Editing Large language models (LLMs) with real-world, unstructured knowledge is essential for correcting and updating their internal parametric knowledge. In this work, we revisit the fundamental next-token prediction (NTP) as a candidate paradigm for unstructured editing. We identify Context Reliance as a critical failure mode of NTP-based approaches, where knowledge acquired from edited text becomes highly dependent on its preceding context, leading to recall failures when that context is absent during inference. This hypothesis is supported by our empirical validation that prepending context during inference recovers knowledge recall. We further theoretically demonstrate that Context Reliance is an inherent consequence of gradient-based optimization, which tends to bind acquired knowledge to a specific aggregated contextual representation. To address this, we propose a simple yet effective COntext-INdependent editing framework (COIN), encouraging model to focus on knowledge within local scope rather than memorizing contextual patterns. Evaluations show that COIN reduces Context Reliance by 45.2% and outperforms strong baselines by 23.6% in editing success rate, highlighting the vital role of mitigating Context Reliance for robust editing.

Generative retrieval models perform document retrieval by autoregressively generating document identifiers (DocIDs). This process naturally forms a sequential decision problem, where each decoding step selects a DocID token and the complete token sequence determines the retrieved document. However, retrieval effectiveness is typically evaluated only after the full DocID is generated, creating a mismatch between token-level generation and document-level relevance supervision. As a result, existing reinforcement learning methods for generative retrieval mostly rely on sequence-level rewards, where the same document-level feedback is propagated to all decoding steps. Such coarse-grained feedback makes it difficult to identify which token decisions are responsible for successful or failed retrieval. In this work, we propose a fine-grained reinforcement learning framework for generative retrieval with token-level relevance rewards. Instead of assigning a single reward to the entire generated DocID, we estimate step-wise rewards by measuring how each token decision changes the expected retrieval quality of the corresponding generation trajectory. This enables more precise credit assignment and encourages the policy to favor token decisions that contribute more directly to document-level relevance. We further develop practical reward estimation strategies tailored to the DocID generation process and incorporate them into a policy optimization framework. Experiments on retrieval benchmarks show that our method consistently outperforms sequence-level reward baselines, demonstrating the effectiveness of fine-grained supervision for aligning autoregressive DocID generation with retrieval objectives.

Knowledge Editing has emerged as a promising solution for efficiently updating embedded knowledge in large language models (LLMs). While existing approaches demonstrate effectiveness in integrating new knowledge and preserving the original capabilities of LLMs, they fail to maintain fine-grained irrelevant knowledge facts that share the same subject as edited knowledge but differ in relation and object. This challenge arises because subject representations inherently encode multiple attributes, causing the target and fine-grained irrelevant knowledge to become entangled in the representation space, and thus vulnerable to unintended alterations during editing. To address this, we propose DiKE, a novel approach that Disentangles Knowledge representations for LLM Editing (DiKE). DiKE consists of two key components: a Knowledge Representation Disentanglement (KRD) module that decomposes the subject representation into target-knowledgerelated and -unrelated components, and a Disentanglement-based Knowledge Edit (DKE) module that updates only the target-related component while explicitly preserving the unrelated one. We further derive a closed-form, rank-one parameter update based on matrix theory to enable efficient and minimally invasive edits. To rigorously evaluate fine-grained irrelevant knowledge preservation, we construct FINE-KED, a new benchmark comprising fine-grained irrelevant knowledge at different levels of relational similarity to the edited knowledge. Extensive experiments across multiple LLMs demonstrate that DiKE substantially improves fine-grained irrelevant knowledge preservation while maintaining competitive general editing performance.

Generative retrieval (GR) reformulates the Information Retrieval (IR) task as the generation of document identifiers (docIDs). Despite its promise, existing GR models exhibit poor generalization to newly added documents, often failing to generate the correct docIDs. While incremental training offers a straightforward remedy, it is computationally expensive, resource-intensive, and prone to catastrophic forgetting, thereby limiting the scalability and practicality of GR. In this paper, we identify the core bottleneck as the decoder's ability to map hidden states to the correct docIDs of newly added documents. Model editing, which enables targeted parameter modifications for docID mapping, represents a promising solution. However, applying model editing to current GR models is not trivial, which is severely hindered by indistinguishable edit vectors across queries, due to the high overlap of shared docIDs in retrieval results. To address this, we propose DOME (docID-oriented model editing), a novel method that effectively and efficiently adapts GR models to unseen documents. DOME comprises three stages: (1) identification of critical layers, (2) optimization of edit vectors, and (3) construction and application of updates. At its core, DOME employs a hybrid-label adaptive training strategy that learns discriminative edit vectors by combining soft labels, which preserve query-specific semantics for distinguishable updates, with hard labels that enforce precise mapping modifications. Experiments on widely used benchmarks, including NQ and MS MARCO, show that our method significantly improves retrieval performance on new documents while maintaining effectiveness on the original collection. Moreover, DOME achieves this with only about 60% of the training time required by incremental training, considerably reducing computational cost and enabling efficient, frequent model updates.

Conversational search needs an understanding of the interaction history between users and systems to retrieve information that aligns with the user’s current query. A prominent challenge in this task lies in addressing irrelevant context, which can obscure the user’s true intent and hinder retrieval accuracy. Existing methods for conversational search rely on the dense retrieval architecture, which encodes the context into fixed-length vectors. However, these approaches are constrained by the inherent limitations of the bi-encoder architecture, making it difficult to mitigate interference from irrelevant contexts effectively. In this paper, we propose Conversational Generative Retrieval for Context Denoising (CGR4CD), a novel framework that leverages a sequence-to-sequence generative retrieval architecture with cross-attention layers to dynamically capture key information during decoding. Our method introduces innovative proposition-based docids and a two-stage training strategy, combining generation and rank losses for enhanced denoising and ranking. Extensive evaluations on three conversational search benchmarks (TopiOCQA, QReCC, and OR-QuAC) demonstrate that CGR4CD achieves state-of-the-art performance, outperforming baselines in 8 out of 9 evaluation metrics. On TopiOCQA, which is characterized by topic switching and high noise, CGR4CD achieves an average improvement of 8.2% over baselines of similar size and surpasses the LLM-based Chatretriever by 4.4%, highlighting its superior context denoising capability. Our approach also maintains parameter efficiency, achieving these results with a T5-large backbone compared to larger LLM-based retrievers. Detailed analysis further validates its superior capability in contextual denoising, particularly in noisy scenarios.

Large Language Models have shown great success in recommender systems. However, the limited and sparse nature of user data often restricts the LLM's ability to effectively model behavior patterns. To address this, existing studies have explored cross-domain solutions by conducting Cross-Domain Recommendation tasks. But previous methods typically assume domains are overlapped and can be accessed readily. None of the LLM methods address the privacy-preserving issues in the CDR settings, that is, Privacy-Preserving Cross-Domain Recommendation. Conducting non-overlapping PPCDR with LLM is challenging since: 1)The inability to share user identity or behavioral data across domains impedes effective cross-domain alignment. 2)The heterogeneity of data modalities across domains complicates knowledge integration. 3)Fusing collaborative filtering signals from traditional recommendation models with LLMs is difficult, as they operate within distinct feature spaces. To address the above issues, we propose SF-UBM, a Semantic-enhanced Federated User Behavior Modeling method. Specifically, to deal with Challenge 1, we leverage natural language as a universal bridge to connect disjoint domains via a semantic-enhanced federated architecture. Here, text-based item representations are encrypted and shared, while user-specific data remains local. To handle Challenge 2, we design a Fact-counter Knowledge Distillation module to integrate domain-agnostic knowledge with domain-specific knowledge, across different data modalities. To tackle Challenge 3, we project pre-learned user preferences and cross-domain item representations into the soft prompt space, aligning behavioral and semantic spaces for effective LLM learning. We conduct extensive experiments on three pairs of real-world domains, and the experimental results demonstrate the effectiveness of SF-UBM compared to the recent SOTA methods.

Instruction fine-tuning is employed to enhance the instruction-following ability of large language models (LLMs). As the amount of instruction fine-tuning data increases, selecting the optimal core set becomes particularly important. However, ensuring the diversity of the core set remains a significant challenge. Existing methods predominantly distinguish different training data based on the text features themselves, decoupled from LLMs' own understanding and representation of the data. To address this issue, we propose a Model-Aware Diverse Core Set Selection method, which distinguishes data features based on the neural activation states during LLM inference. This approach serves as an efficient instantiation of coverage-based selection using model-intrinsic activation features to ensure the diversity in the core set. We extensively evaluate our method on six benchmarks that cover five distinct tasks. In our method, the core set selected by the 3B-parameter LLM performs effectively when utilized to fine-tune larger models with 7B, 8B, and 13B parameters. Experimental results on the Alpaca-GPT4 dataset, which comprises 52K instruction-response pairs, show that the core set, sized at 15% of the original dataset and selected by Llama-3.2-3B-Instruct, achieves an average improvement of 2.5% when fine-tuning four larger base models compared with training on the full dataset. The experimental results demonstrate that our method enhances model performance on multiple downstream tasks while reducing data requirements.

The rapid scaling of large language models (LLMs) has made full fine-tuning increasingly impractical due to its prohibitive computational, storage, and memory costs. Although parameter-efficient fine-tuning (PEFT) methods such as Low-Rank Adaptation (LoRA) significantly reduce the number of trainable parameters, most existing approaches treat layers independently and overlook potential cross-layer redundancy. To address this limitation, this paper aims to develop a more parameter-efficient adaptation strategy that can effectively exploit cross-layer parameter sharing while maintaining model expressiveness. Motivated by our preliminary analysis showing that the optimal sharing granularity varies across tasks, we propose ASLoRA, a cross-layer parameter-sharing framework that combines global sharing with partial adaptive sharing. Specifically, ASLoRA introduces a globally shared low-rank input projection (A) together with selectively merged layer-wise output projections (B), enabling the model to balance global parameter reuse with task-specific adaptation. Extensive experiments on diverse NLP benchmarks demonstrate that ASLoRA consistently outperforms standard LoRA while using only about 20% of the trainable parameters. Further analyses show that our proposed adaptive sharing strategy enhances both adaptation capacity and task-specific adaptability.

Sequential knowledge editing in large language models often causes catastrophic collapse of the model’s general abilities, especially for parameter-modifying methods. Existing approaches mitigate this issue through heuristic constraints on parameter updates, the mechanisms underlying such degradation remain insufficiently understood. In this work, we present a systematic spectral analysis of sequential knowledge editing and show that a model’s general abilities are closely associated with dominant singular directions of pretrained weight matrices. These directions are highly sensitive to perturbations and are progressively disrupted by repeated edits, closely tracking the collapse in both editing efficacy and general performance. Building on this insight, we propose REVIVE, a plug-and-play framework that prevents model collapse by explicitly preserving this dominant subspace. REVIVE analyzes parameter updates in the spectral basis of the original weights and filters out components that would interfere with the dominant subspace. Extensive experiments across multiple models and benchmarks show that REVIVE consistently improves editing efficacy while substantially preserving general abilities under long-horizon sequential editing, including extreme settings with up to 20,000 edits.

Multimodal large language models (MLLMs) have shown strong general capabilities. However, their capabilities, reliability boundaries, and key bottlenecks in automated decision-making for underground engineering remain unclear because domain-specific multimodal datasets are lacking. To advance the development of MLLMs for intelligent construction, this paper introduces UndergrBench, a comprehensive Chinese benchmark. UndergrBench comprehensively evaluates MLLMs across three cognitive levels—knowledge memorization, understanding, and application, using 10,888 samples across 9 tasks. The data sources include engineering codes, geological reports, scientific papers, and disaster case records, with expert annotation ensuring quality. Based on UndergrBench, 59 general-purpose models (20 multimodal and 39 language-only) were evaluated. Results show Doubao-Seed-1.6 and Qwen3-235B-A22B achieve best performance among MLLMs and LLMs. Nevertheless, a considerable gap remains before reliable application in underground engineering tasks. UndergrBench provides a framework for assessing and improving MLLMs in geological investigation, design, and construction, laying a foundation for future domain-specialized large models.

While generative retrieval (GR) demonstrates competitive performance on standard retrieval benchmarks, existing approaches directly map queries to document identifiers (docids) without intermediate deliberation, limiting their effectiveness for complex queries that require multi-step reasoning. As a preliminary study on integrating chain-of-thought (CoT) into generative retrieval, we introduce ThinkGR, a unified framework that interleaves CoT with docid generation, enabling iterative thinking and retrieval within a single generative process. To bridge the gap between free-form thought generation and structured retrieval targets, we design (1) a hybrid decoding strategy that dynamically switches between unconstrained thought generation and constrained docid decoding, and (2) a two-phase training approach that first aligns thought-retrieval patterns through supervised fine-tuning, then optimizes thought quality via retrieval-grounded reinforcement learning. Experiments on four multi-hop retrieval benchmarks demonstrate that ThinkGR achieves state-of-the-art performance with an average improvement of +6.86%. Our work opens new avenues for enhancing generative retrieval with explicit deliberation capabilities, with promising implications for retrieval tasks requiring complex reasoning.

Recommender systems alleviate information overload, yet repeated feedback between recommendations and user interactions can reinforce existing preferences and narrow users' exposure, forming information cocoons. While this phenomenon has been widely studied in traditional sequential recommendation, its impact on generative recommendation remains unclear. By replacing atomic item IDs with Semantic ID (SID) sequences, generative recommenders introduce a different recommendation mechanism whose role in information cocoon formation is not yet understood. To investigate whether generative recommenders deepen information cocoons, we propose \textsc{RecLoop}, a closed-loop simulation framework with LLM-driven user agents. We compare two generative recommenders and two traditional sequential baselines on two Amazon datasets across multiple feedback cycles. In addition to standard exposure-level metrics, we introduce \emph{Code-Space Structural Cocoon}, a model-level metric that measures concentration in the generated SID space. Experimental results show that generative recommenders are generally less prone to exposure-level cocoon formation than traditional baselines, preserving broader exposure diversity and slowing cross-user homogenization. However, feedback loops can still induce concentration within the generated SID space. We further find that cocoon severity depends strongly on tokenization strategy and model scale: collaborative-signal tokenization produces stronger cocoon effects than semantic tokenization, whereas larger models maintain greater code-space diversity and better retain access to niche content. These findings suggest that information cocoons in generative recommendation are shaped not only by recommendation behavior, but also by item tokenization and model capacity. Our code is available at https://github.com/Dregen-Yor/RecLoop.

To address the inefficiency of full fine-tuning of Contrastive Language-Image Pre-training (CLIP) models and the performance loss of adapter-based methods, we propose a novel efficient hybrid fine-tuning strategy (called HFLIP) to achieve a balance of efficiency and performance. HFLIP fine-tunes the key selected ViT blocks with interpretable semantic attention supervision on selected transformer heads via machine-learning methods’ selection, while keeping other blocks adapter-based for efficiency. Specifically, HFLIP introduces two key components: (1) a Dynamic Block-selection Genetic Algorithm (DBGA) that automatically selects a small subset of critical blocks in the ViT for full tuning, while keeping the rest adapter-tuned, ensuring a proper trade-off between fine-tuning effectiveness and efficiency; and (2) a Clustering-based Head-selection with Explainable-attention Guidance (CHEG), where hierarchical clustering is employed to identify representative attention heads, which are then fine-tuned under guidance from explainable attention maps, encouraging semantically consistent and globally diverse attention patterns. Extensive experiments on multiple downstream tasks show that HFLIP achieves comparable or even better performance than full fine-tuning by updating only 30% of the training parameters, while reducing GPU memory consumption by about 16%. In addition, HFLIP makes the CLIP-based ViT attention mechanism more interpretable compared to both the pretrained CLIP and other fine-tuned variants. We release our code at https://github.com/huiye8870/HFLIP.

Retrieval-augmented generation (RAG) integrates large language models ( LLM s) with retrievers to access external knowledge, improving the factuality of LLM generation in knowledge-grounded tasks. To optimize the RAG performance, most previous work independently fine-tunes the retriever to adapt to frozen LLM s or trains the LLMs to use documents retrieved by off-the-shelf retrievers, lacking end-to-end training supervision. Recent work addresses this limitation by jointly training these two components but relies on overly simplifying assumptions of document independence, which has been criticized for being far from real-world scenarios. Thus, effectively optimizing the overall RAG performance remains a critical challenge. We propose a direct retrieval-augmented optimization framework, named DRO, that enables end-to-end training of two key components: (i) a generative knowledge selection model and (ii) an LLM generator. DRO alternates between two phases: (i) document permutation estimation and (ii) re-weighted maximization, progressively improving RAG components through a variational approach. In the estimation step, we treat document permutation as a latent variable and directly estimate its distribution from the selection model by applying an importance sampling strategy. In the maximization step, we calibrate the optimization expectation using importance weights and jointly train the selection model and LLM generator. Our theoretical analysis reveals that DRO is analogous to policy-gradient methods in reinforcement learning. Extensive experiments conducted on five datasets illustrate that DRO outperforms the best baseline with 5 experiments to qualitatively analyze the stability, convergence, and variance of DRO.

An increasing number of studies have been proposed for community healthcare services (CHS) that connect users and physicians. Complex descriptions of diseases and symptoms make CHS difficult to match users’ amateur questions with physicians. In this study, we propose the task of question-physician routing to address this issue by matching questions and physicians. Vocabulary gaps, nonreciprocal commutations, and limited user profiling information in CHS make question-physician routing a challenging research problem. To address the above challenges, we propose a new question-routing approach to match a physician with a given question, where we argue that the assistance of external medical knowledge helps to improve the performance of question-physician routing. Specifically, we propose a medical question-routing framework, namely deep medical question routing (DMQR), which integrates external medical knowledge into the routing process. DMQR can be divided into three main components: First, we construct a medical knowledge extractor to extract knowledge from a medical knowledge base constructed based on external knowledge resources. Then, we propose a knowledge interpreter to integrate medical knowledge into the given question for a better representation. Lastly, we route top-k relevant candidate physicians to the question by employing a hierarchical multi-label classifier. To further enhance the robustness and efficiency of question-physician routing, we propose a virtual adversarial medical question routing optimization method, v-DMQR, and an efficient optimization way that uses anchor question-physician pairs, namely av-DMQR. Since there is no existing data in the literature, we collect a large-scale dataset from real-world CHS scenarios. Extensive experiments performed on our dataset verify the effectiveness of our proposed approach. Moreover, we find that our proposed approach significantly outperforms state-of-the-art baselines on medical question routing in terms of Recall, MAP, and MRR metrics.

Synthetic data has been widely used in training large language models (LLMs), demonstrating its effectiveness in enhancing model performance. However, most existing synthetic data generation methods prioritize quality and diversity, frequently neglecting alignment with the model's underlying data distribution. To mitigate this limitation, we introduce the Prefix-Guided Preference Data Synthesis method. Specifically, we employ prefix content of varying lengths from existing training data as prompts, enabling the LLMs itself to generate new training samples that better align with its own generative distribution. By mitigating distributional discrepancies, our approach reduces KL divergence to 54.8% of its original value. On the Open LLM Leaderboard v2, our method achieves an improvement of up to 71.3% on the MUSR task and an average performance gain of 6% across multiple benchmarks. These results indicate that utilizing prefix content to guide data synthesis effectively mitigates distributional misalignment with target models, thereby improving training outcomes.

The effectiveness of contrastive learning in sequential recommendation hinges on the construction of contrastive views, which ideally should be both semantically consistent and diverse. However, most existing CL-based methods rely on heuristic augmentations that are prone to removing crucial items or disrupting transition patterns, leading to semantic drift. While a few studies have explored learnable augmentations to improve view quality, they often suffer from limited diversity and still necessitate heuristic aids. Furthermore, the quality differences across views are rarely modeled explicitly and adaptively, aggravating the false-positive issue. To address these issues, we propose Quality-aware Collaborative Multi-Positive Contrastive Learning for sequential recommendation. First, we introduce a learnable collaborative sequence augmentation module that generates two augmented views under two complementary collaborative contexts, one based on same-target sequences and the other on similar sequences, thereby enhancing view diversity while preserving intent consistency.Second, we design a quality-aware mechanism, tightly integrated into the model representations, which estimates each view' s quality from the confidence of its augmentation operations and assigns adaptive weights to ensure that high-confidence views contribute more supervision while low-confidence ones contribute less.Extensive experiments on three real-world datasets demonstrate that QCMP-CL outperforms state-of-the-art CL-based sequential recommendation baselines.

As large language models (LLMs) are rapidly advancing and achieving near-human capabilities on specific tasks, aligning them with human values is becoming more urgent. In scenarios where LLMs outperform humans, we face a weak-to-strong alignment problem where we need to effectively align strong student LLMs through weak supervision generated by weak teachers. Existing alignment methods mainly focus on strong-to-weak alignment and self-alignment settings, and it is impractical to adapt them to the much harder weak-to-strong alignment setting. To fill this gap, we propose a multi-agent contrastive preference optimization (MACPO) framework. MACPO facilitates weak teachers and strong students to learn from each other by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. To get this, we devise a mutual positive behavior augmentation strategy to encourage weak teachers and strong students to learn from each other's positive behavior and further provide higher quality positive behavior for the next iteration. Additionally, we propose a hard negative behavior construction strategy to induce weak teachers and strong students to generate familiar negative behavior by fine-tuning on negative behavioral data. Experimental results on the HH-RLHF and PKU-SafeRLHF datasets, evaluated using both automatic metrics and human judgments, demonstrate that MACPO simultaneously improves the alignment performance of strong students and weak teachers. Moreover, as the number of weak teachers increases, MACPO achieves better weak-to-strong alignment performance through more iteration optimization rounds.

Offline reinforcement learning (RL) aims to learn policies without online explorations. To enlarge the training data, model-based offline RL learns a dynamics model which is utilized as a virtual environment to generate simulation data and enhance policy learning. However, existing data augmentation methods for offline RL suffer from (i) trivial improvement from short-horizon simulation; and (ii) the lack of evaluation and correction for generated data, leading to low-qualified augmentation. In this paper, we propose offline trajectory optimization for offline reinforcement learning (OTTO). The key motivation is to conduct long-horizon simulation and then utilize model uncertainty to evaluate and correct the augmented data. Specifically, we propose an ensemble of Transformers, a.k.a. World Transformers, to predict environment state dynamics and the reward function. Three strategies are proposed to use World Transformers to generate long-horizon trajectory simulation by perturbing the actions in the offline data. Then, an uncertainty-based World Evaluator is introduced to firstly evaluate the confidence of the generated trajectories and then perform the correction for low-confidence data. Finally, we jointly use the original data with the corrected augmentation data to train an offline RL algorithm. OTTO serves as a plug-in module and can be integrated with existing model-free offline RL methods. Experiments on various benchmarks show that OTTO can effectively improve the performance of representative offline RL algorithms, including in complex environments with sparse rewards like AntMaze. Codes are available at https://github.com/ZiqiZhao1/OTTO.