Radiology report evaluation is a crucial part of radiologists' training and plays a key role in ensuring diagnostic accuracy. As part of the standard reporting workflow, a junior radiologist typically prepares a preliminary report, which is then reviewed and edited by a senior radiologist to produce the final report. Identifying semantic differences between preliminary and final reports is essential for junior doctors, both as a training tool and to help uncover gaps in clinical knowledge. While AI in radiology is a rapidly growing field, the application of large language models (LLMs) remains challenging due to the need for specialised domain knowledge. In this paper, we explore the ability of LLMs to provide explainable and accurate comparisons of reports in the radiology domain. We begin by comparing the performance of several LLMs in comparing radiology reports. We then assess a more traditional approach based on Named-Entity-Recognition (NER). However, both approaches exhibit limitations in delivering accurate feedback on semantic similarity. To address this, we propose Llama-EntScore, a semantic similarity scoring method using a combination of Llama 3.1 and NER with tunable weights to emphasise or de-emphasise specific types of differences. Our approach generates a quantitative similarity score for tracking progress and also gives an interpretation of the score that aims to offer valuable guidance in reviewing and refining their reporting. We find our method achieves 67% exact-match accuracy and 93% accuracy within +/- 1 when compared to radiologist-provided ground truth scores - outperforming both LLMs and NER used independently. Code is available at: https://github.com/otmive/llama_reports

Compositional generalization, the ability to generate novel combinations of known concepts, is a key ingredient for visual generative models. Yet, not all mechanisms that enable or inhibit it are fully understood. In this work, we conduct a systematic study of how various design choices influence compositional generalization in image and video generation in a positive or negative way. Through controlled experiments, we identify two key factors: (i) whether the training objective operates on a discrete or continuous distribution, and (ii) to what extent conditioning provides information about the constituent concepts during training. Building on these insights, we show that relaxing the MaskGIT discrete loss with an auxiliary continuous JEPA-based objective can improve compositional performance in discrete models like MaskGIT.

With the rapid advancement of AI-generated videos, there is an urgent need for effective detection tools to mitigate societal risks such as misinformation and reputational harm. In addition to accurate classification, it is essential that detection models provide interpretable explanations to ensure transparency for regulators and end users. To address these challenges, we introduce VidGuard-R1, the first video authenticity detector that fine-tunes a multi-modal large language model (MLLM) using group relative policy optimization (GRPO). Our model delivers both highly accurate judgments and insightful reasoning. We curate a challenging dataset of 140k real and AI-generated videos produced by state-of-the-art generation models, carefully designing the generation process to maximize discrimination difficulty. We then fine-tune Qwen-VL using GRPO with two specialized reward models that target temporal artifacts and generation complexity. Extensive experiments demonstrate that VidGuard-R1 achieves state-of-the-art zero-shot performance on existing benchmarks, with additional training pushing accuracy above 95%. Case studies further show that VidGuard-R1 produces precise and interpretable rationales behind its predictions. The code is publicly available at https://VidGuard-R1.github.io.

In this study, we utilized the multimodal capabilities of OpenAI GPT-4o to automatically generate jaw cyst findings on dental panoramic radiographs. To improve accuracy, we constructed a Self-correction Loop with Structured Output (SLSO) framework and verified its effectiveness. A 10-step process was implemented for 22 cases of jaw cysts, including image input and analysis, structured data generation, tooth number extraction and consistency checking, iterative regeneration when inconsistencies were detected, and finding generation with subsequent restructuring and consistency verification. A comparative experiment was conducted using the conventional Chain-of-Thought (CoT) method across seven evaluation items: transparency, internal structure, borders, root resorption, tooth movement, relationships with other structures, and tooth number. The results showed that the proposed SLSO framework improved output accuracy for many items, with 66.9%, 33.3%, and 28.6% improvement rates for tooth number, tooth movement, and root resorption, respectively. In the successful cases, a consistently structured output was achieved after up to five regenerations. Although statistical significance was not reached because of the small size of the dataset, the overall SLSO framework enforced negative finding descriptions, suppressed hallucinations, and improved tooth number identification accuracy. However, the accurate identification of extensive lesions spanning multiple teeth is limited. Nevertheless, further refinement is required to enhance overall performance and move toward a practical finding generation system.

Padding tokens are widely used in large language models (LLMs) to equalize sequence lengths during batched inference. While they should be fully masked, implementation errors can cause them to influence computation, and the extent of this influence is not well understood. We systematically study this effect across three open-source model families (Llama, Gemma, Qwen), inserting controlled amounts of padding and evaluating outcomes along four axes: activations, generation quality, bias, and safety. Even small amounts of padding shift hidden representations, degrade quality in smaller models, alter bias in unpredictable ways, and weaken safety guardrails. These findings demonstrate that padding is not a harmless detail but a robustness risk that must be carefully handled in deployment.

Large Language Models (LLMs) excel at problem solving by generating chain of thoughts in natural language, but such verbal thinking is computationally costly and prone to overthinking. Recent work instead proposes a latent thinking architecture Huginn-3.5B, which represents intermediate reasoning steps as sequence of latent representations. However, latent thoughts lack interpretability and are difficult to supervise, raising concerns about the correctness and reliability of its latent thinking processes. In this paper, we provide a systematic study of how Huginn-3.5B thinks in the latent space and how external supervision signals can improve its latent thinking processes. We show that latent thoughts leading to correct versus incorrect answers exhibit highly distinguishable patterns, and that a latent classifier can reliably predict answer correctness directly from latent thoughts. Leveraging these insights, we propose Latent Thinking Optimization (LTO), a probabilistic algorithm that employs the latent classifier as a Latent Reward Model (LRM) to optimize the latent thinking processes. Extensive experiments across diverse reasoning tasks demonstrate that LRM is highly effective in detecting incorrect latent thinking patterns, and LTO can significantly improve the latent thinking processes. Furthermore, we show that LRM can generalize across diverse domains, and LTO can be seamlessly applied to general LLMs to improve their thinking processes. In contrast to verbal thinking, our method demonstrates that reward modeling and scaling test-time thinking with supervision can be performed directly in the latent space, highlighting its potential as a general, efficient, and domain-agnostic approach to improving the thinking processes of LLMs.

Large language models (LLMs) have recently demonstrated strong capabilities in generating machine learning (ML) code, enabling end-to-end pipeline construction from natural language instructions. However, existing benchmarks for ML code generation are mainly restricted to English, overlooking the global and multilingual nature of ML research and practice. To address this gap, we present ML2B, the first benchmark for evaluating multilingual ML code generation. ML2B consists of 30 Kaggle competitions translated into 13 natural languages, covering tabular, text, and image data types, with structured metadata and validated human-reviewed translations. For evaluation, we employ AIDE, an automated framework for end-to-end assessment of data science pipelines, and provide insights into cross-lingual model performance. Our results reveal substantial 15-45% performance degradation on non-English tasks, highlighting critical challenges in multilingual representation learning for code generation. The benchmark, evaluation framework, and comprehensive results are made available through our GitHub repository to facilitate future research in multilingual ML code generation: https://github.com/enaix/ml2b.

Large language models are rapidly transforming social science research by enabling the automation of labor-intensive tasks like data annotation and text analysis. However, LLM outputs vary significantly depending on the implementation choices made by researchers (e.g., model selection or prompting strategy). Such variation can introduce systematic biases and random errors, which propagate to downstream analyses and cause Type I (false positive), Type II (false negative), Type S (wrong sign), or Type M (exaggerated effect) errors. We call this phenomenon where configuration choices lead to incorrect conclusions LLM hacking. We find that intentional LLM hacking is strikingly simple. By replicating 37 data annotation tasks from 21 published social science studies, we show that, with just a handful of prompt paraphrases, virtually anything can be presented as statistically significant. Beyond intentional manipulation, our analysis of 13 million labels from 18 different LLMs across 2361 realistic hypotheses shows that there is also a high risk of accidental LLM hacking, even when following standard research practices. We find incorrect conclusions in approximately 31% of hypotheses for state-of-the-art LLMs, and in half the hypotheses for smaller language models. While higher task performance and stronger general model capabilities reduce LLM hacking risk, even highly accurate models remain susceptible. The risk of LLM hacking decreases as effect sizes increase, indicating the need for more rigorous verification of LLM-based findings near significance thresholds. We analyze 21 mitigation techniques and find that human annotations provide crucial protection against false positives. Common regression estimator correction techniques can restore valid inference but trade off Type I vs. Type II errors. We publish a list of practical recommendations to prevent LLM hacking.

Large, open datasets can accelerate ecological research, particularly by enabling researchers to develop new insights by reusing datasets from multiple sources. However, to find the most suitable datasets to combine and integrate, researchers must navigate diverse ecological and environmental data provider platforms with varying metadata availability and standards. To overcome this obstacle, we have developed a large language model (LLM)-based metadata harvester that flexibly extracts metadata from any dataset's landing page, and converts these to a user-defined, unified format using existing metadata standards. We validate that our tool is able to extract both structured and unstructured metadata with equal accuracy, aided by our LLM post-processing protocol. Furthermore, we utilise LLMs to identify links between datasets, both by calculating embedding similarity and by unifying the formats of extracted metadata to enable rule-based processing. Our tool, which flexibly links the metadata of different datasets, can therefore be used for ontology creation or graph-based queries, for example, to find relevant ecological and environmental datasets in a virtual research environment.

Reinforcement Learning with Verifiable Reward (RLVR) has emerged as a powerful paradigm for unlocking reasoning capabilities in large language models, yet its full potential is hindered by two under-explored dimensions: Depth-the hardest problem a model can sample; Breadth-the number of instances consumed in a single iteration. We dissect the popular GRPO algorithm and reveal a systematic bias: the cumulative-advantage disproportionately weights samples with medium accuracy, while down-weighting the low-accuracy instances that are crucial for pushing reasoning boundaries. To rectify the depth neglect, we introduce Difficulty Adaptive Rollout Sampling (DARS), which re-weights hard problems through targeted multi-stage rollouts, thereby increasing the number of positive rollouts for hard problems. Empirically, naively enlarging rollout size only accelerates convergence and even hurts Pass@K. Our DARS, in contrast, delivers consistent Pass@K gains without extra inference cost at convergence. Just as we adaptively expanded the depth of exploration, we now ask whether aggressively scaling the breadth of training data can further amplify reasoning gains. To this end, we intensely scale batch size and replace PPO's mini-batch iterations with full-batch updates over multiple epochs. Increasing breadth significantly enhances Pass@1 performance. Large-breadth training sustains high token-level entropy, indicating continued exploration and reduced gradient noise. We further present DARS-B, which augments DARS with large breadth, and demonstrate simultaneous gains in Pass@K and Pass@1. The results confirm that breadth and adaptive exploration across depth operate as orthogonal dimensions in RLVR, which are key to unleashing the reasoning power of RLVR.