Large Language Model (LLM) personalization aims to adapt generation to individual users by capturing implicit preferences from historical interactions. Steering-based methods offer an inference-efficient solution, but existing work follows a difference-based paradigm that estimates personalized steering directions from representation offsets between contrastive historical examples, thus entangling inherent user preference with preference-agnostic contextual bias. This limitation arises because offset-based estimation uses raw contrastive variation, making the derived direction confounded by contextual bias rather than identifying preference-discriminative variation between user-conditioned and generic behavior. To address this, we propose Disentangled Personal Direction Discovery (DiPer), which formulates personalized steering vector estimation as a discriminative direction discovery problem. DiPer learns layer-wise user-specific directions through supervised metric learning, based on Recursive Feature Machine (CITATION), over historical activations, to discover preference-discriminative directions while suppressing contextual bias. Theoretical analysis and extensive experiments demonstrate DiPer’s efficacy and robustness.

Large language models (LLMs) have greatly advanced conversational systems, making Personalized Dialogue Generation (PDG) a core direction for tailoring responses to individual users. However, retrieval-augmented and user-parameterization approaches typically rely on dense user-level persona information and consistent domain-level persona distributions, leading to two key challenges for scalable PDG: sparse personas of unseen users hinder personalization, while domain-bound user parameters require costly retraining across dialogue scenes. Therefore, we argue that these limitations originate from an over-reliance on static data fitting, as opposed to mastering atomic personalization capabilities. To this end, we propose Meta-Collaborative Learning (MCL), a novel learning paradigm that proactively supplements cues via Neighbor Identifier and Booster, and dynamically aligns responses to preferences by Generator. By decomposing PDG into synergistic modules with bi-level joint optimization, MCL internalizes universal personalization skills for acquiring and leveraging contextual signals across diverse users and domains. Extensive experiments validate that MCL delivers robust personalization under practical unseen-user and cross-domain settings.

Recent studies indicate a fundamental incompatibility between ID representations and language model (LM) representations, as they capture behavioral and semantic spaces respectively. This mismatch leads LM representations to consistently underperform ID representations in recommendation tasks. In this work, we revisit this problem and show, from an information-theoretic perspective, that LLM representations retain all discriminative information in ID representations. Based on this, we introduce a Profile-then-Embedding (PtE) framework for recommendation, consisting of a Profile Stage, in which semantic user and item profiles are generated jointly through LLM-based bidirectional reasoning over user-item interactions, and a Personalized Embedding Stage, which encodes these profiles into task-aligned recommendation embeddings. We demonstrate PtE’s effectiveness across three benchmark datasets, including cold-start and long-tail scenarios, achieving substantial gains in both discriminative and generative recommendation models.

The improvement of LLMs’ instruction-following capabilities depends critically on the availability of high-quality instruction-response pairs. While existing automatic data synthetic methods alleviate the burden of manual curation, they often rely heavily on either the quality of seed data or strong assumptions about the structure and content of web documents. To tackle these challenges, we propose Web Reconstruction (WebR), a fully automated framework for synthesizing high-quality instruction-tuning (IT) data directly from raw web documents with minimal assumptions. Leveraging the inherent diversity of raw web content, we conceptualize web reconstruction as an instruction-tuning data synthesis task via a novel dual-perspective paradigm—Web as Instruction and Web as Response—where each web document is designated as either the input or output role to trigger the reconstruction process. Comprehensive experiments show that datasets generated by WebR outperform state-of-the-art baselines by up to 16.65% across four instruction-following benchmarks. Notably, WebR demonstrates superior compatibility, data efficiency, and scalability, enabling enhanced domain adaptation with minimal effort.

Recent advances in LLM-based recommendation have shown promise, yet their cross-domain generalization is hindered by a fundamental mismatch between language-centric pretraining and the recommendation task. Existing methods, relying on language-level knowledge, fail to capture dynamic, item-level user interests across domains. To bridge this gap, we propose RecBase, a domain-agnostic foundational model pretrained with a recommendation-oriented objective. RecBase leverages a large-scale, heterogeneous, cross-domain corpus with unified textual representations and feature mappings to enhance cross-domain generalization. To further align item semantics across domains, we introduce a unified item tokenizer that encodes items into hierarchical concept identifiers, enabling structured representation and efficient vocabulary sharing. The model is trained using an autoregressive objective to capture complex item-level sequential patterns. On eight real-world datasets, our 1.5B-parameter model matches or surpasses the performance of LLM baselines up to 7B parameters in zero-shot and cross-domain recommendation tasks.