Gerard de Melo

Watermarking has emerged as a pivotal solution for content traceability and intellectual property protection in Large Vision-Language Models (LVLMs). However, vision-agnostic watermarks may introduce visually irrelevant tokens and disrupt visual grounding by enforcing indiscriminate pseudo-random biases. Additionally, current vision-specific watermarks rely on a static, one-time estimation of vision critical weights and ignore the weight distribution density when determining the proportion of protected tokens. This design fails to account for dynamic changes in visual dependence during generation and may introduce low-quality tokens in the long tail. To address these challenges, we propose Attention-Guided Dynamic Watermarking (AGMark), a novel framework that embeds detectable signals while strictly preserving visual fidelity. At each decoding step, AGMark first dynamically identifies semantic-critical evidence based on attention weights for visual relevance, together with context-aware coherence cues, resulting in a more adaptive and well-calibrated evidence-weight distribution. It then determines the proportion of semantic-critical tokens by jointly considering uncertainty awareness (token entropy) and evidence calibration (weight density), thereby enabling adaptive vocabulary partitioning to avoid irrelevant tokens. Empirical results confirm that AGMark outperforms conventional methods, observably improving generation quality and yielding particularly strong gains in visual semantic fidelity in the later stages of generation. The framework maintains highly competitive detection accuracy (at least 99.36\% AUC) and robust attack resilience (at least 88.61\% AUC) without sacrificing inference efficiency, effectively establishing a new standard for reliability-preserving multi-modal watermarking.

The availability of structured legal data is important for advancing Natural Language Processing (NLP) techniques for the German legal system. One of the most widely used datasets, Open Legal Data, provides a large-scale collection of German court decisions. While the metadata in this raw dataset is consistently structured, the decision texts themselves are inconsistently formatted and often lack clearly marked sections. Reliable separation of these sections is important not only for rhetorical role classification but also for downstream tasks such as retrieval and citation analysis. In this work, we introduce a cleaned and sectioned dataset of 251,038 German court decisions derived from the official Open Legal Data dataset. We systematically separated three important sections in German court decisions, namely Tenor (operative part of the decision), Tatbestand (facts of the case), and Entscheidungsgründe (judicial reasoning), which are often inconsistently represented in the original dataset. To ensure the reliability of our extraction process, we used Cochran’s formula with a 95% confidence level and a 5% margin of error to draw a statistically representative random sample of 384 cases, and manually verified that all three sections were correctly identified. We also extracted the Rechtsmittelbelehrung (appeal notice) as a separate field, since it is a procedural instruction and not part of the decision itself. The resulting corpus is publicly available in the JSONL format, making it an accessible resource for further research on the German legal system.

Introduction: The rapid advancement of artificial intelligence (AI) has created new opportunities in educational research, particularly in the efficient analysis of complex social interactions within classrooms. One promising area involves the classification of teachers' motivational messages. Traditionally, such assessments have relied on self‐reports and observer evaluations, which require a lot of staff and time resources. Recently, large language models (LLMs) have been employed to classify teachers' motivational messages, offering novel, less labour‐intensive approaches for classification. Aims: Building on these recent developments, this work presents a comprehensive literature overview exploring the applications, potential, and limitations of using LLMs to classify teachers' motivational messages. Results: The present comprehensive literature overview indicates that the use of LLMs for classifying teachers' motivational messages is a promising yet still emerging field of research. Recent studies have applied LLMs in innovative ways, drawing on established motivational theories and employing novel classification techniques, such as zero‐shot and few‐shot prompting or fine‐tuning, to classify motivational messages. Open questions remain, particularly concerning the structure, quantity, and quality of annotated material. Discussion: Whereas recent studies have demonstrated the potential of LLMs to offer scalable and time‐efficient alternatives for classifying motivational messages in the classroom, several challenges persist. These include concerns related to the quality and quantity of training data, model generalisability, the ability to capture the complexity of classroom interactions, and biases involved in integrating LLMs as a classification method. This comprehensive literature overview provides practical recommendations for the responsible use of LLMs in educational research and school practice.

Background: Recent studies have examined the relation between teacher motivation, motivational messages and student learning but are limited to an achievement‐related context, primarily using survey data. Moreover, our understanding of the relation between various teacher characteristics, such as teacher self‐efficacy (TSE), and their motivational message use remains limited. Aims: Our study tested whether teacher speech can be classified into self‐determination (SDT)‐based motivational messages and reliably assessed with a large language model (LLM). Additionally, we analysed the relation between pre‐service TSE and their motivational message use. Sample: For our first aim, we used human‐rater annotations from 119 pre‐service teachers' classroom recordings. For our second aim, we used data from 103 pre‐service teachers (52.69% female; M _age = 22.98, SD _age = 3.26, Min _age = 19, Max _age = 34) who participated in a survey and were video‐recorded while teaching. Methods: First, we manually classified pre‐service teachers' motivational messages based on transcripts and used human‐rater annotations to fine‐tune an LLM. Second, we analysed the relation between pre‐service TSE and motivational message use. Results and Conclusions: The fine‐tuned LLM demonstrated promising performance in assessing SDT‐based motivational messages but needs further refining to assess thwarting messages. The analysis with human annotation showed that pre‐service TSE for classroom management positively affected the frequency of relatedness‐supportive messages. Pre‐service TSE for student engagement increased the likelihood of never using a competence‐ or relatedness‐thwarting message. Pre‐service TSE for instructional strategies reduced the frequency of autonomy‐supportive messages. LLM‐based analyses showed slightly different results but did not contradict human annotation‐based analyses.

Xiechi Zhang, Zetian Ouyang, Linlin Wang, Gerard De Melo, Zhu Cao, Xiaoling Wang, Ya Zhang, Yanfeng Wang, Liang He. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Yue Li, Xin Yi, Dongsheng Shi, Gerard De Melo, Xiaoling Wang, Linlin Wang. Findings of the Association for Computational Linguistics: ACL 2025. 2025.

With the proliferation of Large Language Models (LLMs) in diverse domains, there is a particular need for unified evaluation standards in clinical medical scenarios, where models need to be examined very thoroughly. We present CliMedBench, a comprehensive benchmark with 14 expert-guided core clinical scenarios specifically designed to assess the medical ability of LLMs across 7 pivot dimensions. It comprises 33,735 questions derived from real-world medical reports of top-tier tertiary hospitals and authentic examination exercises. The reliability of this benchmark has been confirmed in several ways. Subsequent experiments with existing LLMs have led to the following findings: (i) Chinese medical LLMs underperform on this benchmark, especially where medical reasoning and factual consistency are vital, underscoring the need for advances in clinical knowledge and diagnostic accuracy. (ii) Several general-domain LLMs demonstrate substantial potential in medical clinics, while the limited input capacity of many medical LLMs hinders their practical use. These findings reveal both the strengths and limitations of LLMs in clinical scenarios and offer critical insights for medical research.

The WallStreetBets (WSB) community on Reddit gained prominence for its role in the GameStop saga and the resulting meme stock phenomenon. Concurrently, this has boosted the popularity of finance-related communities on Reddit, with the top five totaling more than 25 million subscribers at the time of writing. However, little is known about the reliability of the advice disseminated in these communities, which is a relevant research question within the field of social computing. In this study, we examine the collective intelligence of WSB, the largest and most active subreddit focused on the stock market, and assess its potential as a democratizing force in enabling access to financial knowledge. First, we establish that WSB meets several criteria to be considered a collectively intelligent crowd. Then, we test our hypothesis quantitatively by analyzing Reddit posts and financial data from a 28-month period to evaluate how successful an investor relying on WSB recommendations could have been. We define a portfolio of WSB’s most discussed stocks which shows significant growth, outperforming the S&P 500 index over the reviewed time frame. We further find that following buy signals at the time they are posted on WSB leads to positive outcomes over the long run, and that the GameStop hype merely amplified previously existing characteristics. The WSB portfolio underperforms the broader market during downturns, but recovers more quickly and achieves higher profits afterwards. The results of our work can be generalized to comparable finance-related communities, indicating that their original purpose of leisurely entertainment has already been extended towards tangible real-world value.