Large Language Model (LLM)-based agents have achieved promising performance in web task planning. However, they primarily focus on task completion rather than preference modeling, leading to imprecise personalization. This can be attributed to two facts: (1) Contextual scarcity. Vague instructions and insufficient background of users’ inputs make it difficult to accurately identify users’ intents. (2) Preference unpredictability. It’s hard for a single response path to handle users’ sudden intent changes. To address these issues, we propose SAPA, a Situation-Augmented Personalized web Agent. Specifically, a Situational Intent Grounding module is proposed to enrich the contexts of users’ inputs by incorporating historical records and candidate constraints, effectively alleviating the contextual scarcity. Then, we propose a Task-Driven Semantic Expansion module to explore diverse response paths and reduce preference unpredictability in turn, which is achieved by employing a test-time expansion strategy. We construct SAPA with representative LLMs and conduct extensive experiments on PersonalWAB and Amazon Reviews datasets. On PersonalWAB, SAPA achieves new state-of-the-art performance with improvements of +14.0%, +8.8%, and +1.3% in Overall score on personalized recommendation, search, and question answering, respectively. We further evaluate on three single-category scenarios from Amazon Reviews (Office Products, Beauty and Personal Care, Electronics), where SAPA consistently improves Overall scores by 16%–18% with gains across all task types. Ablation studies and case analyses further confirm that the gains stem from jointly addressing contextual scarcity and preference unpredictability.

Providing extensive context via prompting is vital for leveraging the capabilities of Large Language Models (LLMs). However, lengthy contexts significantly increase inference latency. To mitigate this issue, context compression—particularly soft prompt compression—has emerged as a widely studied solution, which converts long contexts into shorter memory embeddings via a trained compressor. Existing methods typically compress the entire context indiscriminately into a set of memory tokens, requiring the compressor to capture global dependencies and necessitating extensive pre-training data to learn effective patterns. Inspired by the chunking mechanism in human working memory and our empirical observation of spatial specialization within memory embeddings, we propose Parallelized Iterative Compression (PIC). By simply modifying the Transformer’s attention mask, PIC explicitly restricts the receptive field of memory tokens to sequential local chunks, thereby significantly reducing the difficulty of compressor training. Experiments across multiple downstream tasks demonstrate that PIC outperforms competitive baselines while dramatically expediting the training process. Finally, we provide an investigation into the emergence of spatial specialization, revealing how tuning strategies and explicit structural priors shape the latent structure of the memory embeddings.

Low-Rank Adaptation (LoRA) is widely used for fine-tuning large language models, but suffers from catastrophic forgetting. Existing continual learning methods mitigate this by estimating per-parameter importance independently. However, this assumption breaks down under LoRA’s factorization 𝛥 W = BA: infinitely many pairs (B, A) yield the same weight update, making importance scores on A and B separately unreliable. We propose HLoRA, which computes path-integral importance in the combined weight update space, bypassing the unreliable separate estimation on A and B. HLoRA applies hierarchical regularization at two levels: element-wise constraints preserve parameters critical for general knowledge, and layer-wise coefficients adapt regularization strength based on each layer’s aggregate importance. Experiments on scientific, medical, and physical tasks using GPT-J and LLaMA-3 show that HLoRA mitigates catastrophic forgetting while enhancing adaptability, running approximately 20x faster and requiring only 10%–15% of the storage compared to prior methods. Code is released on https://anonymous.4open.science/r/HLoRA-915F.

Small Language Models (SLMs) have been widely deployed in edge computing and mobile devices. Generally, these models need to be continuously updated to adapt to dynamic environments and evolving user needs. Therefore, the performance of SLMs in continual learning has become a critical research focus. Compared to LLMs with rich redundant representation spaces, SLMs are more likely to overwrite previous representations during new task learning due to their fewer attention heads and layers. To address this issue, we propose LightLoRA, a sparse fine-tuning method specifically designed for SLMs. This method constructs a conflict-aware importance mechanism by combining raw importance from SVD decomposition with task conflict degrees based on gradient direction consistency. By selecting and updating heads that provide high contributions and low conflicts, LightLoRA successfully maintains sufficient space for knowledge preservation. Experimental results show that the proposed method reaches state-of-the-art performance across multiple benchmarks without adding extra parameters. Code is released on https://anonymous.4open.science/r/LightLoRA-077C.

Despite extensive safety alignment, Large Language Models (LLMs) often fail against jailbreak attacks. While machine unlearning has emerged as a promising defense by erasing specific harmful parameters, current methods remain vulnerable to diverse jailbreaks. We first conduct an empirical study and discover that this failure mechanism is caused by jailbreaks primarily activating non-erased parameters in the intermediate layers. Further, by probing the underlying mechanism through which these circumvented parameters reassemble into the prohibited output, we verify the persistent existence of dynamic jailbreak paths and show that the inability to rectify them constitutes the fundamental gap in existing unlearning defenses. To bridge this gap, we propose Jailbreak Path Unlearning (JPU), which is the first to rectify dynamic jailbreak paths towards safety anchors by dynamically mining on-policy adversarial samples to expose vulnerabilities and identify jailbreak paths. Extensive experiments demonstrate that JPU significantly enhances jailbreak resistance against dynamic attacks while preserving the model’s utility.

Large language models (LLMs) generate human-aligned content under certain safety constraints. However, the current known technique “jailbreak prompt” can circumvent safety-aligned measures and induce LLMs to output malicious content. Research on Jailbreaking can help identify vulnerabilities in LLMs and guide the development of robust security frameworks. To circumvent the issue of attack templates becoming obsolete as models evolve, existing methods adopt iterative mutation and dynamic optimization to facilitate more automated jailbreak attacks. However, these methods face two challenges: inefficiency and repetitive optimization, as they overlook the value of past attack experiences. To better integrate past attack experiences to assist current jailbreak attempts, we propose the JailExpert, an automated jailbreak framework, which is the first to achieve a formal representation of experience structure, group experiences based on semantic drift, and support the dynamic updating of the experience pool. Extensive experiments demonstrate that JailExpert significantly improves both attack effectiveness and efficiency. Compared to the current state-of-the-art black-box jailbreak method, JailExpert achieves an average increase of 24% in attack success rate and 2.7 times improvement in attack efficiency.

Few-shot named entity recognition (NER) enables us to build a NER system for a new domain using very few labeled examples. However, existing prototypical networks for this task suffer from roughly estimated label dependency and closely distributed prototypes, thus often causing misclassifications. To address the above issues, we propose EP-Net, an Entity-level Prototypical Network enhanced by dispersedly distributed prototypes. EP-Net builds entity-level prototypes and considers text spans to be candidate entities, so it no longer requires the label dependency. In addition, EP-Net trains the prototypes from scratch to distribute them dispersedly and aligns spans to prototypes in the embedding space using a space projection. Experimental results on two evaluation tasks and the Few-NERD settings demonstrate that EP-Net consistently outperforms the previous strong models in terms of overall performance. Extensive analyses further validate the effectiveness of EP-Net.

Span-based joint extraction models have shown their efficiency on entity recognition and relation extraction. These models regard text spans as candidate entities and span tuples as candidate relation tuples. Span semantic representations are shared in both entity recognition and relation extraction, while existing models cannot well capture semantics of these candidate entities and relations. To address these problems, we introduce a span-based joint extraction framework with attention-based semantic representations. Specially, attentions are utilized to calculate semantic representations, including span-specific and contextual ones. We further investigate effects of four attention variants in generating contextual semantic representations. Experiments show that our model outperforms previous systems and achieves state-of-the-art results on ACE2005, CoNLL2004 and ADE.