Attack success rate (ASR) is insufficient for evaluating the safety of large language models (LLMs). Many defenses rely on surface-level refusal behavior, which can over-refuse benign-sensitive prompts and remain brittle: once the initial refusal boundary is bypassed, the model may still follow a harmful continuation trajectory. We analyze model activation spaces to examine the limits of static interventions, finding substantial overlap among useful, benign-sensitive, and harmful representations. This suggests that a single global intervention may suppress harmful outputs only by perturbing general utility. We propose IRIS, Iterative Refinement of Internal States, a routing-conditioned internal correction method. To address this, IRIS leaves low-risk states unaltered while routing high-risk states to a shared module for multi-step latent refinement. To make the refinement module learn behavior-level correction, we introduce model-owned base-source prefix recovery, which trains IRIS to recover from base-model failure prefixes toward safe continuations. Extensive evaluations on Llama-3.1-8B and Qwen-2.5-7B demonstrate that IRIS drastically reduces ASR from around 70% to around 10% and successfully defends against prefilling attacks. Furthermore, IRIS preserves general benchmark utility and maintains near-zero over-refusal rates.

Continual multimodal instruction tuning requires multimodal large language models to acquire new task abilities sequentially while preserving previously learned knowledge. LoRA-MoE provides a promising solution by introducing expert-based capacity, but repeatedly learning and maintaining full LoRA experts leads to substantial parameter overhead. This raises a natural question: is full expert expansion necessary for every new task? To answer it, we analyze the SVD of task-specific LoRA updates and observe substantial overlap in their input- and output-side LoRA direction subspaces, with task-specific adaptation largely captured by lightweight coordinates over these subspaces. Motivated by this observation, we propose CoRe-MoE, a Compact Reusable MoE framework for parameter-efficient continual multimodal instruction tuning. CoRe-MoE extracts reusable input- and output-side direction bases from an initial expert bank, and for subsequent tasks trains only compact coordinate experts together with task-specific low-rank routers. Experiments on two representative MLLMs show that CoRe-MoE improves final average performance over the strongest competing baseline by up to 5.90 points, while using less than 1% of the trainable parameters required by sequential LoRA for later tasks.

Although scaling up the number of trainable parameters can effectively improve the training performance of large language models, it also leads to increased computational overhead. When delving into the parameter difference, we find that a subset of parameters, termed advantageous parameters, plays a crucial role in determining model performance. Further analysis reveals that stronger models tend to possess more such parameters. In this paper, we propose **A**dvantageous **P**arameter **EX**pansion Training (APEX), a method that progressively expands advantageous parameters into the space of disadvantageous ones, thereby increasing their proportion and enhancing training effectiveness, while keeping the total parameter count unchanged. Extensive experiments on both instruction tuning and continued pre-training across five base models demonstrate that, in instruction tuning, APEX outperforms full-parameter tuning while using only 52% of the trainable parameters. In continued pre-training, APEX achieves the same perplexity level as conventional training with only approximately 30% of the training data, and yields significant improvements on downstream tasks.

Built upon the existing analysis of retrieval heads in large language models, we propose an alternative reranking framework that trains models to estimate passage–query relevance using the attention scores of selected heads. This approach provides a listwise solution that leverages the holistic information within the entire candidate shortlist during ranking. At the same time, it naturally produces continuous relevance scores, enabling training on arbitrary retrieval datasets without requiring Likert-scale supervision. Our framework is lightweight and effective, requiring only small-scale models (e.g., 4B parameters) to achieve strong performance. Extensive experiments demonstrate that our method outperforms existing state-of-the-art pointwise and listwise rerankers across multiple domains, including Wikipedia and long narrative datasets. It further establishes a new state-of-the-art on the LoCoMo benchmark that assesses the capabilities of dialogue understanding and memory usage. We further demonstrate that our framework supports flexible extensions. For example, augmenting candidate passages with contextual information further improves ranking accuracy, while training attention heads from middle layers enhances efficiency without sacrificing performance.

Humans understand long and complex texts by relying on a holistic semantic representation of the content. This global view helps organize prior knowledge, interpret new information, and integrate evidence dispersed across a document, as revealed by the Mindscape-Aware Capability of humans in psychology. However, current Retrieval-Augmented Generation (RAG) systems often lack explicit global semantic guidance, making it difficult to retrieve and integrate dispersed evidence in long-context tasks. In this paper, we propose Mindscape-Aware RAG (MiA-RAG), the first framework to formulate mindscape-aware retrieval and generation as a unified conditioning paradigm for LLM-based RAG. MiA-RAG builds a mindscape through hierarchical summarization and conditions both retrieval and generation on this global semantic representation. This enables the retriever to form enriched query embeddings and the generator to reason over retrieved evidence within a coherent global context. We evaluate MiA-RAG across diverse long-context and bilingual benchmarks for evidence-based understanding and global sense-making. It consistently surpasses baselines, and further analysis shows that it aligns local details with a coherent global representation, enabling more human-like long-context retrieval and reasoning.

The Key-Value (KV) cache introduces substantial memory overhead during large language model (LLM) inference. Although existing vector quantization (VQ) methods reduce KV cache usage and provide flexible representational capacity across bit-widths, they suffer severe performance degradation at ultra-low bit-widths due to key cache outliers that hinder effective codebook utilization. To address this challenge, we propose VecInfer, a novel VQ method for aggressive KV cache compression while enabling efficient inference. By applying smooth and Hadamard transformations, VecInfer suppresses outliers in the key cache, enabling the codebook to comprehensively cover the original data distribution and thereby reducing quantization difficulty. To facilitate efficient deployment, we design an optimized CUDA kernel that fuses computation with dequantization to minimize memory access overhead. Extensive evaluations demonstrate that VecInfer consistently outperforms existing quantization baselines across both long-context understanding and mathematical reasoning tasks. With only 2-bit quantization, VecInfer achieves performance comparable to full precision, while delivering up to 2.7× speedup in large-batch self-attention computation and 8.3× reduction in single-batch end-to-end latency on Llama-3.1-8B with a 196k sequence length.

Despite the commendable progress of recent LLM-based data synthesis methods, they face two limitations in generating table instruction tuning data. First, they can not thoroughly explore the vast input space of table understanding tasks, leading to limited data diversity. Second, they ignore the weaknesses in table understanding ability of the target LLM and blindly pursue the increase of data quantity, resulting in suboptimal data efficiency. In this paper, we introduce a progressive and weakness-guided data synthesis framework tailored for table instruction tuning, named TableDreamer, to mitigate the above issues. Specifically, we first synthesize diverse tables and related instructions as seed data, and then perform an iterative exploration of the input space under the guidance of the newly identified weakness data, which eventually serve as the final training data for fine-tuning the target LLM. Extensive experiments on 10 tabular benchmarks demonstrate the effectiveness of the proposed framework, which boosts the average accuracy of Llama3.1-8B-instruct by 11.62% (49.07→60.69) with 27K GPT-4o synthetic data and outperforms state-of-the-art data synthesis baselines which use more training data.

The Key-Value (KV) cache in generative large language models (LLMs) introduces substantial memory overhead. Existing works mitigate this burden by offloading or compressing the KV cache. However, loading the entire cache incurs significant latency due to PCIe bandwidth bottlenecks in CPU-GPU communication, while aggressive compression causes notable performance degradation. We identify that certain layers in the LLM need to maintain global information and are unsuitable for selective loading. In contrast, other layers primarily focus on a few tokens with dominant activations that potentially incur substantial quantization error. This observation leads to a key insight that loading dominant tokens and quantizing all tokens can complement each other. Building on this insight, we propose a hybrid compression method, TailorKV, which seamlessly integrates quantization and offloading. TailorKV develops an inference framework along with a hardware-friendly implementation that leverages these complementary characteristics. Extensive long-context evaluations exhibit that TailorKV achieves nearly lossless performance under aggressive compression settings, outperforming the state-of-the-art. Particularly, the Llama-3.1-8B with 128k context can be served within a single RTX 3090 GPU, reaching 82 ms per token during decoding.

The high costs of customizing large language models (LLMs) fundamentally limit their adaptability to user-specific needs. Consequently, LLMs are increasingly offered as cloud-based services, a paradigm that introduces critical limitations: providers struggle to support personalized customization at scale, while users face privacy risks when exposing sensitive data. To address this dual challenge, we propose Customized Black-box Prompt Tuning (CBP-Tuning), a novel framework that facilitates efficient local customization while preserving bidirectional privacy. Specifically, we design a two-stage framework: (1) a prompt generator trained on the server-side to capture domain-specific and task-agnostic capabilities, and (2) user-side gradient-free optimization that tailors soft prompts for individual tasks. This approach eliminates the need for users to access model weights or upload private data, requiring only a single customized vector per task while achieving effective adaptation. Furthermore, the evaluation of CBP-Tuning in the commonsense reasoning, medical and financial domain settings demonstrates superior performance compared to baselines, showcasing its advantages in task-agnostic processing and privacy preservation.

Despite the efficacy of Direct Preference Optimization (DPO) in aligning Large Language Models (LLMs), reward hacking remains a pivotal challenge. This issue emerges when LLMs excessively reduce the probability of rejected completions to achieve high rewards, without genuinely meeting their intended goals. As a result, this leads to overly lengthy generation lacking diversity, as well as catastrophic forgetting of knowledge. We investigate the underlying reason behind this issue, which is representation redundancy caused by neuron collapse in the parameter space. Hence, we propose a novel Weights-Rotated Preference Optimization (RoPO) algorithm, which implicitly constrains the output layer logits with the KL divergence inherited from DPO and explicitly constrains the intermediate hidden states by fine-tuning on a multi-granularity orthogonal matrix. This design prevents the policy model from deviating too far from the reference model, thereby retaining the knowledge and expressive capabilities acquired during pre-training and SFT stages. Our RoPO achieves up to a 3.27-point improvement on AlpacaEval 2, and surpasses the best baseline by 6.2 to 7.5 points on MT-Bench with merely 0.015% of the trainable parameters, demonstrating its effectiveness in alleviating the reward hacking problem of DPO.

Large language models (LLMs) with the Mixture-of-Experts (MoE) architecture achieve high cost-efficiency by selectively activating a subset of the parameters. Despite the inference efficiency of MoE LLMs, the training of extensive experts from scratch incurs substantial overhead, whereas reconstructing a dense LLM into an MoE LLM significantly reduces the training budget. However, existing reconstruction methods often overlook the diversity among experts, leading to potential redundancy. In this paper, we come up with the observation that a specific LLM exhibits notable diversity after being pruned on different calibration datasets, based on which we present a Diversity-Enhanced reconstruction method named DIVE. The recipe of DIVE includes domain affinity mining, pruning-based expert reconstruction, and efficient retraining. Specifically, the reconstruction includes pruning and reassembly of the feed-forward network (FFN) module. After reconstruction, we efficiently retrain the model on routers, experts and normalization modules. We implement DIVE on Llama-style LLMs with open-source training corpora. Experiments show that DIVE achieves training efficiency with minimal accuracy trade-offs, outperforming existing pruning and MoE reconstruction methods with the same number of activated parameters. Code is available at: https://github.com/yuchenblah/DIVE.

Parameter-efficient fine-tuning (PEFT) has become a common method for fine-tuning large language models, where a base model can serve multiple users through PEFT module switching. To enhance user experience, base models require periodic updates. However, once updated, PEFT modules fine-tuned on previous versions often suffer substantial performance degradation on newer versions. Re-tuning these numerous modules to restore performance would incur significant computational costs. Through a comprehensive analysis of the changes that occur during base model updates, we uncover an interesting phenomenon: continual training primarily affects task-specific knowledge stored in Feed-Forward Networks (FFN), while having less impact on the task-specific pattern in the Attention mechanism. Based on these findings, we introduce Trans-PEFT, a novel approach that enhances the PEFT module by focusing on the task-specific pattern while reducing its dependence on certain knowledge in the base model. Further theoretical analysis supports our approach. Extensive experiments across 7 base models and 12 datasets demonstrate that Trans-PEFT trained modules can maintain performance on updated base models without re-tuning, significantly reducing maintenance overhead in real-world applications.

Due to the demand for efficient fine-tuning of large language models, Low-Rank Adaptation (LoRA) has been widely adopted as one of the most effective parameter-efficient fine-tuning methods. Nevertheless, while LoRA improves efficiency, there remains room for improvement in accuracy. Herein, we adopt a novel perspective to assess the characteristics of LoRA ranks. The results reveal that different ranks within the LoRA modules not only exhibit varying levels of importance but also evolve dynamically throughout the fine-tuning process, which may limit the performance of LoRA. Based on these findings, we propose BeamLoRA, which conceptualizes each LoRA module as a beam where each rank naturally corresponds to a potential sub-solution, and the fine-tuning process becomes a search for the optimal sub-solution combination. BeamLoRA dynamically eliminates underperforming sub-solutions while expanding the parameter space for promising ones, enhancing performance with a fixed rank. Extensive experiments across three base models and 12 datasets spanning math reasoning, code generation, and commonsense reasoning demonstrate that BeamLoRA consistently enhances the performance of LoRA, surpassing the other baseline methods.

Large Language Models (LLMs) excel in various domains but pose inherent privacy risks. Existing methods to evaluate privacy leakage in LLMs often use memorized prefixes or simple instructions to extract data, both of which well-alignment models can easily block. Meanwhile, Jailbreak attacks bypass LLM safety mechanisms to generate harmful content, but their role in privacy scenarios remains underexplored. In this paper, we examine the effectiveness of jailbreak attacks in extracting sensitive information, bridging privacy leakage and jailbreak attacks in LLMs. Moreover, we propose PIG, a novel framework targeting Personally Identifiable Information (PII) and addressing the limitations of current jailbreak methods. Specifically, PIG identifies PII entities and their types in privacy queries, uses in-context learning to build a privacy context, and iteratively updates it with three gradient-based strategies to elicit target PII. We evaluate PIG and existing jailbreak methods using two privacy-related datasets. Experiments on four white-box and two black-box LLMs show that PIG outperforms baseline methods and achieves state-of-the-art (SoTA) results. The results underscore significant privacy risks in LLMs, emphasizing the need for stronger safeguards.

Most existing Visual Question Answering (VQA) systems are constrained to support domain-specific questions, i.e., to train different models separately for different VQA tasks, thus generalizing poorly to others. For example, models trained on the reasoning-focused dataset GQA struggle to effectively handle samples from the knowledge-emphasizing dataset OKVQA. Meanwhile, in real-world scenarios, it is user-unfriendly to restrict the domain of questions. Therefore, this paper proposes a necessary task: One-to-Many Visual Question Answering, of which the ultimate goal is to enable a single model to answer as many different domains of questions as possible by the effective integration of available VQA resources. To this end, we first investigate into ten common VQA datasets, and break the task of VQA down into the integration of three key abilities.Then, considering assorted questions rely on different VQA abilities, this paper proposes a novel dynamic Mixture of LoRAs (MoL) strategy. MoL mixes three individually trained LoRA adapters (corresponding to each VQA ability) dynamically for different samples demanding various VQA abilities. The proposed MoL strategy is verified to be highly effective by experiments, establishing SOTAs on four datasets. In addition, MoL generalizes well to three extra zero-shot datasets.Data and codes will be released.

Structured pruning fundamentally reduces computational and memory overheads of large language models (LLMs) and offers a feasible solution for end-side LLM deployment. Structurally pruned models remain dense and high-precision, highly compatible with further tuning and compression. However, as the coarse-grained structured pruning poses large damage to the highly interconnected model, achieving a high compression ratio for scaled-up LLMs remains a challenge. In this paper, we introduce a task-agnostic structured pruning approach coupled with a compact Transformer architecture design. The proposed approach, named TransAct, reduces transitional activations inside multi-head attention (MHA) and multi-layer perceptron (MLP) modules, while preserving the inter-module activations that are sensitive to perturbations. Hence, the LLM is pruned into an intra-module low-rank architecture, significantly reducing weights, KV Cache and attention computation. TransAct is implemented on the LLaMA model and evaluated on downstream benchmarks. Results verify the optimality of our approach at high compression with respect to both efficiency and performance. Further, ablation studies reveal the strength of activation-guided iterative pruning and provide experimental analysis on the redundancy of MHA and MLP modules.

Parameter-efficient fine-tuning (PEFT) has emerged as the predominant technique for fine-tuning in the era of large language models. However, existing PEFT methods still have inadequate training efficiency. Firstly, the utilization of large-scale foundation models during the training process is excessively redundant for certain fine-tuning tasks. Secondly, as the model size increases, the growth in trainable parameters of empirically added PEFT modules becomes non-negligible and redundant, leading to inefficiency. To achieve task-specific efficient fine-tuning, we propose the Light-PEFT framework, which includes two methods: Masked Early Pruning of the Foundation Model and Multi-Granularity Early Pruning of PEFT. The Light-PEFT framework allows for the simultaneous estimation of redundant parameters in both the foundation model and PEFT modules during the early stage of training. These parameters can then be pruned for more efficient fine-tuning. We validate our approach on GLUE, SuperGLUE, QA tasks, and various models. With Light-PEFT, parameters of the foundation model can be pruned by up to over 40%, while still controlling trainable parameters to be only 25% of the original PEFT method. Compared to utilizing the PEFT method directly, Light-PEFT achieves training and inference speedup, reduces memory usage, and maintains comparable performance and the plug-and-play feature of PEFT.

Recent studies on the lottery ticket hypothesis (LTH) show that pre-trained language models (PLMs) like BERT contain matching subnetworks that have similar transfer learning performance as the original PLM. These subnetworks are found using magnitude-based pruning. In this paper, we find that the BERT subnetworks have even more potential than these studies have shown. Firstly, we discover that the success of magnitude pruning can be attributed to the preserved pre-training performance, which correlates with the downstream transferability. Inspired by this, we propose to directly optimize the subnetwork structure towards the pre-training objectives, which can better preserve the pre-training performance. Specifically, we train binary masks over model weights on the pre-training tasks, with the aim of preserving the universal transferability of the subnetwork, which is agnostic to any specific downstream tasks. We then fine-tune the subnetworks on the GLUE benchmark and the SQuAD dataset. The results show that, compared with magnitude pruning, mask training can effectively find BERT subnetworks with improved overall performance on downstream tasks. Moreover, our method is also more efficient in searching subnetworks and more advantageous when fine-tuning within a certain range of data scarcity. Our code is available at https://github.com/llyx97/TAMT.

In zero-shot multilingual extractive text summarization, a model is typically trained on English summarization dataset and then applied on summarization datasets of other languages. Given English gold summaries and documents, sentence-level labels for extractive summarization are usually generated using heuristics. However, these monolingual labels created on English datasets may not be optimal on datasets of other languages, for that there is the syntactic or semantic discrepancy between different languages. In this way, it is possible to translate the English dataset to other languages and obtain different sets of labels again using heuristics. To fully leverage the information of these different sets of labels, we propose NLSSum (Neural Label Search for Summarization), which jointly learns hierarchical weights for these different sets of labels together with our summarization model. We conduct multilingual zero-shot summarization experiments on MLSUM and WikiLingua datasets, and we achieve state-of-the-art results using both human and automatic evaluations across these two datasets.