Transformer-based models primarily rely on Next Token Prediction (NTP), which predicts the next token in a sequence based on the preceding context. However, NTP's focus on single-token prediction often limits a model's ability to plan ahead or maintain long-range coherence, raising questions about how well LLMs can predict longer contexts, such as full sentences within structured documents. While NTP encourages local fluency, it provides no explicit incentive to ensure global coherence across sentence boundaries--an essential skill for reconstructive or discursive tasks. To investigate this, we evaluate three commercial LLMs (GPT -4o, Claude 3.5 Sonnet, and Gemini 2.0 Flash) on Masked Sentence Prediction (MSP) -- the task of infilling a randomly removed sentence -- from three domains: ROCStories (narrative), Recipe1M (procedural), and Wikipedia (expository). We assess both fidelity (similarity to the original sentence) and cohesiveness (fit within the surrounding context). Our key finding reveals that commercial LLMs, despite their superlative performance in other tasks, are poor at predicting masked sentences in low-structured domains, highlighting a gap in current model capabilities.

Temporal Video Grounding (TVG) aims to precisely localize video segments corresponding to natural language queries, which is a critical capability for long-form video understanding. Although existing reinforcement learning approaches encourage models to generate reasoning chains before predictions, they fail to explicitly constrain the reasoning process to ensure the quality of the final temporal predictions. To address this limitation, we propose Timestamp Anchor-constrained Reasoning for Temporal Video Grounding (TAR-TVG), a novel framework that introduces timestamp anchors within the reasoning process to enforce explicit supervision to the thought content. These anchors serve as intermediate verification points. More importantly, we require each reasoning step to produce increasingly accurate temporal estimations, thereby ensuring that the reasoning process contributes meaningfully to the final prediction. To address the challenge of low-probability anchor generation in models (e.g., Qwen2.5-VL-3B), we develop an efficient self-distillation training strategy: (1) initial GRPO training to collect 30K high-quality reasoning traces containing multiple timestamp anchors, (2) supervised fine-tuning (SFT) on distilled data, and (3) final GRPO optimization on the SFT-enhanced model. This three-stage training strategy enables robust anchor generation while maintaining reasoning quality. Experiments show that our model achieves state-of-the-art performance while producing interpretable, verifiable reasoning chains with progressively refined temporal estimations.

Sepsis, a life-threatening inflammatory response to infection, causes organ dysfunction, making early detection and optimal management critical. Previous reinforcement learning (RL) approaches to sepsis management rely primarily on structured data, such as lab results or vital signs, and on a dearth of a comprehensive understanding of the patient's condition. In this work, we propose a Multimodal Offline REinforcement learning for Clinical notes Leveraged Enhanced stAte Representation (MORE-CLEAR) framework for sepsis control in intensive care units. MORE-CLEAR employs pre-trained large-scale language models (LLMs) to facilitate the extraction of rich semantic representations from clinical notes, preserving clinical context and improving patient state representation. Gated fusion and cross-modal attention allow dynamic weight adjustment in the context of time and the effective integration of multimodal data. Extensive cross-validation using two public (MIMIC-III and MIMIC-IV) and one private dataset demonstrates that MORE-CLEAR significantly improves estimated survival rate and policy performance compared to single-modal RL approaches. To our knowledge, this is the first to leverage LLM capabilities within a multimodal offline RL for better state representation in medical applications. This approach can potentially expedite the treatment and management of sepsis by enabling reinforcement learning models to propose enhanced actions based on a more comprehensive understanding of patient conditions.

With the increase in maritime traffic and the mandatory implementation of the Automatic Identification System (AIS), the importance and diversity of maritime traffic analysis tasks based on AIS data, such as vessel trajectory prediction, anomaly detection, and collision risk assessment, is rapidly growing. However, existing approaches tend to address these tasks individually, making it difficult to holistically consider complex maritime situations. To address this limitation, we propose a novel framework, AIS-LLM, which integrates time-series AIS data with a large language model (LLM). AIS-LLM consists of a Time-Series Encoder for processing AIS sequences, an LLM-based Prompt Encoder, a Cross-Modality Alignment Module for semantic alignment between time-series data and textual prompts, and an LLM-based Multi-Task Decoder. This architecture enables the simultaneous execution of three key tasks: trajectory prediction, anomaly detection, and risk assessment of vessel collisions within a single end-to-end system. Experimental results demonstrate that AIS-LLM outperforms existing methods across individual tasks, validating its effectiveness. Furthermore, by integratively analyzing task outputs to generate situation summaries and briefings, AIS-LLM presents the potential for more intelligent and efficient maritime traffic management.

For transradial amputees, robotic prosthetic hands promise to regain the capability to perform daily living activities. To advance next-generation prosthetic hand control design, it is crucial to address current shortcomings in robustness to out of lab artifacts, and generalizability to new environments. Due to the fixed number of object to interact with in existing datasets, contrasted with the virtually infinite variety of objects encountered in the real world, current grasp models perform poorly on unseen objects, negatively affecting users' independence and quality of life. To address this: (i) we define semantic projection, the ability of a model to generalize to unseen object types and show that conventional models like YOLO, despite 80% training accuracy, drop to 15% on unseen objects. (ii) we propose Grasp-LLaVA, a Grasp Vision Language Model enabling human-like reasoning to infer the suitable grasp type estimate based on the object's physical characteristics resulting in a significant 50.2% accuracy over unseen object types compared to 36.7% accuracy of an SOTA grasp estimation model. Lastly, to bridge the performance-latency gap, we propose Hybrid Grasp Network (HGN), an edge-cloud deployment infrastructure enabling fast grasp estimation on edge and accurate cloud inference as a fail-safe, effectively expanding the latency vs. accuracy Pareto. HGN with confidence calibration (DC) enables dynamic switching between edge and cloud models, improving semantic projection accuracy by 5.6% (to 42.3%) with 3.5x speedup over the unseen object types. Over a real-world sample mix, it reaches 86% average accuracy (12.2% gain over edge-only), and 2.2x faster inference than Grasp-LLaVA alone.

While defenses against single-turn jailbreak attacks on Large Language Models (LLMs) have improved significantly, multi-turn jailbreaks remain a persistent vulnerability, often achieving success rates exceeding 70% against models optimized for single-turn protection. This work presents an empirical analysis of automated multi-turn jailbreak attacks across state-of-the-art models including GPT-4, Claude, and Gemini variants, using the StrongREJECT benchmark. Our findings challenge the perceived sophistication of multi-turn attacks: when accounting for the attacker's ability to learn from how models refuse harmful requests, multi-turn jailbreaking approaches are approximately equivalent to simply resampling single-turn attacks multiple times. Moreover, attack success is correlated among similar models, making it easier to jailbreak newly released ones. Additionally, for reasoning models, we find surprisingly that higher reasoning effort often leads to higher attack success rates. Our results have important implications for AI safety evaluation and the design of jailbreak-resistant systems. We release the source code at https://github.com/diogo-cruz/multi_turn_simpler

Aligning Large Language Models (LLMs) with diverse human values requires moving beyond a single holistic "better-than" preference criterion. While collecting fine-grained, aspect-specific preference data is more reliable and scalable, existing methods like Direct Preference Optimization (DPO) struggle with the severe noise and conflicts inherent in such aggregated datasets. In this paper, we tackle this challenge from a data-centric perspective. We first derive the Direct Multi-Preference Optimization (DMPO) objective, and uncover a key Preference Divergence (PD) term that quantifies inter-aspect preference conflicts. Instead of using this term for direct optimization, we leverage it to formulate a novel, theoretically-grounded data selection principle. Our principle advocates for selecting a subset of high-consensus data--identified by the most negative PD values--for efficient DPO training. We prove the optimality of this strategy by analyzing the loss bounds of the DMPO objective in the selection problem. To operationalize our approach, we introduce practical methods of PD term estimation and length bias mitigation, thereby proposing our PD selection method. Evaluation on the UltraFeedback dataset with three varying conflict levels shows that our simple yet effective strategy achieves over 10% relative improvement against both the standard holistic preference and a stronger oracle using aggregated preference signals, all while boosting training efficiency and obviating the need for intractable holistic preference annotating, unlocking the potential of robust LLM alignment via fine-grained preference signals.

Attribution explanation is a typical approach for explaining deep neural networks (DNNs), inferring an importance or contribution score for each input variable to the final output. In recent years, numerous attribution methods have been developed to explain DNNs. However, a persistent concern remains unresolved, i.e., whether and which attribution methods faithfully reflect the actual contribution of input variables to the decision-making process. The faithfulness issue undermines the reliability and practical utility of attribution explanations. We argue that these concerns stem from three core challenges. First, difficulties arise in comparing attribution methods due to their unstructured heterogeneity, differences in heuristics, formulations, and implementations that lack a unified organization. Second, most methods lack solid theoretical underpinnings, with their rationales remaining absent, ambiguous, or unverified. Third, empirically evaluating faithfulness is challenging without ground truth. Recent theoretical advances provide a promising way to tackle these challenges, attracting increasing attention. We summarize these developments, with emphasis on three key directions: (i) Theoretical unification, which uncovers commonalities and differences among methods, enabling systematic comparisons; (ii) Theoretical rationale, clarifying the foundations of existing methods; (iii) Theoretical evaluation, rigorously proving whether methods satisfy faithfulness principles. Beyond a comprehensive review, we provide insights into how these studies help deepen theoretical understanding, inform method selection, and inspire new attribution methods. We conclude with a discussion of promising open problems for further work.

We introduce InterChart, a diagnostic benchmark that evaluates how well vision-language models (VLMs) reason across multiple related charts, a task central to real-world applications such as scientific reporting, financial analysis, and public policy dashboards. Unlike prior benchmarks focusing on isolated, visually uniform charts, InterChart challenges models with diverse question types ranging from entity inference and trend correlation to numerical estimation and abstract multi-step reasoning grounded in 2-3 thematically or structurally related charts. We organize the benchmark into three tiers of increasing difficulty: (1) factual reasoning over individual charts, (2) integrative analysis across synthetically aligned chart sets, and (3) semantic inference over visually complex, real-world chart pairs. Our evaluation of state-of-the-art open and closed-source VLMs reveals consistent and steep accuracy declines as chart complexity increases. We find that models perform better when we decompose multi-entity charts into simpler visual units, underscoring their struggles with cross-chart integration. By exposing these systematic limitations, InterChart provides a rigorous framework for advancing multimodal reasoning in complex, multi-visual environments.

The future of poultry production depends on a paradigm shift replacing subjective, labor-intensive welfare checks with data-driven, intelligent monitoring ecosystems. Traditional welfare assessments-limited by human observation and single-sensor data-cannot fully capture the complex, multidimensional nature of laying hen welfare in modern farms. Multimodal Artificial Intelligence (AI) offers a breakthrough, integrating visual, acoustic, environmental, and physiological data streams to reveal deeper insights into avian welfare dynamics. This investigation highlights multimodal As transformative potential, showing that intermediate (feature-level) fusion strategies achieve the best balance between robustness and performance under real-world poultry conditions, and offer greater scalability than early or late fusion approaches. Key adoption barriers include sensor fragility in harsh farm environments, high deployment costs, inconsistent behavioral definitions, and limited cross-farm generalizability. To address these, we introduce two novel evaluation tools - the Domain Transfer Score (DTS) to measure model adaptability across diverse farm settings, and the Data Reliability Index (DRI) to assess sensor data quality under operational constraints. We also propose a modular, context-aware deployment framework designed for laying hen environments, enabling scalable and practical integration of multimodal sensing. This work lays the foundation for a transition from reactive, unimodal monitoring to proactive, precision-driven welfare systems that unite productivity with ethical, science based animal care.