Pruning is an effective method for compressing Large Language Models, but finding an optimal, non-uniform layer-wise sparsity allocation remains a key challenge. While heuristic methods are fast but yield suboptimal performance, more powerful search-based approaches like Reinforcement Learning are often hindered by prohibitive computational costs on large-scale models. To overcome this efficiency barrier, we propose FastForward Pruning. Its core is a decoupled, single-step RL framework that separates policy optimization from the complex budget satisfaction problem. Such a decoupling is crucial for efficiently searching the vast policy space of LLMs. This curriculum-based strategy begins with low-cost, simple tasks and gradually increases in complexity, significantly reducing the search's computational overhead. Evaluated on the LLaMA, Mistral, and OPT model families, our framework discovers pruning policies that achieve superior performance over strong heuristic baselines. Crucially, when compared to other search-based algorithms, our method achieves competitive or superior results at a fraction of the computational cost, demonstrating a clear advantage in search efficiency.

Vision-Language-Action (VLA) models have emerged as a powerful paradigm in Embodied AI. However, the significant computational overhead of processing redundant visual tokens remains a critical bottleneck for real-time robotic deployment. While standard token pruning techniques can alleviate this, these task-agnostic methods struggle to preserve task-critical visual information. To address this challenge, simultaneously preserving both the holistic context and fine-grained details for precise action, we propose Compressor-VLA, a novel hybrid instruction-conditioned token compression framework designed for efficient, task-oriented compression of visual information in VLA models. The proposed Compressor-VLA framework consists of two token compression modules: a Semantic Task Compressor (STC) that distills holistic, task-relevant context, and a Spatial Refinement Compressor (SRC) that preserves fine-grained spatial details. This compression is dynamically modulated by the natural language instruction, allowing for the adaptive condensation of task-relevant visual information. Experimentally, extensive evaluations demonstrate that Compressor-VLA achieves a competitive success rate on the LIBERO benchmark while reducing FLOPs by 59% and the visual token count by over 3x compared to its baseline. The real-robot deployments on a dual-arm robot platform validate the model's sim-to-real transferability and practical applicability. Moreover, qualitative analyses reveal that our instruction guidance dynamically steers the model's perceptual focus toward task-relevant objects, thereby validating the effectiveness of our approach.

Large Language Models (LLMs) face a significant bottleneck during autoregressive inference due to the massive memory footprint of the Key-Value (KV) cache. Existing compression techniques like token eviction, quantization, or other low-rank methods often risk information loss, have fixed limits, or introduce significant computational overhead from explicit decompression steps. In this work, we introduce SWAN, a novel, fine-tuning-free framework that eliminates this overhead. Our method uses an offline orthogonal matrix to rotate and prune the KV-cache, which is then used directly in the attention computation without any reconstruction. Our extensive experiments demonstrate that SWAN, augmented with a small dense buffer, offers a robust trade-off, maintaining performance close to the uncompressed baseline even at aggressive 50-60% memory savings per-token on KV-cache. A key advantage is its runtime-tunable compression level, allowing operators to dynamically adjust the memory footprint, a flexibility absent in methods requiring fixed offline configurations. This combination of a decompression-free design, high performance under compression, and adaptability makes SWAN a practical and efficient solution for serving LLMs with long contexts.

The task of translating natural language questions into query languages has long been a central focus in semantic parsing. Recent advancements in Large Language Models (LLMs) have significantly accelerated progress in this field. However, existing studies typically focus on a single query language, resulting in methods with limited generalizability across different languages. In this paper, we formally define the Text-to-Query task paradigm, unifying semantic parsing tasks across various query languages. We identify query skeletons as a shared optimization target of Text-to-Query tasks, and propose a general dynamic data augmentation framework that explicitly diagnoses model-specific weaknesses in handling these skeletons to synthesize targeted training data. Experiments on four Text-to-Query benchmarks demonstrate that our method achieves state-of-the-art performance using only a small amount of synthesized data, highlighting the efficiency and generality of our approach and laying a solid foundation for unified research on Text-to-Query tasks. We release our code at https://github.com/jjjycaptain/Skeletron.

Large Language Models (LLMs) remain susceptible to jailbreak exploits that bypass safety filters and induce harmful or unethical behavior. This work presents a systematic taxonomy of existing jailbreak defenses across prompt-level, model-level, and training-time interventions, followed by three proposed defense strategies. First, a Prompt-Level Defense Framework detects and neutralizes adversarial inputs through sanitization, paraphrasing, and adaptive system guarding. Second, a Logit-Based Steering Defense reinforces refusal behavior through inference-time vector steering in safety-sensitive layers. Third, a Domain-Specific Agent Defense employs the MetaGPT framework to enforce structured, role-based collaboration and domain adherence. Experiments on benchmark datasets show substantial reductions in attack success rate, achieving full mitigation under the agent-based defense. Overall, this study highlights how jailbreaks pose a significant security threat to LLMs and identifies key intervention points for prevention, while noting that defense strategies often involve trade-offs between safety, performance, and scalability. Code is available at: https://github.com/Kuro0911/CS5446-Project

Modern large language models integrate web search to provide real-time answers, yet it remains unclear whether they are efficiently calibrated to use search when it is actually needed. We introduce a benchmark evaluating both the necessity and effectiveness of web access across commercial models with no access to internal states or parameters. The dataset includes a static split of 783 temporally anchored questions answerable from pre-cutoff knowledge, aimed at testing whether models invoke search based on low internal confidence, and a dynamic split of 288 post-cutoff queries designed to test whether models recognise when search is required and retrieve updated information. Web access substantially improves static accuracy for GPT-5-mini and Claude Haiku 4.5, though confidence calibration worsens. On dynamic queries, both models frequently invoke search yet remain below 70 percent accuracy due to weak query formulation. Costs per accuracy-improving call remain low, but returns diminish once initial retrieval fails. Selective invocation helps, but models become overconfident and inconsistent after search. Overall, built-in web search meaningfully improves factual accuracy and can be invoked selectively, yet models remain overconfident, skip retrieval when it is essential, and falter once initial search queries underperform. Taken together, internal web search works better as a good low-latency verification layer than a reliable analytical tool, with clear room for improvement.

With the rapid development of Large Language Models, dialogue systems are shifting from information tools to emotional companions, heralding the era of Emotional Companionship Dialogue Systems (ECDs) that provide personalized emotional support for users. However, the field lacks clear definitions and systematic evaluation standards for ECDs. To address this, we first propose a definition of ECDs with formal descriptions. Then, based on this theory and the design principle of "Ability Layer-Task Layer (three level)-Data Layer-Method Layer", we design and implement the first ECD evaluation benchmark - MoodBench 1.0. Through extensive evaluations of 30 mainstream models, we demonstrate that MoodBench 1.0 has excellent discriminant validity and can effectively quantify the differences in emotional companionship abilities among models. Furthermore, the results reveal current models' shortcomings in deep emotional companionship, guiding future technological optimization and significantly aiding developers in enhancing ECDs' user experience.

Linux kernel evolution breaks drivers through API/ABI changes, semantic shifts, and security-hardening updates. We introduce DRIVEBENCH, an executable corpus of kernel$\rightarrow$driver co-evolution cases, and AUTODRIVER, a closed-loop, LLM-driven system for automating driver maintenance. The system integrates prompt engineering, multi-agent collaboration, static analysis, and iterative validation to ensure that generated patches are not only syntactically correct but also functionally and semantically consistent with kernel conventions. The corpus spans v5.10-v6.10 with 235 validated cases drawn from 612 candidates. In evaluation across 55 cases, AUTODRIVER achieves 56.4% compilation success; QEMU-based boot verification indicates that compiled patches preserve driver initialization in most instances. By releasing DRIVEBENCH and tooling, we enable reproducible research and a practical route to continuous, safe co-evolution of drivers with the Linux kernel.

Due to the scarcity of high-quality data, large language models (LLMs) are often trained on mixtures of data with varying quality levels, even after sophisticated data curation. A natural approach to better leverage high-quality data is curriculum-based pretraining, where the model is trained on data sorted in ascending order of quality as determined by a quality metric. However, prior studies have reported limited improvements from such curriculum-based pretraining strategies. This work identifies a critical factor constraining these methods: the incompatibility between the ascending data quality order and the decaying learning rate (LR) schedule. We find that while curriculum-based training substantially outperforms random shuffling when using a constant LR, its advantage diminishes under standard LR decay schedules. Our experiments show this incompatibility can be mitigated by two simple strategies: (1) employing a more moderate LR decay schedule, where the final LR is only moderately smaller than the peak LR, and (2) replacing LR decay with model averaging, i.e., computing a weighted average of the final few checkpoints. By combining these strategies, we improve the average score on a suite of standard benchmarks by 1.64% over random shuffling, without additional data refinement. Validated on 1.5B-parameter models trained over 30B tokens with various data-quality metrics, our findings call for a re-evaluation of curriculum-based LLM pretraining and underscore the potential of co-designing data curricula with optimization methods.

Group-based policy optimization methods like GRPO and GSPO have become standard for training multimodal models, leveraging group-wise rollouts and relative advantage estimation. However, they suffer from a critical \emph{gradient vanishing} problem when all responses within a group receive identical rewards, causing advantage estimates to collapse and training signals to diminish. Existing attempts to mitigate this issue fall into two paradigms: filtering-based and sampling-based methods. Filtering-based methods first generate rollouts broadly and then retroactively filter out uninformative groups, leading to substantial computational overhead. Sampling-based methods proactively select effective samples before rollout but rely on static criteria or prior dataset knowledge, lacking real-time adaptability. To address these issues, we propose \textbf{VADE}, a \textbf{V}ariance-\textbf{A}ware \textbf{D}ynamic sampling framework via online sample-level difficulty \textbf{E}stimation. Our framework integrates three key components: online sample-level difficulty estimation using Beta distributions, a Thompson sampler that maximizes information gain through the estimated correctness probability, and a two-scale prior decay mechanism that maintains robust estimation under policy evolution. This three components design enables VADE to dynamically select the most informative samples, thereby amplifying training signals while eliminating extra rollout costs. Extensive experiments on multimodal reasoning benchmarks show that VADE consistently outperforms strong baselines in both performance and sample efficiency, while achieving a dramatic reduction in computational overhead. More importantly, our framework can serves as a plug-and-play component to be seamlessly integrated into existing group-based RL algorithms. Code and models are available at https://VADE-RL.github.io.