Large Language Models (LLMs) have become powerful, but hallucinations remain a vital obstacle to their trustworthy use. While previous works improved the capability of hallucination detection by measuring uncertainty, they all lack the ability to explain the provenance behind why hallucinations occur, i.e., which part of the inputs tends to trigger hallucinations. Recent works on the prompt attack indicate that uncertainty exists in semantic propagation, where attention mechanisms gradually fuse local token information into high-level semantics across layers. Meanwhile, uncertainty also emerges in language generation, due to its probability-based selection of high-level semantics for sampled generations. Based on that, we propose RePPL to recalibrate uncertainty measurement by these two aspects, which dispatches explainable uncertainty scores to each token and aggregates in Perplexity-style Log-Average form as total score. Experiments show that our method achieves the best comprehensive detection performance across various QA datasets on advanced models (average AUC of 0.833), and our method is capable of producing token-level uncertainty scores as explanations for the hallucination. Leveraging these scores, we preliminarily find the chaotic pattern of hallucination and showcase its promising usage.

Large Language Models face security threats from jailbreak attacks. Existing research has predominantly focused on prompt-level attacks while largely ignoring the underexplored attack surface of user-controlled response prefilling. This functionality allows an attacker to dictate the beginning of a model's output, thereby shifting the attack paradigm from persuasion to direct state manipulation.In this paper, we present a systematic black-box security analysis of prefill-level jailbreak attacks. We categorize these new attacks and evaluate their effectiveness across fourteen language models. Our experiments show that prefill-level attacks achieve high success rates, with adaptive methods exceeding 99% on several models. Token-level probability analysis reveals that these attacks work through initial-state manipulation by changing the first-token probability from refusal to compliance.Furthermore, we show that prefill-level jailbreak can act as effective enhancers, increasing the success of existing prompt-level attacks by 10 to 15 percentage points. Our evaluation of several defense strategies indicates that conventional content filters offer limited protection. We find that a detection method focusing on the manipulative relationship between the prompt and the prefill is more effective. Our findings reveal a gap in current LLM safety alignment and highlight the need to address the prefill attack surface in future safety training.

Machine Learning (ML) has been integrated into various software and systems. Two main components are essential for training an ML model: the training data and the ML algorithm. Given the critical role of data in ML system development, it has become increasingly important to assess the quality of data attributes and ensure that the data meets specific requirements before its utilization. This work proposes an approach to guide non-experts in identifying data requirements for ML systems using goal modeling. In this approach, we first develop the Data Requirement Goal Model (DRGM) by surveying the white literature to identify and categorize the issues and challenges faced by data scientists and requirement engineers working on ML-related projects. An initial DRGM was built to accommodate common tasks that would generalize across projects. Then, based on insights from both white and gray literature, a customization mechanism is built to help adjust the tasks, KPIs, and goals' importance of different elements within the DRGM. The generated model can aid its users in evaluating different datasets using GRL evaluation strategies. We then validate the approach through two illustrative examples based on real-world projects. The results from the illustrative examples demonstrate that the data requirements identified by the proposed approach align with the requirements of real-world projects, demonstrating the practicality and effectiveness of the proposed framework. The proposed dataset selection customization mechanism and the proposed DRGM are helpful in guiding non-experts in identifying the data requirements for machine learning systems tailored to a specific ML problem. This approach also aids in evaluating different dataset alternatives to choose the optimum dataset for the problem. For future work, we recommend implementing tool support to generate the DRGM based on a chatbot interface.

Multimodal in-context learning (ICL) equips Large Vision-language Models (LVLMs) with the ability to adapt to new tasks via multiple user-provided demonstrations, without requiring any model parameter updates. However, its effectiveness is constrained by the token-intensive nature of multimodal inputs and the complexity of cross-modal few-shot reasoning, which together hinder LVLMs from extracting useful patterns from demonstrations. To address these challenges, we propose \textbf{M$^2$IV}, a novel representation engineering approach that replaces explicit token-level demonstrations with a set of learnable Multimodal In-context Vectors directly injected into the residual streams of LVLMs. By analyzing the distinct roles of multi-head attention (MHA) and multi-layer perceptrons (MLP) in the ICL process, we design a training strategy that enables M$^2$IV to perform fine-grained semantic distillation and robust cross-modal representation learning. M$^2$IV not only improves performance across diverse tasks and LVLMs but also significantly reduces token overhead, enabling graceful scaling to many-shot scenarios. To further enhance usability, we introduce \textbf{VLibrary}, a repository that stores trained M$^2$IVs for flexible retrieval and injection. With VLibrary, users can steer pre-trained LVLMs in a customized manner that meets diverse requirements. Extensive experiments demonstrate that M$^2$IV consistently outperforms vanilla ICL and prior representation engineering baselines, achieving an average accuracy gain of 3.74\% with substantial improvements in overall efficiency.

Temporal context is essential for robotic manipulation because such tasks are inherently non-Markovian, yet mainstream VLA models typically overlook it and struggle with long-horizon, temporally dependent tasks. Cognitive science suggests that humans rely on working memory to buffer short-lived representations for immediate control, while the hippocampal system preserves verbatim episodic details and semantic gist of past experience for long-term memory. Inspired by these mechanisms, we propose MemoryVLA, a Cognition-Memory-Action framework for long-horizon robotic manipulation. A pretrained VLM encodes the observation into perceptual and cognitive tokens that form working memory, while a Perceptual-Cognitive Memory Bank stores low-level details and high-level semantics consolidated from it. Working memory retrieves decision-relevant entries from the bank, adaptively fuses them with current tokens, and updates the bank by merging redundancies. Using these tokens, a memory-conditioned diffusion action expert yields temporally aware action sequences. We evaluate MemoryVLA on 150+ simulation and real-world tasks across three robots. On SimplerEnv-Bridge, Fractal, and LIBERO-5 suites, it achieves 71.9%, 72.7%, and 96.5% success rates, respectively, all outperforming state-of-the-art baselines CogACT and pi-0, with a notable +14.6 gain on Bridge. On 12 real-world tasks spanning general skills and long-horizon temporal dependencies, MemoryVLA achieves 84.0% success rate, with long-horizon tasks showing a +26 improvement over state-of-the-art baseline. Project Page: https://shihao1895.github.io/MemoryVLA

The success of deep learning-based speaker verification systems is largely attributed to access to large-scale and diverse speaker identity data. However, collecting data from more identities is expensive, challenging, and often limited by privacy concerns. To address this limitation, we propose INSIDE (Interpolating Speaker Identities in Embedding Space), a novel data expansion method that synthesizes new speaker identities by interpolating between existing speaker embeddings. Specifically, we select pairs of nearby speaker embeddings from a pretrained speaker embedding space and compute intermediate embeddings using spherical linear interpolation. These interpolated embeddings are then fed to a text-to-speech system to generate corresponding speech waveforms. The resulting data is combined with the original dataset to train downstream models. Experiments show that models trained with INSIDE-expanded data outperform those trained only on real data, achieving 3.06\% to 5.24\% relative improvements. While INSIDE is primarily designed for speaker verification, we also validate its effectiveness on gender classification, where it yields a 13.44\% relative improvement. Moreover, INSIDE is compatible with other augmentation techniques and can serve as a flexible, scalable addition to existing training pipelines.

In safety-critical deep learning applications, robustness measures the ability of neural models that handle imperceptible perturbations in input data, which may lead to potential safety hazards. Existing pre-deployment robustness assessment methods typically suffer from significant trade-offs between computational cost and measurement precision, limiting their practical utility. To address these limitations, this paper conducts a comprehensive comparative analysis of existing robustness definitions and associated assessment methodologies. We propose tower robustness to evaluate robustness, which is a novel, practical metric based on hypothesis testing to quantitatively evaluate probabilistic robustness, enabling more rigorous and efficient pre-deployment assessments. Our extensive comparative evaluation illustrates the advantages and applicability of our proposed approach, thereby advancing the systematic understanding and enhancement of model robustness in safety-critical deep learning applications.

Despite the growing use of large language models (LLMs) in clinical dialogue systems, existing evaluations focus on task completion or fluency, offering little insight into the behavioral and risk management requirements essential for safety-critical systems. This paper presents MATRIX (Multi-Agent simulaTion fRamework for safe Interactions and conteXtual clinical conversational evaluation), a structured, extensible framework for safety-oriented evaluation of clinical dialogue agents. MATRIX integrates three components: (1) a safety-aligned taxonomy of clinical scenarios, expected system behaviors and failure modes derived through structured safety engineering methods; (2) BehvJudge, an LLM-based evaluator for detecting safety-relevant dialogue failures, validated against expert clinician annotations; and (3) PatBot, a simulated patient agent capable of producing diverse, scenario-conditioned responses, evaluated for realism and behavioral fidelity with human factors expertise, and a patient-preference study. Across three experiments, we show that MATRIX enables systematic, scalable safety evaluation. BehvJudge with Gemini 2.5-Pro achieves expert-level hazard detection (F1 0.96, sensitivity 0.999), outperforming clinicians in a blinded assessment of 240 dialogues. We also conducted one of the first realism analyses of LLM-based patient simulation, showing that PatBot reliably simulates realistic patient behavior in quantitative and qualitative evaluations. Using MATRIX, we demonstrate its effectiveness in benchmarking five LLM agents across 2,100 simulated dialogues spanning 14 hazard scenarios and 10 clinical domains. MATRIX is the first framework to unify structured safety engineering with scalable, validated conversational AI evaluation, enabling regulator-aligned safety auditing. We release all evaluation tools, prompts, structured scenarios, and datasets.

Contemporary artificial intelligence (AI) development largely centers on rational decision-making, valued for its measurability and suitability for objective evaluation. Y et in real-world contexts, an intelligent agent's decisions are shaped not only by logic but also by deeper influences such as beliefs, values, and preferences. The diversity of human decision-making styles emerges from these differences, highlighting that "style" is an essential but often overlooked dimension of intelligence. This dissertation introduces playstyle as an alternative lens for observing and analyzing the decision-making behavior of intelligent agents, and examines its foundational meaning and historical context from a philosophical perspective. By analyzing how beliefs and values drive intentions and actions, we construct a two-tier framework for style formation: the external interaction loop with the environment and the internal cognitive loop of deliberation. On this basis, we formalize style-related characteristics and propose measurable indicators such as style capacity, style popularity, and evolutionary dynamics. The study focuses on three core research directions: (1) Defining and measuring playstyle, proposing a general playstyle metric based on discretized state spaces, and extending it to quantify strategic diversity and competitive balance; (2) Expressing and generating playstyle, exploring how reinforcement learning and imitation learning can be used to train agents exhibiting specific stylistic tendencies, and introducing a novel approach for human-like style learning and modeling; and (3) Practical applications, analyzing the potential of these techniques in domains such as game design and interactive entertainment. Finally, the dissertation outlines future extensions, including the role of style as a core element in building artificial general intelligence (AGI). By investigating stylistic variation, we aim to rethink autonomy, value expression, and even offer a tangible perspective on the ultimate i philosophical question: What is the soul?

Implicit Neural Point Cloud (INPC) is a recent hybrid representation that combines the expressiveness of neural fields with the efficiency of point-based rendering, achieving state-of-the-art image quality in novel view synthesis. However, as with other high-quality approaches that query neural networks during rendering, the practical usability of INPC is limited by comparatively slow rendering. In this work, we present a collection of optimizations that significantly improve both the training and inference performance of INPC without sacrificing visual fidelity. The most significant modifications are an improved rasterizer implementation, more effective sampling techniques, and the incorporation of pre-training for the convolutional neural network used for hole-filling. Furthermore, we demonstrate that points can be modeled as small Gaussians during inference to further improve quality in extrapolated, e.g., close-up views of the scene. We design our implementations to be broadly applicable beyond INPC and systematically evaluate each modification in a series of experiments. Our optimized INPC pipeline achieves up to 25% faster training, 2x faster rendering, and 20% reduced VRAM usage paired with slight image quality improvements.