Automatically structuring posology instructions is essential for improving medication safety and enabling clinical decision support. In French prescriptions, these instructions are often ambiguous, irregular, or colloquial, limiting the effectiveness of classic ML pipelines. We explore the use of Large Language Models (LLMs) to convert free-text posologies into structured formats, comparing prompt-based methods and fine-tuning against a "pre-LLM" system based on Named Entity Recognition and Linking (NERL). Our results show that while prompting improves performance, only fine-tuned LLMs match the accuracy of the baseline. Through error analysis, we observe complementary strengths: NERL offers structural precision, while LLMs better handle semantic nuances. Based on this, we propose a hybrid pipeline that routes low-confidence cases from NERL (<0.8) to the LLM, selecting outputs based on confidence scores. This strategy achieves 91% structuration accuracy while minimizing latency and compute. Our results show that this hybrid approach improves structuration accuracy while limiting computational cost, offering a scalable solution for real-world clinical use.

Despite the remarkable multimodal capabilities of Large Vision-Language Models (LVLMs), discrepancies often occur between visual inputs and textual outputs--a phenomenon we term visual hallucination. This critical reliability gap poses substantial risks in safety-critical Artificial Intelligence (AI) applications, necessitating a comprehensive evaluation benchmark and effective detection methods. Firstly, we observe that existing visual-centric hallucination benchmarks mainly assess LVLMs from a perception perspective, overlooking hallucinations arising from advanced reasoning capabilities. We develop the Perception-Reasoning Evaluation Hallucination (PRE-HAL) dataset, which enables the systematic evaluation of both perception and reasoning capabilities of LVLMs across multiple visual semantics, such as instances, scenes, and relations. Comprehensive evaluation with this new benchmark exposed more visual vulnerabilities, particularly in the more challenging task of relation reasoning. To address this issue, we propose, to the best of our knowledge, the first Dempster-Shafer theory (DST)-based visual hallucination detection method for LVLMs through uncertainty estimation. This method aims to efficiently capture the degree of conflict in high-level features at the model inference phase. Specifically, our approach employs simple mass functions to mitigate the computational complexity of evidence combination on power sets. We conduct an extensive evaluation of state-of-the-art LVLMs, LLaVA-v1.5, mPLUG-Owl2 and mPLUG-Owl3, with the new PRE-HAL benchmark. Experimental results indicate that our method outperforms five baseline uncertainty metrics, achieving average AUROC improvements of 4%, 10%, and 7% across three LVLMs. Our code is available at https://github.com/HT86159/Evidential-Conflict.

This paper presents the approach of our team called heiDS for the ArchEHR-QA 2025 shared task. A pipeline using a retrieval augmented generation (RAG) framework is designed to generate answers that are attributed to clinical evidence from the electronic health records (EHRs) of patients in response to patient-specific questions. We explored various components of a RAG framework, focusing on ranked list truncation (RLT) retrieval strategies and attribution approaches. Instead of using a fixed top-k RLT retrieval strategy, we employ a query-dependent-k retrieval strategy, including the existing surprise and autocut methods and two new methods proposed in this work, autocut* and elbow. The experimental results show the benefits of our strategy in producing factual and relevant answers when compared to a fixed-$k$.

Large Reasoning Models (LRMs) have achieved remarkable performance on complex tasks by engaging in extended reasoning before producing final answers, yet this strength introduces the risk of overthinking, where excessive token generation occurs even for simple tasks. While recent work in efficient reasoning seeks to reduce reasoning length while preserving accuracy, it remains unclear whether such optimization is truly a free lunch. Drawing on the intuition that compressing reasoning may reduce the robustness of model responses and lead models to omit key reasoning steps, we investigate whether efficient reasoning strategies introduce behavioral inconsistencies. To systematically assess this, we introduce $ICBENCH$, a benchmark designed to measure inconsistency in LRMs across three dimensions: inconsistency across task settings (ITS), inconsistency between training objectives and learned behavior (TR-LB), and inconsistency between internal reasoning and self-explanations (IR-SE). Applying $ICBENCH$ to a range of open-source LRMs, we find that while larger models generally exhibit greater consistency than smaller ones, they all display widespread "scheming" behaviors, including self-disagreement, post-hoc rationalization, and the withholding of reasoning cues. Crucially, our results demonstrate that efficient reasoning strategies such as No-Thinking and Simple Token-Budget consistently increase all three defined types of inconsistency. These findings suggest that although efficient reasoning enhances token-level efficiency, further investigation is imperative to ascertain whether it concurrently introduces the risk of models evading effective supervision.

This paper provides preliminary results on exploring the task of performing turn-level data augmentation for dialogue system based on different types of commonsense relationships, and the automatic evaluation of the generated synthetic turns. The proposed methodology takes advantage of the extended knowledge and zero-shot capabilities of pretrained Large Language Models (LLMs) to follow instructions, understand contextual information, and their commonsense reasoning capabilities. The approach draws inspiration from methodologies like Chain-of-Thought (CoT), applied more explicitly to the task of prompt-based generation for dialogue-based data augmentation conditioned on commonsense attributes, and the automatic evaluation of the generated dialogues. To assess the effectiveness of the proposed approach, first we extracted 200 randomly selected partial dialogues, from 5 different well-known dialogue datasets, and generate alternative responses conditioned on different event commonsense attributes. This novel dataset allows us to measure the proficiency of LLMs in generating contextually relevant commonsense knowledge, particularly up to 12 different specific ATOMIC [10] database relations. Secondly, we propose an evaluation framework to automatically detect the quality of the generated dataset inspired by the ACCENT [26] metric, which offers a nuanced approach to assess event commonsense. However, our method does not follow ACCENT's complex eventrelation tuple extraction process. Instead, we propose an instruction-based prompt for each commonsense attribute and use state-of-the-art LLMs to automatically detect the original attributes used when creating each augmented turn in the previous step. Preliminary results suggest that our approach effectively harnesses LLMs capabilities for commonsense reasoning and evaluation in dialogue systems.

--In recent years, significant progress has been made in the field of surgical scene understanding, particularly in the task of Visual Question Localized-Answering in robotic surgery (Surgical-VQLA). However, existing Surgical-VQLA models lack deep reasoning capabilities and interpretability in surgical scenes, which limits their reliability and potential for development in clinical applications. T o address this issue, inspired by the development of Reasoning Multimodal Large Language Models (MLLMs), we first build the Surgery-R1-54k dataset, including paired data for Visual-QA, Grounding-QA, and Chain-of-Thought (CoT). Then, we propose the first Reasoning MLLM for Surgical-VQLA (Surgery-R1). In our Surgery-R1, we design a two-stage fine-tuning mechanism to enable the basic MLLM with complex reasoning abilities by utilizing supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). Furthermore, for an efficient and high-quality rule-based reward system in our RFT, we design a Multimodal Coherence reward mechanism to mitigate positional illusions that may arise in surgical scenarios. Experiment results demonstrate that Surgery-R1 outperforms other existing state-of-the-art (SOT A) models in the Surgical-VQLA task and widely-used MLLMs, while also validating its reasoning capabilities and the effectiveness of our approach. The code and dataset will be organized in https://github.com/ ITH the rapid development of Robot-Assisted Surgery [1], surgical safety and precision have been significantly improved. Surgical scene understanding plays a vital role in this progress, serving as a key prerequisite for achieving intelligent surgery [2], [3]. In recent years, the rapid development of artificial intelligence has led to substantial progress in surgical scene understanding [3], [4].

Multilingual large language models (LLMs) are advancing rapidly, with new models frequently claiming support for an increasing number of languages. However, existing evaluation datasets are limited and lack cross-lingual alignment, leaving assessments of multilingual capabilities fragmented in both language and skill coverage. To address this, we introduce MuBench, a benchmark covering 61 languages and evaluating a broad range of capabilities. We evaluate several state-of-the-art multilingual LLMs and find notable gaps between claimed and actual language coverage, particularly a persistent performance disparity between English and low-resource languages. Leveraging MuBench's alignment, we propose Multilingual Consistency (MLC) as a complementary metric to accuracy for analyzing performance bottlenecks and guiding model improvement. Finally, we pretrain a suite of 1.2B-parameter models on English and Chinese with 500B tokens, varying language ratios and parallel data proportions to investigate cross-lingual transfer dynamics.

Not every causal relation between variables is equal, and this can be leveraged for the task of causal discovery. Recent research shows that pairs of variables with particular type assignments induce a preference on the causal direction of other pairs of variables with the same type. Although useful, this assignment of a specific type to a variable can be tricky in practice. We propose a tag-based causal discovery approach where multiple tags are assigned to each variable in a causal graph. Existing causal discovery approaches are first applied to direct some edges, which are then used to determine edge relations between tags. Then, these edge relations are used to direct the undirected edges. Doing so improves upon purely type-based relations, where the assumption of type consistency lacks robustness and flexibility due to being restricted to single types for each variable. Our experimental evaluations show that this boosts causal discovery and that these high-level tag relations fit common knowledge.

Tokens are fundamental processing units of generative AI (GenAI) and large language models (LLMs), and token communication (TC) is essential for enabling remote AI-generate content (AIGC) and wireless LLM applications. Unlike traditional bits, each of which is independently treated, the semantics of each token depends on its surrounding context tokens. This inter-token dependency makes TC vulnerable to outage channels, where the loss of a single token can significantly distort the original message semantics. Motivated by this, this paper focuses on optimizing token packetization to maximize the average token similarity (ATS) between the original and received token messages under outage channels. Due to inter-token dependency, this token grouping problem is combinatorial, with complexity growing exponentially with message length. To address this, we propose a novel framework of semantic packet aggregation with lookahead search (SemPA-Look), built on two core ideas. First, it introduces the residual semantic score (RSS) as a token-level surrogate for the message-level ATS, allowing robust semantic preservation even when a certain token packet is lost. Second, instead of full search, SemPA-Look applies a lookahead search-inspired algorithm that samples intra-packet token candidates without replacement (fixed depth), conditioned on inter-packet token candidates sampled with replacement (fixed width), thereby achieving linear complexity. Experiments on a remote AIGC task with the MS-COCO dataset (text captioned images) demonstrate that SemPA-Look achieves high ATS and LPIPS scores comparable to exhaustive search, while reducing computational complexity by up to 40$\times$. Compared to other linear-complexity algorithms such as the genetic algorithm (GA), SemPA-Look achieves 10$\times$ lower complexity, demonstrating its practicality for remote AIGC and other TC applications.

Most existing methods rely on the LLM's hidden information (e.g., model parameters or token probabilities), making them ineffective in the black-box setting, where only input and output texts are accessible. Although some methods have been proposed for the black-box setting, they rely on massive manual efforts such as designing complicated questions or instructions. To address these issues, we propose V eilProbe, the first framework for automatically detecting LLMs' pre-training texts in a black-box setting without human intervention. V eilProbe utilizes a sequence-to-sequence mapping model to infer the latent mapping feature between the input text and the corresponding output suffix generated by the LLM. Then it performs the key token perturbations to obtain more distinguishable membership features. Additionally, considering real-world scenarios where the ground-truth training text samples are limited, a prototype-based membership classifier is introduced to alleviate the overfitting issue. Extensive evaluations on three widely used datasets demonstrate that our framework is effective and superior in the black-box setting.