Large Language Models (LLMs) with safe-alignment training are powerful instruments with robust language comprehension capabilities. These models typically undergo meticulous alignment procedures involving human feedback to ensure the acceptance of safe inputs while rejecting harmful or unsafe ones. However, despite their massive scale and alignment efforts, LLMs remain vulnerable to jailbreak attacks, where malicious users manipulate the model to produce harmful outputs that it was explicitly trained to avoid. In this study, we find that the safety mechanisms in LLMs are predominantly embedded in the middle-to-late layers. Building on this insight, we introduce a novel white-box jailbreak method, SABER (Safety Alignment Bypass via Extra Residuals), which connects two intermediate layers $s$ and $e$ such that $s < e$, through a residual connection. Our approach achieves a 51% improvement over the best-performing baseline on the HarmBench test set. Furthermore, SABER induces only a marginal shift in perplexity when evaluated on the HarmBench validation set. The source code is publicly available at https://github.com/PalGitts/SABER.

A key requirement for generalist robots is compositional generalization - the ability to combine atomic skills to solve complex, long-horizon tasks. While prior work has primarily focused on synthesizing a planner that sequences pre-learned skills, robust execution of the individual skills themselves remains challenging, as visuomotor policies often fail under distribution shifts induced by scene composition. To address this, we introduce a scene graph-based representation that focuses on task-relevant objects and relations, thereby mitigating sensitivity to irrelevant variation. Building on this idea, we develop a scene-graph skill learning framework that integrates graph neural networks with diffusion-based imitation learning, and further combine "focused" scene-graph skills with a vision-language model (VLM) based task planner. Experiments in both simulation and real-world manipulation tasks demonstrate substantially higher success rates than state-of-the-art baselines, highlighting improved robustness and compositional generalization in long-horizon tasks.

Spoken dialogue systems increasingly employ large language models (LLMs) to leverage their advanced reasoning capabilities. However, direct application of LLMs in spoken communication often yield suboptimal results due to mismatches between optimal textual and verbal delivery. While existing approaches adapt LLMs to produce speech-friendly outputs, their impact on reasoning performance remains underexplored. In this work, we propose Think-Verbalize-Speak, a framework that decouples reasoning from spoken delivery to preserve the full reasoning capacity of LLMs. Central to our method is verbalizing, an intermediate step that translates thoughts into natural, speech-ready text. We also introduce ReVerT, a latency-efficient verbalizer based on incremental and asynchronous summarization. Experiments across multiple benchmarks show that our method enhances speech naturalness and conciseness with minimal impact on reasoning. The project page with the dataset and the source code is available at https://yhytoto12.github.io/TVS-ReVerT

Abstract--Recent years have witnessed a growing interest in automating labor-intensive and complex activities, i.e., those consisting of multiple atomic tasks, by deploying robots in dynamic and unpredictable environments such as industrial and agricultural settings. A key characteristic of these contexts is that activities are not predefined: while they involve a limited set of possible tasks, their combinations may vary depending on the situation. Moreover, despite recent advances in robotics, the ability for humans to monitor the progress of high-level activities - in terms of past, present, and future actions - remains fundamental to ensure the correct execution of safety-critical processes. In this paper, we introduce a general architecture that integrates Large Language Models (LLMs) with automated planning, enabling humans to specify high-level activities (also referred to as processes) using natural language, and to monitor their execution by querying a robot. We also present an implementation of this architecture using state-of-the-art components and quantitatively evaluate the approach in a real-world precision agriculture scenario. I. INTRODUCTION In recent years, there has been a significant increase in the interest and demand for automating complex and labor-intensive activities through the deployment of robotic systems. These activities, often encountered in industrial and agricultural domains, are typically composed of multiple, smaller atomic subtasksthat must be coordinated to achieve a larger goal. What makes these environments particularly challenging is their dynamic and unpredictable nature: the specific sequence and combination of tasks required can change based on real-time conditions, external events, or evolving objectives. Importantly, while the range of possible jobs is usually limited and known in advance, the structure and flow of the overall activity are not fixed and therefore must be adapted on the fly. In such contexts, human operators must maintain a clear understanding of ongoing high-level activities, both to ensure correctness and to make timely decisions. This includes being aware of what the system has done (past), what it is currently doing (present), and what it intends to do next (future). However, this kind of situational awareness is difficult to maintain in the absence of mechanisms for querying the system in a human-friendly way.

Bias-only fine-tuning has the potential for unprecedented parameter efficiency. However, the link between fine-tuning different bias terms (i.e., bias terms in the query, key, or value projections) and downstream performance remains unclear. The existing approaches, e.g., based on the magnitude of bias change or empirical Fisher information, provide limited guidance for selecting the particular bias term for effective fine-tuning. In this paper, we propose an approach for selecting the bias term to be fine-tuned, forming the foundation of our bias-efficient fine-tuning (BEFT). We extensively evaluate our bias-efficient approach against other bias-selection approaches, across a wide range of large language models (LLMs) spanning encoder-only and decoder-only architectures from 110M to 6.7B parameters. Our results demonstrate the effectiveness and superiority of our bias-efficient approach on diverse downstream tasks, including classification, multiple-choice, and generation tasks.

Background: Large language models (LLMs) show promise in medicine, but their deployment in hospitals is limited by restricted access to electronic health record (EHR) systems. The Model Context Protocol (MCP) enables integration between LLMs and external tools. Objective: To evaluate whether an LLM connected to an EHR database via MCP can autonomously retrieve clinically relevant information in a real hospital setting. Methods: We developed EHR-MCP, a framework of custom MCP tools integrated with the hospital EHR database, and used GPT-4.1 through a LangGraph ReAct agent to interact with it. Six tasks were tested, derived from use cases of the infection control team (ICT). Eight patients discussed at ICT conferences were retrospectively analyzed. Agreement with physician-generated gold standards was measured. Results: The LLM consistently selected and executed the correct MCP tools. Except for two tasks, all tasks achieved near-perfect accuracy. Performance was lower in the complex task requiring time-dependent calculations. Most errors arose from incorrect arguments or misinterpretation of tool results. Responses from EHR-MCP were reliable, though long and repetitive data risked exceeding the context window. Conclusions: LLMs can retrieve clinical data from an EHR via MCP tools in a real hospital setting, achieving near-perfect performance in simple tasks while highlighting challenges in complex ones. EHR-MCP provides an infrastructure for secure, consistent data access and may serve as a foundation for hospital AI agents. Future work should extend beyond retrieval to reasoning, generation, and clinical impact assessment, paving the way for effective integration of generative AI into clinical practice.

Turning garments right-side out is a challenging manipulation task: it is highly dynamic, entails rapid contact changes, and is subject to severe visual occlusion. We introduce Right-Side-Out, a zero-shot sim-to-real framework that effectively solves this challenge by exploiting task structures. We decompose the task into Drag/Fling to create and stabilize an access opening, followed by Insert&Pull to invert the garment. Each step uses a depth-inferred, keypoint-parameterized bimanual primitive that sharply reduces the action space while preserving robustness. Efficient data generation is enabled by our custom-built, high-fidelity, GPU-parallel Material Point Method (MPM) simulator that models thin-shell deformation and provides robust and efficient contact handling for batched rollouts. Built on the simulator, our fully automated pipeline scales data generation by randomizing garment geometry, material parameters, and viewpoints, producing depth, masks, and per-primitive keypoint labels without any human annotations. With a single depth camera, policies trained entirely in simulation deploy zero-shot on real hardware, achieving up to 81.3% success rate. By employing task decomposition and high fidelity simulation, our framework enables tackling highly dynamic, severely occluded tasks without laborious human demonstrations.

Automated Essay Scoring (AES) systems now reach near human agreement on some public benchmarks, yet real-world adoption, especially in high-stakes examinations, remains limited. A principal obstacle is that most models output a single score without any accompanying measure of confidence or explanation. We address this gap with conformal prediction, a distribution-free wrapper that equips any classifier with set-valued outputs and formal coverage guarantees. Two open-source large language models (Llama-3 8B and Qwen-2.5 3B) are fine-tuned on three diverse corpora (ASAP, TOEFL11, Cambridge-FCE) and calibrated at a 90 percent risk level. Reliability is assessed with UAcc, an uncertainty-aware accuracy that rewards models for being both correct and concise. To our knowledge, this is the first work to combine conformal prediction and UAcc for essay scoring. The calibrated models consistently meet the coverage target while keeping prediction sets compact, indicating that open-source, mid-sized LLMs can already support teacher-in-the-loop AES; we discuss scaling and broader user studies as future work.

While reinforcement learning from scratch has shown impressive results in solving sequential decision-making tasks with efficient simulators, real-world applications with expensive interactions require more sample-efficient agents. Foundation models (FMs) are natural candidates to improve sample efficiency as they possess broad knowledge and reasoning capabilities, but it is yet unclear how to effectively integrate them into the reinforcement learning framework. In this paper, we anticipate and, most importantly, evaluate two promising strategies. First, we consider the use of foundation world models (FWMs) that exploit the prior knowledge of FMs to enable training and evaluating agents with simulated interactions. Second, we consider the use of foundation agents (FAs) that exploit the reasoning capabilities of FMs for decision-making. We evaluate both approaches empirically in a family of grid-world environments that are suitable for the current generation of large language models (LLMs). Our results suggest that improvements in LLMs already translate into better FWMs and FAs; that FAs based on current LLMs can already provide excellent policies for sufficiently simple environments; and that the coupling of FWMs and reinforcement learning agents is highly promising for more complex settings with partial observability and stochastic elements.

ABSTRACT Recent retrieval-augmented image captioning methods incorporate external knowledge to compensate for the limitations in comprehending complex scenes. However, current approaches face challenges in relation modeling: (1) the representation of semantic prompts is too coarse-grained to capture fine-grained relationships; (2) these methods lack explicit modeling of image objects and their semantic relationships. To address these limitations, we propose RACap, a relation-aware retrieval-augmented model for image captioning, which not only mines structured relation semantics from retrieval captions, but also identifies heterogeneous objects from the image. RACap effectively retrieves structured relation features that contain heterogeneous visual information to enhance the semantic consistency and relational expressiveness. Experimental results show that RACap, with only 10.8M trainable parameters, achieves superior performance compared to previous lightweight captioning models. Index T erms-- Image captioning, Lightweight vision-language models, Retrieval-augmented generation, Cross-Modal 1. INTRODUCTION Approaches based on large language models (LLMs) have shown promising results in image captioning tasks [1, 2].