As large language models (LLMs) are increasingly deployed across diverse domains, ensuring their safety has become a critical concern. In response, studies on jailbreak attacks have been actively growing. Existing approaches typically rely on iterative prompt engineering or semantic transformations of harmful instructions to evade detection. In this work, we introduce PUZZLED, a novel jailbreak method that leverages the LLM's reasoning capabilities. It masks keywords in a harmful instruction and presents them as word puzzles for the LLM to solve. We design three puzzle types-word search, anagram, and crossword-that are familiar to humans but cognitively demanding for LLMs. The model must solve the puzzle to uncover the masked words and then proceed to generate responses to the reconstructed harmful instruction. We evaluate PUZZLED on five state-of-the-art LLMs and observe a high average attack success rate (ASR) of 88.8%, specifically 96.5% on GPT-4.1 and 92.3% on Claude 3.7 Sonnet. PUZZLED is a simple yet powerful attack that transforms familiar puzzles into an effective jailbreak strategy by harnessing LLMs' reasoning capabilities.

The reasoning and planning abilities of Large Language Models (LLMs) have been a frequent topic of discussion in recent years. Their ability to take unstructured planning problems as input has made LLMs' integration into AI planning an area of interest. Nevertheless, LLMs are still not reliable as planners, with the generated plans often containing mistaken or hallucinated actions. Existing benchmarking and evaluation methods investigate planning with LLMs, focusing primarily on success rate as a quality indicator in various planning tasks, such as validating plans or planning in relaxed conditions. In this paper, we approach planning with LLMs as a natural language processing (NLP) task, given that LLMs are NLP models themselves. We propose a recovery pipeline consisting of an NLP-based evaluation of the generated plans, along with three stages to recover the plans through NLP manipulation of the LLM-generated plans, and eventually complete the plan using a symbolic planner. This pipeline provides a holistic analysis of LLM capabilities in the context of AI task planning, enabling a broader understanding of the quality of invalid plans. Our findings reveal no clear evidence of underlying reasoning during plan generation, and that a pipeline comprising an NLP-based analysis of the plans, followed by a recovery mechanism, still falls short of the quality and reliability of classical planners. On average, only the first 2.65 actions of the plan are executable, with the average length of symbolically generated plans being 8.4 actions. The pipeline still improves action quality and increases the overall success rate from 21.9% to 27.5%.

Retrieval-Augmented Generation (RAG) shows impressive performance by supplementing and substituting parametric knowledge in Large Language Models (LLMs). Retrieved knowledge can be divided into three types: explicit answer evidence, implicit answer clue, and insufficient answer context which can be further categorized into totally irrelevant and partially relevant information. Effectively utilizing partially relevant knowledge remains a key challenge for RAG systems, especially in incomplete knowledge base retrieval. Contrary to the conventional view, we propose a new perspective: LLMs can be awakened via partially relevant knowledge already embedded in LLMs. To comprehensively investigate this phenomenon, the triplets located in the gold reasoning path and their variants are used to construct partially relevant knowledge by removing the path that contains the answer. We provide theoretical analysis of the awakening effect in LLMs and support our hypothesis with experiments on two Knowledge Graphs (KGs) Question Answering (QA) datasets. Furthermore, we present a new task, Unseen Entity KGQA, simulating real-world challenges where entity linking fails due to KG incompleteness. Our awakening-based approach demonstrates greater efficacy in practical applications, outperforms traditional methods that rely on embedding-based similarity which are prone to returning noisy information.

Automated bioacoustic analysis is essential for biodiversity monitoring and conservation, requiring advanced deep learning models that can adapt to diverse bioacoustic tasks. This article presents a comprehensive review of large-scale pretrained bioacoustic foundation models and systematically investigates their transferability across multiple bioacoustic classification tasks. We overview bioacoustic representation learning including major pretraining data sources and benchmarks. On this basis, we review bioacoustic foundation models by thoroughly analysing design decisions such as model architecture, pretraining scheme, and training paradigm. Additionally, we evaluate selected foundation models on classification tasks from the BEANS and BirdSet benchmarks, comparing the generalisability of learned representations under both linear and attentive probing strategies. Our comprehensive experimental analysis reveals that BirdMAE, trained on large-scale bird song data with a self-supervised objective, achieves the best performance on the BirdSet benchmark. On BEANS, BEATs$_{NLM}$, the extracted encoder of the NatureLM-audio large audio model, is slightly better. Both transformer-based models require attentive probing to extract the full performance of their representations. ConvNext$_{BS}$ and Perch models trained with supervision on large-scale bird song data remain competitive for passive acoustic monitoring classification tasks of BirdSet in linear probing settings. Training a new linear classifier has clear advantages over evaluating these models without further training. While on BEANS, the baseline model BEATs trained with self-supervision on AudioSet outperforms bird-specific models when evaluated with attentive probing. These findings provide valuable guidance for practitioners selecting appropriate models to adapt them to new bioacoustic classification tasks via probing.

Multimodal Large Language Models (MLLMs) process visual, acoustic, and textual inputs, addressing the limitations of single-modality LLMs. However, existing benchmarks often overlook tri-modal evaluation in Traditional Chinese and do not consider inference latency. To address this, we introduce Multi-TW, the first Traditional Chinese benchmark for evaluating the performance and latency of any-to-any multimodal models. Multi-TW includes 900 multiple-choice questions (image and text, audio and text pairs) sourced from official proficiency tests developed with the Steering Committee for the Test of Proficiency-Huayu (SC-TOP). We evaluated various any-to-any models and vision-language models (VLMs) with audio transcription. Our results show that closed-source models generally outperform open-source ones across modalities, although open-source models can perform well in audio tasks. End-to-end any-to-any pipelines offer clear latency advantages compared to VLMs using separate audio transcription. Multi-TW presents a comprehensive view of model capabilities and highlights the need for Traditional Chinese fine-tuning and efficient multimodal architectures.

We present MoE-MLA-RoPE, a novel architecture combination that combines Mixture of Experts (MoE) with Multi-head Latent Attention (MLA) and Rotary Position Embeddings (RoPE) for efficient language modeling. Our approach addresses the fundamental trade-off between model capacity and computational efficiency through three key innovations: (1) fine-grained expert routing with 64 micro-experts and top-$k$ selection, enabling flexible specialization through 3.6 * 10^7 possible expert combinations; (2) shared expert isolation that dedicates 2 always active experts for common patterns while routing to 6 of 62 specialized experts; and (3) gradient-conflict-free load balancing that maintains expert utilization without interfering with primary loss optimization. Extensive experiments on models ranging from 17M to 202M parameters demonstrate that MoE-MLA-RoPE with compression ratio r=d/2 achieves 68% KV cache memory reduction and 3.2x inference speedup while maintaining competitive perplexity (0.8% degradation). Compared to the parameters with 53.9M parameters, MoE-MLA-RoPE improves the validation loss by 6.9% over the vanilla transformers while using 42% fewer active parameters per forward pass. FLOP-matched experiments reveal even larger gains: 11.1% improvement with 3.2x inference acceleration. Automated evaluation using GPT-4 as a judge confirms quality improvements in generation, with higher scores on coherence (8.1/10), creativity (7.9/10) and grammatical correctness (8.2/10). Our results establish that architectural novelty, not parameter scaling, defines the efficiency frontier for resource-constrained language model deployment.

Hand-drawn sketches are a natural and efficient medium for capturing and conveying ideas. Despite significant advancements in controllable natural image generation, translating freehand sketches into structured, machine-readable diagrams remains a labor-intensive and predominantly manual task. The primary challenge stems from the inherent ambiguity of sketches, which lack the structural constraints and semantic precision required for automated diagram generation. To address this challenge, we introduce SketchAgent, a multi-agent system designed to automate the transformation of hand-drawn sketches into structured diagrams. SketchAgent integrates sketch recognition, symbolic reasoning, and iterative validation to produce semantically coherent and structurally accurate diagrams, significantly reducing the need for manual effort. To evaluate the effectiveness of our approach, we propose the Sketch2Diagram Benchmark, a comprehensive dataset and evaluation framework encompassing eight diverse diagram categories, such as flowcharts, directed graphs, and model architectures. The dataset comprises over 6,000 high-quality examples with token-level annotations, standardized preprocessing, and rigorous quality control. By streamlining the diagram generation process, SketchAgent holds great promise for applications in design, education, and engineering, while offering a significant step toward bridging the gap between intuitive sketching and machine-readable diagram generation. The benchmark is released at this link.

With the rapid advancement of large language models , code generation has become a key benchmark for evaluating LLM capabilities. However, existing benchmarks face two major challenges: (1) the escalating cost of constructing high-quality test suites and reference solutions, and (2) the increasing risk of data contamination, which undermines the reliability of benchmark-based evaluations. In this paper, we propose BIS, a prompt-centric evaluation framework that enables ground-truth-free prediction of LLM performance on code generation tasks. Rather than executing generated code, BIS estimates performance metrics by analyzing the prompt distribution alone. Built on importance sampling theory and implemented using Importance Weighted Autoencoders, our method reweights samples from existing annotated benchmarks to estimate performance on new, unseen benchmarks. To stabilize the estimation, we introduce weight truncation strategies and compute marginal expectations across the fitted distributions. BIS serves as a complementary tool that supports benchmark development and validation under constrained resources, offering actionable and quick feedback for prompt selection and contamination assessment. We conduct extensive experiments involving 8,000 evaluation points across 4 CodeLlama models and 9 diverse benchmarks. Our framework achieves an average absolute prediction error of 1.1% for code correctness scores, with best- and worst-case errors of 0.3% and 1.9%, respectively. It also generalizes well to other metrics, attaining average absolute errors of 2.15% for pass@1. These results demonstrate the reliability and broad applicability of BIS, which can significantly reduce the cost and effort of benchmarking LLMs in code-related tasks.

Large Language Models (LLMs) have demonstrated significant success across diverse applications. However, enforcing content restrictions remains a significant challenge due to their expansive output space. One aspect of content restriction is preventing LLMs from generating harmful content via model alignment approaches such as supervised fine-tuning (SFT). Yet, the need for content restriction may vary significantly across user groups, change rapidly over time, and not always align with general definitions of harmfulness. Applying SFT to each of these specific use cases is impractical due to the high computational, data, and storage demands. Motivated by this need, we propose a new task called \textit{Adaptive Content Restriction} (AdaCoRe), which focuses on lightweight strategies -- methods without model fine-tuning -- to prevent deployed LLMs from generating restricted terms for specific use cases. We propose the first method for AdaCoRe, named \textit{Suffix Optimization (SOP)}, which appends a short, optimized suffix to any prompt to a) prevent a target LLM from generating a set of restricted terms, while b) preserving the output quality. To evaluate AdaCoRe approaches, including our SOP, we create a new \textit{Content Restriction Benchmark} (CoReBench), which contains 400 prompts for 80 restricted terms across 8 carefully selected categories. We demonstrate the effectiveness of SOP on CoReBench, which outperforms the system-level baselines such as system suffix by 15\%, 17\%, 10\%, 9\%, and 6\% on average restriction rates for Gemma2-2B, Mistral-7B, Vicuna-7B, Llama3-8B, and Llama3.1-8B, respectively. We also demonstrate that SOP is effective on POE, an online platform hosting various commercial LLMs, highlighting its practicality in real-world scenarios.

In the age of large language models (LLMs), autonomous agents have emerged as a powerful paradigm for achieving general intelligence. These agents dynamically leverage tools, memory, and reasoning capabilities to accomplish user-defined goals. As agent systems grow in complexity, agent workflows-structured orchestration frameworks-have become central to enabling scalable, controllable, and secure AI behaviors. This survey provides a comprehensive review of agent workflow systems, spanning academic frameworks and industrial implementations. We classify existing systems along two key dimensions: functional capabilities (e.g., planning, multi-agent collaboration, external API integration) and architectural features (e.g., agent roles, orchestration flows, specification languages). By comparing over 20 representative systems, we highlight common patterns, potential technical challenges, and emerging trends. We further address concerns related to workflow optimization strategies and security. Finally, we outline open problems such as standardization and multimodal integration, offering insights for future research at the intersection of agent design, workflow infrastructure, and safe automation.