Large Language Models (LLMs) are primarily constrained by memory and bandwidth bottlenecks during deployment. Although Vector Quantization (VQ) has emerged as a promising solution, existing methods incur inference overhead due to massive codebook storage and intensive index lookups. Moreover, these methods typically suffer from non-negligible performance degradation under ultra-low bitwidth regimes. To bridge this gap, we propose Sparse-Compensated Vector Quantization (SCVQ), a novel framework designed for high-efficiency LLM vector quantization. SCVQ introduces a salience-aware weighted K-means clustering scheme with symmetry constraints to reduces codebook size and indexing costs. Central to our approach is a unified structured representation that consolidates outliers, salient weights, and quantization residuals into a single sparse compensation matrix. This design effectively preserves critical model information while leveraging VQ-specific properties to enable efficient custom kernels. Extensive experiments across multiple benchmarks demonstrate SCVQ’s superior performance. Specifically, SCVQ achieves a perplexity of 5.78 on WikiText-2 for LLaMA-2-7B at 2-bit quantization, while delivering a 1.4× end-to-end inference speedup over existing baselines.

Safety alignment—training large language models (LLMs) to refuse harmful requests while remaining helpful—is critical for responsible deployment. Prior work established that safety behaviors are governed by low-rank structures, suggesting parameter-efficient fine-tuning (PEFT) should be well-suited for alignment. However, Low-Rank Adaptation (LoRA) consistently underperforms full fine-tuning and reinforcement learning on safety benchmarks. We attribute this gap to semantic entanglement: safety-relevant directions are intertwined with unrelated concepts due to polysemanticity, impeding implicit subspace identification. To address this, we propose SAILS (Safety Alignment via Interpretable Low-rank Subspace), which leverages Sparse Autoencoders (SAEs) to disentangle representations into monosemantic features, constructs an interpretable safety subspace from SAE decoder directions, and uses it to initialize LoRA adapters. Theoretically, we prove that SAE-based identification achieves arbitrarily small recovery error under monosemanticity assumptions, while direct identification suffers an irreducible error floor. Empirically, SAILS achieves up to 99.6% safety rates across multiple model families and scales, exceeding full fine-tuning and matching RLHF-based models, with only 0.2% of parameters updated and providing interpretability.

Large language models (LLMs) have achieved remarkable performance across diverse tasks, yet ensuring output safety remains a fundamental challenge. Existing defense methods often suffer from limited generalization, high computational overhead, or significant utility degradation. In this work, we present SecDecoding, a lightweight decoding-time defense framework that significantly improves output safety without compromising model helpfulness. SecDecoding leverages a pair of small contrastive models, namely a base model and a safety fine-tuned expert, to estimate token-level safety signals by measuring divergence in their output distributions. These signals dynamically steer the target model’s generation toward safer trajectories, effectively suppressing unsafe content. Experimental results show that SecDecoding achieves near-zero attack success rates against a wide spectrum of advanced jailbreak attacks across multiple LLMs, while maintaining the model’s helpfulness with minimal degradation. Additionally, SecDecoding is a modular and resource-efficient approach that requires only an auxiliary 1-billion-parameter model and is compatible with speculative decoding, offering up to 1.5× inference speedup.

Large language models (LLMs) have demonstrated remarkable reasoning and planning capabilities, driving extensive research into task decomposition. Existing task decomposition methods focus primarily on memory, tool usage, and feedback mechanisms, achieving notable success in specific domains, but they often overlook the trade-off between performance and cost. In this study, we first conduct a comprehensive investigation on task decomposition, identifying six categorization schemes. Then, we perform an empirical analysis of three factors that influence the performance and cost of task decomposition: categories of approaches, characteristics of tasks, and configuration of decomposition and execution models, uncovering three critical insights and summarizing a set of practical principles. Building on this analysis, we propose the Select-Then-Decompose strategy, which establishes a closed-loop problem-solving process composed of three stages: selection, execution, and verification. This strategy dynamically selects the most suitable decomposition approach based on task characteristics and enhances the reliability of the results through a verification module. Comprehensive evaluations across multiple benchmarks show that the Select-Then-Decompose consistently lies on the Pareto frontier, demonstrating an optimal balance between performance and cost. Our code is publicly available at https://github.com/summervvind/Select-Then-Decompose.

Recent advancements in large language models (LLMs) have revealed their potential for achieving autonomous agents possessing human-level intelligence. However, existing benchmarks for evaluating LLM Agents either use static datasets, potentially leading to data leakage or focus only on single-agent scenarios, overlooking the complexities of multi-agent interactions. There is a lack of a benchmark that evaluates the diverse capabilities of LLM agents in multi-agent, dynamic environments. To this end, we introduce LLMArena, a novel and easily extensible framework for evaluating the diverse capabilities of LLM in multi-agent dynamic environments. LLMArena encompasses seven distinct gaming environments, employing Trueskill scoring to assess crucial abilities in LLM agents, including spatial reasoning, strategic planning, numerical reasoning, risk assessment, communication, opponent modeling, and team collaboration. We conduct an extensive experiment and human evaluation among different sizes and types of LLMs, showing that LLMs still have a significant journey ahead in their development towards becoming fully autonomous agents, especially in opponent modeling and team collaboration. We hope LLMArena could guide future research towards enhancing these capabilities in LLMs, ultimately leading to more sophisticated and practical applications in dynamic, multi-agent settings.

The Mixture of Experts (MoE) for language models has been proven effective in augmenting the capacity of models by dynamically routing each input token to a specific subset of experts for processing. Despite the success, most existing methods face a challenge for balance between sparsity and the availability of expert knowledge: enhancing performance through increased use of expert knowledge often results in diminishing sparsity during expert selection. To mitigate this contradiction, we propose HyperMoE, a novel MoE framework built upon Hypernetworks. This framework integrates the computational processes of MoE with the concept of knowledge transferring in multi-task learning. Specific modules generated based on the information of unselected experts serve as supplementary information, which allows the knowledge of experts not selected to be used while maintaining selection sparsity. Our comprehensive empirical evaluations across multiple datasets and backbones establish that HyperMoE significantly outperforms existing MoE methods under identical conditions concerning the number of experts. Our code is publicly available at https://github.com/Bumble666/Hyper_MoE

Drawing upon the intuition that aligning different modalities to the same semantic embedding space would allow models to understand states and actions more easily, we propose a new perspective to the offline reinforcement learning (RL) challenge. More concretely, we transform it into a supervised learning task by integrating multimodal and pre-trained language models. Our approach incorporates state information derived from images and action-related data obtained from text, thereby bolstering RL training performance and promoting long-term strategic thinking. We emphasize the contextual understanding of language and demonstrate how decision-making in RL can benefit from aligning states’ and actions’ representation with languages’ representation. Our method significantly outperforms current baselines as evidenced by evaluations conducted on Atari and OpenAI Gym environments. This contributes to advancing offline RL performance and efficiency while providing a novel perspective on offline RL.

The de facto way of utilizing black-box large language models (LLMs) to perform various downstream tasks is prompting. However, obtaining suitable prompts for specific tasks is still a challenging problem. While existing LLM-based methods demonstrate promising performance in task-oriented dialogue (TOD) task, they often require manual adjustment in prompt selection, or focus solely on dialogue understanding or generation. To address these issues, we propose an adaptive prompt generation framework to fully unleash the potential of LLMs for the comprehensive TOD system. Firstly, we design a trainable slot generator (TSG) that can generate domain and slot information in the belief state, which serves as prior knowledge for subsequent prompt generation. Next, we propose an adaptive prompt generator (APG) that utilizes the prior knowledge to generate prompts for the LLM, deriving the belief state and system response of the dialogue for evaluation. Finally, we evaluate our framework on the MultiWOZ 2.0 dataset. Extensive experiments demonstrate that our method outperforms existing methods. Our code and data will be released.

This paper explores interactive facial image editing through dialogue and presents the ChatEdit benchmark dataset for evaluating image editing and conversation abilities in this context. ChatEdit is constructed from the CelebA-HQ dataset, incorporating annotated multi-turn dialogues corresponding to user editing requests on the images. The dataset is challenging, as it requires the system to dynamically track and edit images based on user requests, while generating appropriate natural language responses. To address these challenges, we propose a framework comprising a dialogue module for tracking user requests as well as generating responses, and an image editing module for editing images accordingly. Unlike previous approaches, our framework directly tracks the user request of the current turn from the entire dialogue history and edits the initial image instead of manipulating the output from the previous turn, mitigating error accumulation and attribute forgetting issues. Extensive experiments on the ChatEdit dataset demonstrate the superiority of our framework over previous methods and also improvement rooms, encouraging future research. We will release the code and data publicly to facilitate advancements in complex interactive facial image editing.

Parameter-efficient fine-tuning (PEFT) has shown its effectiveness in adapting the pre-trained language models to downstream tasks while only updating a small number of parameters. Despite the success, most existing methods independently adapt to each task without considering knowledge transfer between tasks and are limited to low-data regimes. To overcome this issue, we propose Prototype-based HyperAdapter (PHA), a novel framework built on the adapter-tuning and hypernetwork. It introduces an instance-dense retriever and a prototypical hypernetwork to generate the conditional modules in a sample-efficient manner. This leads to comparable performance improvements against existing PEFT methods on multi-task learning and few-shot transfer learning. More importantly, when the available data size gets smaller, our method outperforms other strong baselines by a large margin. Based on our extensive empirical experiments across various datasets, we demonstrate that PHA strikes a better trade-off between trainable parameters, accuracy on stream tasks, and sample efficiency. Our code is publicly available at https://github.com/Bumble666/PHA