Due to the disparity in the amount of semantic information conveyed by images and texts, image-text retrieval (ITR) faces significant difficulties related to semantic capacity imbalance. To tackle semantic capacity imbalance, there are three key challenges in ITR tasks: 1) How to construct a self-adjusting and discriminative semantic measurement unit for semantic capacity balance assessment. 2) How to design a reliable semantic capacity balancing system to accurately measure the degree of semantic capacity balance. 3) How to adaptively supplement missing semantic elements guided by the degree of semantic capacity balance. Therefore, we propose the Adaptive Semantic Supplementation Network (ASSN), which comprises four modules: 1) Construction, 2) Measurement, 3) Supplementation, and 4) Alignment. For the Construction module, dynamic semantic prototypes are constructed as the semantic measurement units for balance quantification. For the Measurement module, visual and textual semantic prototypes are placed on opposite sides of a lever. The distance of each semantic prototype from the pivot is considered as the lever arm, while its cross-modal transformed semantic capacity is regarded as the force. The semantic capacity balance is measured by calculating the torque difference between all semantic prototypes from the two modalities. For the Supplementation module, ASSN iteratively supplements semantic prototypes on the side with smaller total torque. Moreover, ASSN supplements missing semantic elements by retrieving relevant cross-modal neighbors while enforcing a maximum capacity constraint to prevent over-supplementation. This iterative process continues until balance is achieved. For the Alignment module, token features and adaptively supplemented semantic features are aggregated into semantic units to further improve alignment accuracy. Extensive experiments demonstrate the effectiveness of ASSN over state-of-the-art methods.

The emergence of reasoning multimodal large language models (MLLMs), which generate explicit chain-of-thought (CoT) reasoning before producing answers, has introduced a new challenge for knowledge editing: methods that appear successful under traditional metrics (teacher-forcing accuracy up to 100 We propose CRANE, a retrieval-augmented framework that requires no per-edit parameter modification. CRANE combines a modality-aware dual-library retrieval system with a two-phase training strategy: Supervised Fine-Tuning (SFT) for structural initialization, followed by GRPO with a Cognitive Routing Reward that trains the model to arbitrate between visual priors and injected edit facts. On ReasonEdit-Bench, CRANE achieves 96.9

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.

Multimodal knowledge editing (MKE) aims to correct the internal knowledge of large vision-language models after deployment, yet the behavioral patterns of post-edit models remain underexplored. In this paper, we identify a systemic failure mode in edited models, termed Entity Identity Confusion (EIC): edited models exhibit an absurd behavior where text-only queries about the original entity's identity unexpectedly return information about the new entity. To rigorously investigate EIC, we construct EC-Bench, a diagnostic benchmark that directly probes how image-entity bindings shift before and after editing. Our analysis reveals that EIC stems from existing methods failing to distinguish between Image-Entity (I-E) binding and Entity-Entity (E-E) relational knowledge in the model, causing models to overfit E-E associations as a shortcut: the image is still perceived as the original entity, with the new entity's name serving only as a spurious identity label. We further explore potential mitigation strategies, showing that constraining edits to the model's I-E processing stage encourages edits to act more faithfully on I-E binding, thereby substantially reducing EIC. Based on these findings, we discuss principled desiderata for faithful MKE and provide methodological guidance for future research.

Despite the strong reasoning capabilities of recent large language models (LLMs), achieving reliable performance on challenging tasks often requires post-training or computationally expensive sampling strategies, limiting their practical efficiency. In this work, we first show that a small subset of neurons in LLMs exhibits strong predictive correlations with reasoning correctness. Based on this observation, we propose AdaRAS (Adaptive Reasoning Activation Steering), a lightweight test-time framework that improves reasoning reliability by selectively intervening on neuron activations. AdaRAS identifies Reasoning-Critical Neurons (RCNs) via a polarity-aware mean-difference criterion and adaptively steers their activations during inference, enhancing incorrect reasoning traces while avoiding degradation on already-correct cases. Experiments on 10 mathematics and coding benchmarks demonstrate consistent improvements, including over 13

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.

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 Model (LLM) unlearning aims to erase or suppress undesirable knowledge within the model, offering promise for controlling harmful or private information to prevent misuse. However, recent studies highlight its limited efficacy in real-world scenarios, hindering practical adoption. In this study, we identify a pervasive issue underlying many downstream failures: the effectiveness of existing unlearning methods heavily depends on the form of training samples and frequently fails to generalize to alternate expressions of the same knowledge. We formally characterize this problem as Form-Dependent Bias and systematically investigate its specific manifestation patterns across various downstream tasks. To quantify its prevalence and support future research, we introduce ORT, a novel benchmark designed to evaluate the robustness of unlearning methods against variations in knowledge expression. Results reveal that Form-Dependent Bias is both widespread and severe among current techniques. We argue that LLM unlearning should be form-independent to address the endless forms of downstream tasks encountered in real-world security-critical scenarios. Towards this goal, we introduce Rank-one Concept Redirection (ROCR), a novel training-free method, as a promising solution path. ROCR performs unlearning by targeting the invariants in downstream tasks, specifically the activated dangerous concepts. It is capable of modifying model parameters within seconds to redirect the model's perception of a specific unlearning target concept to another harmless concept. Extensive experiments demonstrate that ROCR significantly improves unlearning effectiveness compared to traditional methods while generating highly natural outputs.

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.

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.

In embodied AI, safety risk can be latent: a benign instruction and a safe scene become hazardous only when composed. Prior work has advanced embodied safety by varying visual contexts or evaluating execution-time dynamics, but the complementary axis of fixing the scene and varying only the instruction remains underexplored. We introduce GuardianBench, an instruction-contrastive benchmark grounded in international safety standards that isolates this latent contextual risk through 3,024 instruction-scene examples organized as same-scene Safe/Unsafe contrastive pairs across various hazard categories. Benchmarking state-of-the-art vision-language models (VLMs) reveals instruction-insensitive verdicts: models disproportionately approve both instructions under a given scene; across the primary models, average pair accuracy is only 24.1

Counterfactual tuning (CFT) has emerged as a promising paradigm for Large Language Model (LLM) unlearning by training models to generate alternative fictitious knowledge in place of undesired content. However, in this work, we find that this paradigm still underperforms other paradigms in some aspects, and identify two previously overlooked pitfalls underlying this gap: (1) knowledge conflict, where mutual inconsistencies within counterfactual corpora induce conflicting gradients that disrupt parameter optimization, and (2) hallucination spillover, where fitting false targets instills a persistent fabrication bias, inflating hallucination rates on unrelated domains. To systematically diagnose these issues, we introduce RWKU+, an extended benchmark equipped with novel trade-off metrics and gradient-level diagnostic tools. Our work further discusses the limitations and overhead of the paradigm, aiming to provide insights and actionable guidance for more rigorous LLM unlearning research.

Large language models (LLMs) face challenges with internal knowledge inaccuracies and outdated information. Knowledge editing has emerged as a pivotal approach to mitigate these issues. Although current knowledge editing techniques exhibit promising performance in single-hop reasoning tasks, they show limitations when applied to multi-hop reasoning. Drawing on cognitive neuroscience and the operational mechanisms of LLMs, we hypothesize that the residual single-hop knowledge after editing causes edited models to revert to their original answers when processing multihop questions, thereby undermining their performance in multi-hop reasoning tasks. To validate this hypothesis, we conduct a series of experiments that empirically confirm our assumptions. Building on the validated hypothesis, we propose a novel knowledge editing method that incorporates a Knowledge Erasure mechanism for Large language model Editing (KELE). Specifically, we design an erasure function for residual knowledge and an injection function for new knowledge. Through joint optimization, we derive the optimal recall vector, which is subsequently utilized within a rank-one editing framework to update the parameters of targeted model layers. Extensive experiments on GPT-J (6B) and LLaMA-2 (7B) demonstrate that KELE substantially enhances the multi-hop reasoning capability of edited LLMs.

Out-of-context news is a common type of misinformation on online media platforms. This involves posting caption, alongside a mismatched news image. Reflecting its importance, researchers have developed models to detect such misinformation. However, a common limitation of these models is that they only consider the scenario where pre-labelled data is available for each news topic or agency, failing to address the out of-context news detection on unverified news of other topics or agencies. In this work, we therefore focus on domain adaptive out-of-context news detection. We regard news topic or news agency as the domain. order to effectively adapt the detection model to unlabelled news topics or agencies, we propose Contrastive Domain Adaptation with Test-Time Training (ConDA-TTT). It first applies contrastive learning to learn a more separable representation space for news inputs, and then uses maximum mean discrepancy (MMD) to remove the domain-specific features so as to keep the domain-invariant features. During test time, it uses the trained model to predict pseudo labels for the target domain test data, and selects those with higher confidence scores to train the classifier of the model, in order to further adapt the model to the target domain data distribution. This approach adapts the model at both training and test phase, making the domain adaptation more robust distribution shifts. Experimental results demonstrate that our approach outperforms state-of-the-art baselines in all the domain adaptation settings on two benchmark datasets, by as much as 2.6% in F1 and 2.4% accuracy.

Generative retrieval (GR) re-frames document retrieval as a sequence-based document identifier (DocID) generation task, memorizing documents with model parameters and enabling end-to-end retrieval without explicit indexing. Existing GR methods are based on auto-regressive generative models, i.e., the token generation is performed from left to right. However, such auto-regressive methods suffer from: (1) mismatch between DocID generation and natural language generation, e.g., an incorrect DocID token generated in early left steps would lead to totally erroneous retrieval; and (2) failure to balance the trade-off between retrieval efficiency and accuracy dynamically, which is crucial for practical applications. To address these limitations, we propose generative document retrieval with diffusion language models, dubbed DiffuGR. It models DocID generation as a discrete diffusion process: during training, DocIDs are corrupted through a stochastic masking process, and a diffusion language model is learned to recover them under a retrieval-aware objective. For inference, DiffuGR attempts to generate DocID tokens in parallel and refines them through a controllable number of denoising steps. In contrast to conventional left-to-right auto-regressive decoding, DiffuGR provides a novel mechanism to first generate more confident DocID tokens and refine the generation through diffusion-based denoising. Moreover, DiffuGR also offers explicit runtime control over the qualitylatency tradeoff. Extensive experiments on benchmark retrieval datasets show that DiffuGR is competitive with strong auto-regressive generative retrievers, while offering flexible speed and accuracy tradeoffs through variable denoising budgets. Overall, our results indicate that non-autoregressive diffusion models are a practical and effective alternative for generative document retrieval.

Knowledge is fundamental to the overall capabilities of Large Language Models (LLMs). The knowledge paradigm of a model, which dictates how it encodes and utilizes knowledge, significantly affects its performance. Despite the continuous development of LLMs under existing knowledge paradigms, issues within these frameworks continue to constrain model potential. This blog post highlight three critical open problems limiting model capabilities: (1) challenges in knowledge updating for LLMs, (2) the failure of reverse knowledge generalization (the reversal curse), and (3) conflicts in internal knowledge. We review recent progress made in addressing these issues and discuss potential general solutions. Based on observations in these areas, we propose a hypothetical paradigm based on Contextual Knowledge Scaling, and further outline implementation pathways that remain feasible within contemporary techniques. Evidence suggests this approach holds potential to address current shortcomings, serving as our vision for future model paradigms. This blog post aims to provide researchers with a brief overview of progress in LLM knowledge systems, while provide inspiration for the development of next-generation model architectures.

Knowledge editing has been proposed as an effective method for updating and correcting the internal knowledge of Large Language Models (LLMs). However, existing editing methods often struggle with complex tasks, such as multi-hop reasoning. In this paper, we identify and investigate the phenomenon of Editing Overfit, where edited models assign disproportionately high probabilities to the edit target, hindering the generalization of new knowledge in complex scenarios. We attribute this issue to the current editing paradigm, which places excessive emphasis on the direct correspondence between the input prompt and the edit target for each edit sample. To further explore this issue, we introduce a new benchmark, EVOKE (EValuation of Editing Overfit in Knowledge Editing), along with fine-grained evaluation metrics. Through comprehensive experiments and analysis, we demonstrate that Editing Overfit is prevalent in current editing methods and that common overfitting mitigation strategies are ineffective in knowledge editing. To overcome this, inspired by LLMs’ knowledge recall mechanisms, we propose a new plug-and-play strategy called Learn the Inference (LTI), which introduce a Multi-stage Inference Constraint module to guide the edited models in recalling new knowledge similarly to how unedited LLMs leverage knowledge through in-context learning. Extensive experimental results across a wide range of tasks validate the effectiveness of LTI in mitigating Editing Overfit.

Large Language Models (LLMs) inevitably acquire harmful information during training on massive datasets. LLM unlearning aims to eliminate the influence of such harmful information while maintaining the model's overall performance. Existing unlearning methods, represented by gradient ascent-based approaches, primarily focus on forgetting target data while overlooking the crucial impact of logically related knowledge on the effectiveness of unlearning. In this paper, through both theoretical and experimental analyses, we first demonstrate that a key reason for the suboptimal unlearning performance is that models can reconstruct the target content through reasoning with logically related knowledge. To address this issue, we propose Unlearning Improvement via Parameter Extrapolation (UIPE), a method that removes knowledge highly correlated with the forgetting targets. Experimental results show that UIPE significantly enhances the performance of various mainstream LLM unlearning methods on the TOFU benchmark.

Exclusion is an important and universal linguistic skill that humans use to express what they do not want. There is little research on exclusionary retrieval, where users express what they do not want to be part of the results produced for their queries. We investigate the scenario of exclusionary retrieval in document retrieval for the first time. We present ExcluIR, a set of resources for exclusionary retrieval, consisting of an evaluation benchmark and a training set for helping retrieval models to comprehend exclusionary queries. The evaluation benchmark includes 3,452 high-quality exclusionary queries, each of which has been manually annotated. The training set contains 70,293 exclusionary queries, each paired with a positive document and a negative document. We conduct detailed experiments and analyses, obtaining three main observations: (i) existing retrieval models with different architectures struggle to comprehend exclusionary queries effectively; (ii) although integrating our training data can improve the performance of retrieval models on exclusionary retrieval, there still exists a gap compared to human performance; and (iii) generative retrieval models have a natural advantage in handling exclusionary queries.