Recent progress in Vision-Language-Action (VLA) models has enabled embodied agents to interpret multimodal instructions and perform complex tasks. However, existing VLAs are mostly confined to short-horizon, table-top manipulation, lacking the memory and reasoning capability required for long-horizon mobile manipulation, where agents must coordinate navigation and manipulation under changing spatial contexts. In this work, we present EchoVLA, a memory-aware VLA model for long-horizon mobile manipulation. EchoVLA incorporates a synergistic declarative memory inspired by the human brain, consisting of a scene memory that maintains a collection of spatial-semantic maps and an episodic memory that stores task-level experiences with multimodal contextual features. During both training and inference, the two memories are individually stored, updated, and retrieved based on current observations, task history, and instructions, and their retrieved representations are fused via coarse- and fine-grained attention to guide mobile-arm diffusion policies. To support large-scale training and evaluation, we further introduce MoMani, an automated benchmark that generates expert-level long-horizon trajectories through multimodal large language model (MLLM)-guided planning and feedback-driven refinement, supplemented with real-robot demonstrations. Experiments in simulated and real-world settings show that EchoVLA improves long-horizon performance, reaching 0.52 SR on manipulation/navigation and 0.31 on mobile manipulation, exceeding $π_{0.5}$ by +0.08 and +0.11.

This paper presents an empirical investigation into the capabilities of Large Language Models (LLMs) to perform automated Attribute-based Access Control (ABAC) policy mining. While ABAC provides fine-grained, context-aware access management, the increasing number and complexity of access policies can make their formulation and evaluation rather challenging. To address the task of synthesizing concise yet accurate policies, we evaluate the performance of some of the state-of-the-art LLMs, specifically Google Gemini (Flash and Pro) and OpenAI ChatGPT, as potential policy mining engines. An experimental framework was developed in Python to generate randomized access data parameterized by varying numbers of subjects, objects, and initial policy sets. The baseline policy sets, which govern permission decisions between subjects and objects, serve as the ground truth for comparison. Each LLM-generated policy was evaluated against the baseline policy using standard performance metrics. The results indicate that LLMs can effectively infer compact and valid ABAC policies for small-scale scenarios. However, as the system size increases, characterized by higher numbers of subjects and objects, LLM outputs exhibit declining accuracy and precision, coupled with significant increase in the size of policy generated, which is beyond the optimal size. These findings highlight both the promise and limitations of current LLM architectures for scalable policy mining in access control domains. Future work will explore hybrid approaches that combine prompt optimization with classical rule mining algorithms to improve scalability and interpretability in complex ABAC environments.

Large language models (LLMs) are increasingly adopted in clinical decision support, yet aligning them with the multifaceted reasoning pathways of real-world medicine remains a major challenge. Using more than 8,000 infertility treatment records, we systematically evaluate four alignment strategies: Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), Group Relative Policy Optimization (GRPO), and In-Context Learning (ICL) through a dual-layer framework combining automatic benchmarks with blinded doctor-in-the-loop assessments. GRPO achieves the highest algorithmic accuracy across multiple decision layers, confirming the value of reinforcement-based optimization for structured prediction tasks. However, clinicians consistently prefer the SFT model, citing clearer reasoning processes (p = 0.035) and higher therapeutic feasibility (p = 0.019). In blinded pairwise comparisons, SFT attains the highest winning rate (51.2%), outperforming both GRPO (26.2%) and even physicians' original decisions (22.7%). These results reveal an alignment paradox: algorithmic improvements do not necessarily translate into higher clinical trust, and may diverge from human-centered preferences. Our findings highlight the need for alignment strategies that prioritize clinically interpretable and practically feasible reasoning, rather than solely optimizing decision-level accuracy.

Recent advances in text-driven image editing have been significant, yet the task of accurately evaluating these edited images continues to pose a considerable challenge. Different from the assessment of text-driven image generation, text-driven image editing is characterized by simultaneously conditioning on both text and a source image. The edited images often retain an intrinsic connection to the original image, which dynamically change with the semantics of the text. However, previous methods tend to solely focus on text-image alignment or have not well aligned with human perception. In this work, we introduce the Text-driven Image Editing Benchmark suite (IE-Bench) to enhance the assessment of text-driven edited images. IE-Bench includes a database contains diverse source images, various editing prompts and the corresponding edited results from different editing methods, and nearly 4,000 samples with corresponding Mean Opinion Scores (MOS) provided by 15 human subjects. Furthermore, we introduce IE-Critic-R1, which, benefiting from Reinforcement Learning from Verifiable Rewards (RLVR), provides more comprehensive and explainable quality assessment for text-driven image editing that aligns with human perception. Extensive experiments demonstrate IE-Critic-R1's superior subjective-alignments on the text-driven image editing task compared with previous metrics. Related data and codes are available to the public.

Large Language Models (LLMs) encode safety-aligned behaviors during pretraining, but these safeguards deteriorate during task-specific fine-tuning, a phenomenon we identify as safety alignment drift. Studies demonstrate that even minimal fine-tuning can compromise safety mechanisms, with models like GPT -3.5 Turbo becoming consistently unsafe after adaptation on just 10 adversarial examples [Qi et al., 2023a]. Attempts to address this issue by modifying model behavior generally fall into two main categories, both of which suffer from inherent limitations. Behavioral unlearning methods attempt to remove undesirable knowledge or responses [Cao and Y ang, 2015, Bourtoule et al., 2021a], but often require costly retraining or risk catastrophic forgetting. Model editing approaches aim to update factual associations or local behaviors through direct parameter intervention [Meng et al., 2022, Mitchell et al., 2022], yet struggle to generalize beyond narrow scopes or isolated prompts. To solve these issues, we propose a new direction that treats safety behavior as an intrinsic property of the model's geometry and seeks to restore alignment through curvature-aware navigation of the loss landscape. Our key insight, supported by extensive empirical analysis (Section 2), is that models preserve notable structural properties in their loss landscapes with respect to harmful content after fine-tuning. Specifically, we observe high correlations in models' responses to harmful inputs before and after fine-tuning, despite substantial divergence in other functional behaviors. This suggests that safety mechanisms remain largely preserved in the parameter space, merely shifted to less dominant regions during task-specific optimization.

Testing RESTful API is increasingly important in quality assurance of cloud-native applications. Recent advances in machine learning (ML) techniques have demonstrated that various testing activities can be performed automatically by large language models (LLMs) with reasonable accuracy. This paper develops a multi-agent system called MASTEST that combines LLM-based and programmed agents to form a complete tool chain that covers the whole workflow of API test starting from generating unit and system test scenarios from API specification in the OpenAPI Swagger format, to generating of Pytest test scripts, executing test scripts to interact with web services, to analysing web service response messages to determine test correctness and calculate test coverage. The system also supports the incorporation of human testers in reviewing and correcting LLM generated test artefacts to ensure the quality of testing activities. MASTEST system is evaluated on two LLMs, GPT-4o and DeepSeek V3.1 Reasoner with five public APIs. The performances of LLMs on various testing activities are measured by a wide range of metrics, including unit and system test scenario coverage and API operation coverage for the quality of generated test scenarios, data type correctness, status code coverage and script syntax correctness for the quality of LLM generated test scripts, as well as bug detection ability and usability of LLM generated test scenarios and scripts. Experiment results demonstrated that both DeepSeek and GPT-4o achieved a high overall performance. DeepSeek excels in data type correctness and status code detection, while GPT-4o performs best in API operation coverage. For both models, LLM generated test scripts maintained 100\% syntax correctness and only required minimal manual edits for semantic correctness. These findings indicate the effectiveness and feasibility of MASTEST.

With the rapid advancement of Large Language Models (LLMs), recent studies have drawn attention to their potential for handling not only simple question-answer tasks but also more complex conversational abilities and performing human-like behavioral imitations. In particular, there is considerable interest in how accurately LLMs can reproduce real human emotions and behaviors, as well as whether such reproductions can function effectively in real-world scenarios. However, existing benchmarks focus primarily on knowledge-based assessment and thus fall short of sufficiently reflecting social interactions and strategic dialogue capabilities. To address these limitations, this work proposes a methodology to quantitatively evaluate the human emotional and behavioral imitation and strategic decision-making capabilities of LLMs by employing a Buy and Sell negotiation simulation. Specifically, we assign different personas to multiple LLMs and conduct negotiations between a Buyer and a Seller, comprehensively analyzing outcomes such as win rates, transaction prices, and SHAP values. Our experimental results show that models with higher existing benchmark scores tend to achieve better negotiation performance overall, although some models exhibit diminished performance in scenarios emphasizing emotional or social contexts. Moreover, competitive and cunning traits prove more advantageous for negotiation outcomes than altruistic and cooperative traits, suggesting that the assigned persona can lead to significant variations in negotiation strategies and results. Consequently, this study introduces a new evaluation approach for LLMs' social behavior imitation and dialogue strategies, and demonstrates how negotiation simulations can serve as a meaningful complementary metric to measure real-world interaction capabilities-an aspect often overlooked in existing benchmarks.

Diffusion Transformers have demonstrated remarkable capabilities in visual synthesis, yet they often struggle with high-level semantic reasoning and long-horizon planning. This limitation frequently leads to visual hallucinations and mis-alignments with user instructions, especially in scenarios involving complex scene understanding, human-object interactions, multi-stage actions, and in-context motion reasoning. To address these challenges, we propose Plan-X, a framework that explicitly enforces high-level semantic planning to instruct video generation process. At its core lies a Semantic Planner, a learnable multimodal language model that reasons over the user's intent from both text prompts and visual context, and autoregressively generates a sequence of text-grounded spatio-temporal semantic tokens. These semantic tokens, complementary to high-level text prompt guidance, serve as structured "semantic sketches" over time for the video diffusion model, which has its strength at synthesizing high-fidelity visual details. Plan-X effectively integrates the strength of language models in multimodal in-context reasoning and planning, together with the strength of diffusion models in photorealistic video synthesis. Extensive experiments demonstrate that our framework substantially reduces visual hallucinations and enables fine-grained, instruction-aligned video generation consistent with multimodal context.

Safety- and security-critical systems have to be thoroughly tested against their specifications. The state of practice is to have _natural language_ specifications, from which test cases are derived manually - a process that is slow, error-prone, and difficult to scale. _Formal_ specifications, on the other hand, are well-suited for automated test generation, but are tedious to write and maintain. In this work, we propose a two-stage pipeline that uses large language models (LLMs) to bridge the gap: First, we extract _protocol elements_ from natural-language specifications; second, leveraging a protocol implementation, we synthesize and refine a formal _protocol specification_ from these elements, which we can then use to massively test further implementations. We see this two-stage approach to be superior to end-to-end LLM-based test generation, as 1. it produces an _inspectable specification_ that preserves traceability to the original text; 2. the generation of actual test cases _no longer requires an LLM_; 3. the resulting formal specs are _human-readable_, and can be reviewed, version-controlled, and incrementally refined; and 4. over time, we can build a _corpus_ of natural-language-to-formal-specification mappings that can be used to further train and refine LLMs for more automatic translations. Our prototype, AUTOSPEC, successfully demonstrated the feasibility of our approach on five widely used _internet protocols_ (SMTP, POP3, IMAP, FTP, and ManageSieve) by applying its methods on their _RFC specifications_ written in natural-language, and the recent _I/O grammar_ formalism for protocol specification and fuzzing. In its evaluation, AUTOSPEC recovers on average 92.8% of client and 80.2% of server message types, and achieves 81.5% message acceptance across diverse, real-world systems.

State-space models (SSMs), particularly Mamba, have become powerful architectures for sequence modeling, yet their internal dynamics remain poorly understood compared to attention-based models. We introduce and validate the Influence Score, a controllability-based metric derived from the discretized state-space parameters of Mamba and computed through a backward recurrence analogous to system observability. The score quantifies how strongly a token at position k affects all later states and outputs. We evaluate this measure across three Mamba variants: mamba-130m, mamba-2.8b, and mamba-2.8b-slimpj, using six experiments that test its sensitivity to temperature, prompt complexity, token type, layer depth, token position, and input perturbations. The results show three main insights: (1) the Influence Score increases with model size and training data, reflecting model capacity; (2) Mamba exhibits consistent architectural patterns, including recency bias and concentrated influence in mid-to-late layers; and (3) emergent behaviors appear only at scale, with mamba-2.8b-slimpj uniquely prioritizing content words and reducing internal influence in the presence of noise. These findings establish the Influence Score as a practical diagnostic tool for interpreting and comparing SSM-based language models.