as a case study, we perform a comparative analysis of these methods across dimensions like cost, agility, and knowledge incorporation. We propose a hybrid update strategy that leverages the long-term knowledge adaptation of periodic fine-tuning with the agility of low-cost RAG. We demonstrate through live A/B experiments on a billion-user platform that this hybrid approach yields statistically significant improvements in user satisfaction, offering a practical and cost-effective framework for maintaining high-quality LLM-powered recommender systems.

cs.IR↗

arXiv2025

PLUM: Adapting Pre-trained Language Models for Industrial-scale Generative Recommendations

Large Language Models (LLMs) pose a new paradigm of modeling and computation for information tasks. Recommendation systems are a critical application domain poised to benefit significantly from the sequence modeling capabilities and world knowledge inherent in these large models. In this paper, we introduce PLUM, a framework designed to adapt pre-trained LLMs for industry-scale recommendation tasks. PLUM consists of item tokenization using Semantic IDs, continued pre-training (CPT) on domain-specific data, and task-specific fine-tuning for recommendation objectives. For fine-tuning, we focus particularly on generative retrieval, where the model is directly trained to generate Semantic IDs of recommended items based on user context. We conduct comprehensive experiments on large-scale internal video recommendation datasets. Our results demonstrate that PLUM achieves substantial improvements for retrieval compared to a heavily-optimized production model built with large embedding tables. We also present a scaling study for the model's retrieval performance, our learnings about CPT, a few enhancements to Semantic IDs, along with an overview of the training and inference methods that enable launching this framework to billions of users in YouTube.

cs.IR↗

arXiv2025

ReMem: Mutual Information-Aware Fine-tuning of Pretrained Vision Transformers for Effective Knowledge Distillation

Knowledge distillation from pretrained visual representation models offers an effective approach to improve small, task-specific production models. However, the effectiveness of such knowledge transfer drops significantly when distilling from strong models that are pretrained in a large scale. In this paper, we address this challenge for pretrained Vision Transformers (ViTs) by exploring methods to fine-tune them for more effective knowledge transfer. Motivated by the connection between mutual information and distillation effectiveness, we propose to employ mutual information-aware optimization during finetuning. For small or highly-imbalanced downstream datasets where such optimization becomes less effective, we introduce a simple yet effective heuristic of reweighting MLP blocks. This approach is inspired by our observation that top MLP blocks are primarily responsible for mutual information loss. Our method enables small student models to benefit from those pretrained models among the strongest.

cs.LG↗

arXiv2025

Serendipitous Recommendation with Multimodal LLM

Conventional recommendation systems succeed in identifying relevant content but often fail to provide users with surprising or novel items. Multimodal Large Language Models (MLLMs) possess the world knowledge and multimodal understanding needed for serendipity, but their integration into billion-item-scale platforms presents significant challenges. In this paper, we propose a novel hierarchical framework where fine-tuned MLLMs provide high-level guidance to conventional recommendation models, steering them towards more serendipitous suggestions. This approach leverages MLLM strengths in understanding multimodal content and user interests while retaining the efficiency of traditional models for item-level recommendation. This mitigates the complexity of applying MLLMs directly to vast action spaces. We also demonstrate a chain-of-thought strategy enabling MLLMs to discover novel user interests by first understanding video content and then identifying relevant yet unexplored interest clusters. Through live experiments within a commercial short-form video platform serving billions of users, we show that our MLLM-powered approach significantly improves both recommendation serendipity and user satisfaction.

cs.IR↗

arXiv2025

User Feedback Alignment for LLM-powered Exploration in Large-scale Recommendation Systems

Exploration, the act of broadening user experiences beyond their established preferences, is challenging in large-scale recommendation systems due to feedback loops and limited signals on user exploration patterns. Large Language Models (LLMs) offer potential solutions by leveraging their world knowledge to recommend novel content outside these loops. A key challenge is aligning LLMs with user preferences while preserving their knowledge and reasoning. To enhance planning for new user interests using LLMs, this paper introduces a novel approach that combines hierarchical planning with LLM inference-time scaling. This method aims to improve recommendation relevancy without compromising novelty. We decouple novelty and user-alignment, training separate LLMs for each objective. We then scale up the novelty-focused LLM's inference and select the best-of-n predictions using the user-aligned LLM. Live experiments demonstrate efficacy, showing significant gains in both user satisfaction (measured by watch activity and active user counts) and exploration diversity.

cs.IR↗

arXiv2024

STAR: A Simple Training-free Approach for Recommendations using Large Language Models

Recent progress in large language models (LLMs) offers promising new approaches for recommendation system tasks. While the current state-of-the-art methods rely on fine-tuning LLMs to achieve optimal results, this process is costly and introduces significant engineering complexities. Conversely, methods that directly use LLMs without additional fine-tuning result in a large drop in recommendation quality, often due to the inability to capture collaborative information. In this paper, we propose a Simple Training-free Approach for Recommendation (STAR), a framework that utilizes LLMs and can be applied to various recommendation tasks without the need for fine-tuning, while maintaining high quality recommendation performance. Our approach involves a retrieval stage that uses semantic embeddings from LLMs combined with collaborative user information to retrieve candidate items. We then apply an LLM for pairwise ranking to enhance next-item prediction. Experimental results on the Amazon Review dataset show competitive performance for next item prediction, even with our retrieval stage alone. Our full method achieves Hits@10 performance of +23.8% on Beauty, +37.5% on Toys & Games, and -1.8% on Sports & Outdoors relative to the best supervised models. This framework offers an effective alternative to traditional supervised models, highlighting the potential of LLMs in recommendation systems without extensive training or custom architectures.

cs.IR↗

arXiv2024

Bridging the Gap: Unpacking the Hidden Challenges in Knowledge Distillation for Online Ranking Systems

Knowledge Distillation (KD) is a powerful approach for compressing a large model into a smaller, more efficient model, particularly beneficial for latency-sensitive applications like recommender systems. However, current KD research predominantly focuses on Computer Vision (CV) and NLP tasks, overlooking unique data characteristics and challenges inherent to recommender systems. This paper addresses these overlooked challenges, specifically: (1) mitigating data distribution shifts between teacher and student models, (2) efficiently identifying optimal teacher configurations within time and budgetary constraints, and (3) enabling computationally efficient and rapid sharing of teacher labels to support multiple students. We present a robust KD system developed and rigorously evaluated on multiple large-scale personalized video recommendation systems within Google. Our live experiment results demonstrate significant improvements in student model performance while ensuring consistent and reliable generation of high quality teacher labels from a continuous data stream of data.

cs.IR↗

arXiv2024

Leveraging LLM Reasoning Enhances Personalized Recommender Systems

Recent advancements have showcased the potential of Large Language Models (LLMs) in executing reasoning tasks, particularly facilitated by Chain-of-Thought (CoT) prompting. While tasks like arithmetic reasoning involve clear, definitive answers and logical chains of thought, the application of LLM reasoning in recommendation systems (RecSys) presents a distinct challenge. RecSys tasks revolve around subjectivity and personalized preferences, an under-explored domain in utilizing LLMs' reasoning capabilities. Our study explores several aspects to better understand reasoning for RecSys and demonstrate how task quality improves by utilizing LLM reasoning in both zero-shot and finetuning settings. Additionally, we propose RecSAVER (Recommender Systems Automatic Verification and Evaluation of Reasoning) to automatically assess the quality of LLM reasoning responses without the requirement of curated gold references or human raters. We show that our framework aligns with real human judgment on the coherence and faithfulness of reasoning responses. Overall, our work shows that incorporating reasoning into RecSys can improve personalized tasks, paving the way for further advancements in recommender system methodologies.

cs.IR↗

arXiv2024

Aligning Large Language Models with Recommendation Knowledge

Large language models (LLMs) have recently been used as backbones for recommender systems. However, their performance often lags behind conventional methods in standard tasks like retrieval. We attribute this to a mismatch between LLMs' knowledge and the knowledge crucial for effective recommendations. While LLMs excel at natural language reasoning, they cannot model complex user-item interactions inherent in recommendation tasks. We propose bridging the knowledge gap and equipping LLMs with recommendation-specific knowledge to address this. Operations such as Masked Item Modeling (MIM) and Bayesian Personalized Ranking (BPR) have found success in conventional recommender systems. Inspired by this, we simulate these operations through natural language to generate auxiliary-task data samples that encode item correlations and user preferences. Fine-tuning LLMs on such auxiliary-task data samples and incorporating more informative recommendation-task data samples facilitates the injection of recommendation-specific knowledge into LLMs. Extensive experiments across retrieval, ranking, and rating prediction tasks on LLMs such as FLAN-T5-Base and FLAN-T5-XL show the effectiveness of our technique in domains such as Amazon Toys & Games, Beauty, and Sports & Outdoors. Notably, our method outperforms conventional and LLM-based baselines, including the current SOTA, by significant margins in retrieval, showcasing its potential for enhancing recommendation quality.

cs.IR↗

arXiv2024

Wisdom of Committee: Diverse Distillation from Large Foundation Models and Domain Experts

Knowledge distillation from foundation models to compact domain models is challenging due to substantial gaps in capacity, architecture, and modality. For example, in our experiments, distilling from a 76M-parameter language model to a 2M-parameter recommender closes less than 40% of the performance gap between the undistilled student and the teacher. We show that introducing domain-specific experts -- which share the student's architectural characteristics -- alongside the foundation model as a diverse teacher committee significantly improves transfer. However, standard multi-teacher methods fail to exploit this diversity: naively combining heterogeneous teachers can degrade performance below single-teacher distillation. To address this, we propose DiverseDistill, an interactive distillation framework that employs a learnable Question-Answer mechanism to generate teacher-conditioned queries and align heterogeneous teacher outputs into the student's representation space. Unlike methods requiring gradient-based co-optimization or architectural modification of teachers, DiverseDistill operates with frozen teachers using only forward-pass inference through their intermediate layers: no parameter updates, no co-training, and no architectural surgery. A dynamic teacher importance mechanism further reduces training cost by filtering low-relevance teachers per sample (e.g., ~30% fewer forward passes with no quality loss for recommendation tasks), while the entire Distillation Module is discarded after training, adding zero inference overhead. Evaluations on recommendation (38x compression) and vision (3.6x compression) tasks demonstrate that DiverseDistill recovers 73-114% of the teacher-student performance gap, consistently outperforming all single- and multi-teacher baselines.

cs.LG↗