In high-stakes scenarios—such as self-harm, legal, or medical queries—LLMs must be both trustworthy and helpful. However, these goals often conflict. We propose priority alignment, a new alignment paradigm that enforces a strict “trustworthy-before-helpful” ordering: optimization of helpfulness is conditioned on first meeting trustworthy thresholds (e.g., harmlessness or honesty). To realize this, we introduce Self-Priority Alignment (SPA)—a fully unsupervised framework that generates diverse responses, self-evaluates them and refines them by the model itself, and applies dual-criterion denoising to remove inconsistency and control variance. From this, SPA constructs lexicographically ordered preference pairs and fine-tunes the model using an uncertainty-weighted alignment loss that emphasizes high-confidence, high-gap decisions. Experiments across multiple benchmarks show that SPA improves helpfulness without compromising safety, outperforming strong baselines while preserving general capabilities. Our results demonstrate that SPA provides a scalable and interpretable alignment strategy for critical LLM applications.

Tianyu Yang, Lisen Dai, Xiangqi Wang, Minhao Cheng, Yapeng Tian, Xiangliang Zhang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Vision Language Models (VLMs) have demonstrated remarkable success in downstream applications, yet they often exhibit biases, raising ethical concerns. While previous efforts have aimed to evaluate and improve the moral reasoning capabilities of VLMs, existing approaches are limited by simplified, unimodal settings or overly static visual scenarios. We propose a novel multi-image-based dataset pipeline MIST (Moral Inference through Storytelling with Text and Images) designed to assess moral reasoning in complex, dynamic scenarios to address these limitations. To ensure better alignment between these modalities, we introduce the concept of ''text-image flow,'' which seamlessly integrates visual and textual information across complex scenarios. Using this dataset, we evaluate seven widely used VLMs, offering critical insights into their performance in moral reasoning tasks.

LLMs often need effective configurations, like temperature and reasoning steps, to handle tasks requiring sophisticated reasoning and problem-solving, ranging from joke generation to mathematical reasoning. Existing prompting approaches usually adopt general-purpose, fixed configurations that work “well enough” across tasks but seldom achieve task-specific optimality. To address this gap, we introduce AdaReasoner, an LLM-agnostic plugin designed for any LLM to automate adaptive reasoning configurations for tasks requiring different types of thinking. AdaReasoner is trained using a reinforcement learning (RL) framework, combining a factorized action space with a targeted exploration strategy, along with a pretrained reward model to optimize the policy model for reasoning configurations with only a few-shot guide. AdaReasoner is backed by theoretical guarantees and experiments of fast convergence and a sublinear policy gap. Across six different LLMs and a variety of reasoning tasks, it consistently outperforms standard baselines, preserves out-of-distribution robustness, and yield gains on knowledge-intensive tasks through tailored prompts.

Agent systems based on large language models (LLMs) have shown great potential in complex reasoning tasks, but building efficient and generalizable workflows remains a major challenge. Most existing approaches rely on manually designed processes, which limits their adaptability across different tasks. While a few methods attempt automated workflow generation, they are often tied to specific datasets or query types and make limited use of intermediate feedback, reducing system robustness and reasoning depth. Moreover, their operations are typically predefined and inflexible. To address these limitations, we propose **DyFlow**, a dynamic workflow generation framework that adaptively constructs and adjusts reasoning procedures based on task requirements and real-time intermediate feedback, thereby enhancing cross-task generalization. DyFlow consists of two core components: a designer and an executor. The designer decomposes complex problems into a sequence of sub-goals defined by high-level objectives and dynamically plans the next steps based on intermediate outputs and feedback. These plans are then carried out by the executor, which executes each operation using dynamic operators with context-aware parameterization, enabling flexible and semantically grounded reasoning. We systematically evaluate DyFlow across diverse domains, including social reasoning, biomedical tasks, mathematical problem solving, and code generation. Results demonstrate that DyFlow significantly outperforms existing baselines, achieving substantial Pass@k improvements and exhibiting robust generalization across diverse domains.

Wildlife management is increasingly reliant on data-driven insights to address the impacts of climate change on species and ecosystems. However, the complexity of accessing and querying large, multimodal datasets often limits the ability of non-technical users, such as wildlife managers and conservationists, to make informed decisions. To address this challenge, we present WildlifeLookup, a public accessible, intelligent chatbot designed to facilitate natural language interaction with a novel knowledge graph (KN-Wildlife) that houses critical wildlife and environmental data. WildlifeLookup simplifies access to species distributions, habitat interactions, and climate-related events by converting user queries into precise graph queries, reducing the technical barriers for end users. The chatbot WildlifeLookup is available at https://oknbot.ngrok.dev/

Audio is an important medium in people’s daily life, secret information can be embedded into audio for covert communication. However, traditional audio information hiding techniques cannot achieve large hiding capacity and good imperceptibility at the same time, and rely on complex encryption, which limits their applicability in resource-constrained Internet of Things (IoT) environments. In this article, we propose a new audio information hiding method, named AdvAudio, which can achieve large high capacity, as well as good imperceptibility, without reliance on cryptographic encryption. Specifically, AdvAudio leverages adversarial example technique to train a well-designed perturbation for cover audio and the secret information can only be extracted by the private automatic speech recognition (ASR) model. To achieve this, we implement two adversarial example algorithms tailored for both online transmission and physical-world transmission scenarios. In particular, our embedding algorithm dynamically adjusts the addition of simulated environmental noise depending on whether the audio is intended to propagate in the physical world. The iterative optimization process is guided by targeted adversarial attack objectives, ensuring that the private ASR model decodes the embedded secret information accurately. Taking DeepSpeech as the private model, we implement a prototype of AdvAudio, which achieves a high embedding capacity of 383.8 bps with excellent imperceptibility, yielding a Perceptual Evaluation of Speech Quality (PESQ) score of 2.351. Furthermore, it offers robust security, achieving a 100% defense success rate against both internal and external attacks. In the physical world, AdvAudio still maintains effectiveness across six different types of noise and retaining 82% accuracy even under sudden loud noises. Additionally, the secret information can only be extracted in the target environment, with a success rate of 26%, and 0% in non-target environments. In the future, we aim at enhancing the steganalysis resistance of AdvAudio and explore its potential applications in various environments or with alternative ASR models.

The data privacy and security of Internet of Things (IoT) applications are progressively crucial and cryptography is frequently used to ensure security. Cryptographic hardware implementation is commonly adopted for high throughput and low-computational resources. The crypto circuits have to be strictly tested to guarantee the correctness of data. Scan-based design-for-testability (DFT) widely employed in the chip industry improves the controllability and observability of circuits. However, it facilitates illegal users to steal the internal data for cracking cryptographic keys. Many researchers have recently suggested effective defense technologies against scan-based attacks, but each technology has its own negative aspects. In this article, we propose an effective DFT authorization framework (DAF) based on obfuscation mechanisms. The authentication module is embedded to verify users’ keys, and it can empower each chip distinctive key to minimize the loss of key divulgence. If the test authorization key is accurate, the typical scan operation can be carried out. Conversely, if the key is inaccurate, the inserted obfuscation module comes into play to scramble the scan data. Additionally, a random number generation circuit is used to increase the uncertainty of data obfuscation, effectively preventing attackers from inferring sensitive information. Simulation results show that this design has a low overhead, and theoretical analysis demonstrates that the design has high security with no impact on the testability of the chip.

Cryptography chips are often used in some applications, such as smart grids and Internet of Things (IoT) to ensure their security. Cryptographic chips must be strictly tested to guarantee the correctness of the encryption and decryption. Scan-based design-for-testability (DFT) provides high test quality. However, it can also be misused to steal the cipher key of cryptographic chips by hackers. In this article, we present a new scan design methodology that can resist scan-based side-channel attacks by the dynamical obfuscation of scan input data and scan output data. The scan test is managed by a test password, which consists of load password and scan password. When the chip enters into the test mode, it is required to apply the test password via some external input ports. Once the correct load password is delivered, the scan password can be loaded into a special shift register. If the scan password is also correct, the chip testing can proceed normally. In case the load password or the scan password is wrong, the data in scan chains cannot be propagated correctly. Specifically, some elusory bits are sneaked into scan chains dynamically. The advantage of the proposed method is that it has no negative impact on design performance and test flow when powerfully protecting cryptographic chips. The area penalty is also acceptably low compared with other schemes.

With the rapid advancements of the artificial intelligence, machine learning, especially neural networks, have shown huge superiority over humans in image recognition, autonomous vehicles and medical diagnosis. However, its opacity and inexplicability provide many chances for malicious attackers. Recent researches have shown that neural networks are vulnerable to adversarial example (AE) attacks. In the testing stage, it fools the model by adding subtle perturbations to the original sample to misclassify the input, which poses a serious threat to safety-critical areas such as autonomous driving. In order to mitigate this threat, this paper proposes a hardware-assisted randomization method against AEs, where an approximate computing technique in hardware, voltage over-scaling (VOS), is used to randomize the training set of the model, then the processed data are used to generate multiple neural network models, finally multiple redundant models are used for the integrated classification and detection of the AEs. Various AE attacks on the proposed defense are evaluated to prove its effectiveness.