As visual generation technologies continue to advance, the scale of video datasets has expanded rapidly, and the quality of these datasets is critical to the performance of video generation models. We argue that temporal splitting, detailed captions, and video quality filtering are three key factors that determine dataset quality. However, existing datasets exhibit various limitations in these areas. To address these challenges, we introduce Koala-36M, a large-scale, high-quality video dataset featuring accurate temporal splitting, detailed captions, and superior video quality. The core of our approach lies in improving the consistency between fine-grained conditions and video content. Specifically, we employ a linear classifier on probability distributions to enhance the accuracy of transition detection, ensuring better temporal consistency. We then provide structured captions for the splitted videos, with an average length of 200 words, to improve text-video alignment. Additionally, we develop a Video Training Suitability Score (VTSS) that integrates multiple sub-metrics, allowing us to filter high-quality videos from the original corpus. Finally, we incorporate several metrics into the training process of the generation model, further refining the fine-grained conditions. Our experiments demonstrate the effectiveness of our data processing pipeline and the quality of the proposed Koala-36M dataset. Our dataset and code will be released at https://koala36m.github.io/.

We show that these attractor representations optimize generalization to unseen examples by exploiting properties of knowledge graph relations (e.g. We find experimental support for such universality by showing that LLMs and simpler neural networks can be stitched, i.e., by stitching the first part of one model to the last part of another, mediated only by an affine or almost affine transformation. We hypothesize that this dynamic toward simplicity and generalization is driven by "intelligence from starvation": where overfitting is minimized by pressure to minimize the use of resources that are either scarce or competed for against other tasks. Large Language Models (LLMs), despite being primarily trained for next-token predictions, have shown impressive reasoning capabilities (Bubeck et al., 2023; Anthropic, 2024; Team et al., 2023). However, despite recent progress reviewed below, it is not well understood what knowledge LLMs represent internally and how they represent it. Improving such understanding could enable valuable progress relevant to transparency, interpretability, fairness and robustness, for example discovering and correcting inaccuracies to improve model reliability.

We introduce Pixtral-12B, a 12--billion-parameter multimodal language model. Pixtral-12B is trained to understand both natural images and documents, achieving leading performance on various multimodal benchmarks, surpassing a number of larger models. Unlike many open-source models, Pixtral is also a cutting-edge text model for its size, and does not compromise on natural language performance to excel in multimodal tasks. Pixtral uses a new vision encoder trained from scratch, which allows it to ingest images at their natural resolution and aspect ratio. This gives users flexibility on the number of tokens used to process an image. Pixtral is also able to process any number of images in its long context window of 128K tokens. Pixtral 12B substanially outperforms other open models of similar sizes (Llama-3.2 11B \& Qwen-2-VL 7B). It also outperforms much larger open models like Llama-3.2 90B while being 7x smaller. We further contribute an open-source benchmark, MM-MT-Bench, for evaluating vision-language models in practical scenarios, and provide detailed analysis and code for standardized evaluation protocols for multimodal LLMs. Pixtral-12B is released under Apache 2.0 license.

We present VoxelPrompt, an agent-driven vision-language framework that tackles diverse radiological tasks through joint modeling of natural language, image volumes, and analytical metrics. VoxelPrompt is multi-modal and versatile, leveraging the flexibility of language interaction while providing quantitatively grounded image analysis. Given a variable number of 3D medical volumes, such as MRI and CT scans, VoxelPrompt employs a language agent that iteratively predicts executable instructions to solve a task specified by an input prompt. These instructions communicate with a vision network to encode image features and generate volumetric outputs (e.g., segmentations). VoxelPrompt interprets the results of intermediate instructions and plans further actions to compute discrete measures (e.g., tumor growth across a series of scans) and present relevant outputs to the user. We evaluate this framework in a sandbox of diverse neuroimaging tasks, and we show that the single VoxelPrompt model can delineate hundreds of anatomical and pathological features, measure many complex morphological properties, and perform open-language analysis of lesion characteristics. VoxelPrompt carries out these objectives with accuracy similar to that of fine-tuned, single-task models for segmentation and visual question-answering, while facilitating a much larger range of tasks. Therefore, by supporting accurate image processing with language interaction, VoxelPrompt provides comprehensive utility for numerous imaging tasks that traditionally require specialized models to address.

Patients with schizophrenia often present with cognitive impairments that may hinder their ability to learn about their condition. These individuals could benefit greatly from education platforms that leverage the adaptability of Large Language Models (LLMs) such as GPT-4. While LLMs have the potential to make topical mental health information more accessible and engaging, their black-box nature raises concerns about ethics and safety. Prompting offers a way to produce semi-scripted chatbots with responses anchored in instructions and validated information, but prompt-engineered chatbots may drift from their intended identity as the conversation progresses. We propose a Critical Analysis Filter for achieving better control over chatbot behavior. In this system, a team of prompted LLM agents are prompt-engineered to critically analyze and refine the chatbot's response and deliver real-time feedback to the chatbot. To test this approach, we develop an informational schizophrenia chatbot and converse with it (with the filter deactivated) until it oversteps its scope. Once drift has been observed, AI-agents are used to automatically generate sample conversations in which the chatbot is being enticed to talk about out-of-bounds topics. We manually assign to each response a compliance score that quantifies the chatbot's compliance to its instructions; specifically the rules about accurately conveying sources and being transparent about limitations. Activating the Critical Analysis Filter resulted in an acceptable compliance score (>=2) in 67.0% of responses, compared to only 8.7% when the filter was deactivated. These results suggest that a self-reflection layer could enable LLMs to be used effectively and safely in mental health platforms, maintaining adaptability while reliably limiting their scope to appropriate use cases.

Language models are extensively evaluated, but correctly interpreting evaluation results requires knowledge of train-test overlap which refers to the extent to which the language model is trained on the very data it is being tested on. The public currently lacks adequate information about train-test overlap: most models have no public train-test overlap statistics, and third parties cannot directly measure train-test overlap since they do not have access to the training data. To make this clear, we document the practices of 30 model developers, finding that just 9 developers report train-test overlap: 4 developers release training data under open-source licenses, enabling the community to directly measure train-test overlap, and 5 developers publish their train-test overlap methodology and statistics. By engaging with language model developers, we provide novel information about train-test overlap for three additional developers. Overall, we take the position that language model developers should publish train-test overlap statistics and/or training data whenever they report evaluation results on public test sets. We hope our work increases transparency into train-test overlap to increase the community-wide trust in model evaluations.

The improvement of economic policymaking presents an opportunity for broad societal benefit, a notion that has inspired research towards AI-driven policymaking tools. AI policymaking holds the potential to surpass human performance through the ability to process data quickly at scale. However, existing RL-based methods exhibit sample inefficiency, and are further limited by an inability to flexibly incorporate nuanced information into their decision-making processes. Thus, we propose a novel method in which we instead utilize pre-trained Large Language Models (LLMs), as sample-efficient policymakers in socially complex multi-agent reinforcement learning (MARL) scenarios. We demonstrate significant efficiency gains, outperforming existing methods across three environments. Our code is available at https://github.com/hegasz/large-legislative-models.

Language models (LMs) are known to suffer from hallucinations and misinformation. Retrieval augmented generation (RAG) that retrieves verifiable information from an external knowledge corpus to complement the parametric knowledge in LMs provides a tangible solution to these problems. However, the generation quality of RAG is highly dependent on the relevance between a user's query and the retrieved documents. Inaccurate responses may be generated when the query is outside of the scope of knowledge represented in the external knowledge corpus or if the information in the corpus is out-of-date. In this work, we establish a statistical framework that assesses how well a query can be answered by an RAG system by capturing the relevance of knowledge. We introduce an online testing procedure that employs goodness-of-fit (GoF) tests to inspect the relevance of each user query to detect out-of-knowledge queries with low knowledge relevance. Additionally, we develop an offline testing framework that examines a collection of user queries, aiming to detect significant shifts in the query distribution which indicates the knowledge corpus is no longer sufficiently capable of supporting the interests of the users. We demonstrate the capabilities of these strategies through a systematic evaluation on eight question-answering (QA) datasets, the results of which indicate that the new testing framework is an efficient solution to enhance the reliability of existing RAG systems.

As the cultural heritage sector increasingly adopts technologies like Retrieval-Augmented Generation (RAG) to provide more personalised search experiences and enable conversations with collections data, the demand for specialised evaluation datasets has grown. While end-to-end system testing is essential, it's equally important to assess individual components. We target the final, answering task, which is well-suited to Machine Reading Comprehension (MRC). Although existing MRC datasets address general domains, they lack the specificity needed for cultural heritage information. Unfortunately, the manual creation of such datasets is prohibitively expensive for most heritage institutions. This paper presents a cost-effective approach for generating domain-specific MRC datasets with increased difficulty using Reinforcement Learning from Human Feedback (RLHF) from synthetic preference data. Our method leverages the performance of existing question-answering models on a subset of SQuAD to create a difficulty metric, assuming that more challenging questions are answered correctly less frequently. This research contributes: (1) A methodology for increasing question difficulty using PPO and synthetic data; (2) Empirical evidence of the method's effectiveness, including human evaluation; (3) An in-depth error analysis and study of emergent phenomena; and (4) An open-source codebase and set of three llama-2-chat adapters for reproducibility and adaptation.

Citation networks are critical in modern science, and predicting which previous papers (candidates) will a new paper (query) cite is a critical problem. However, the roles of a paper's citations vary significantly, ranging from foundational knowledge basis to superficial contexts. Distinguishing these roles requires a deeper understanding of the logical relationships among papers, beyond simple edges in citation networks. The emergence of LLMs with textual reasoning capabilities offers new possibilities for discerning these relationships, but there are two major challenges. First, in practice, a new paper may select its citations from gigantic existing papers, where the texts exceed the context length of LLMs. Second, logical relationships between papers are implicit, and directly prompting an LLM to predict citations may result in surface-level textual similarities rather than the deeper logical reasoning. In this paper, we introduce the novel concept of core citation, which identifies the critical references that go beyond superficial mentions. Thereby, we elevate the citation prediction task from a simple binary classification to distinguishing core citations from both superficial citations and non-citations. To address this, we propose $\textbf{HLM-Cite}$, a $\textbf{H}$ybrid $\textbf{L}$anguage $\textbf{M}$odel workflow for citation prediction, which combines embedding and generative LMs. We design a curriculum finetune procedure to adapt a pretrained text embedding model to coarsely retrieve high-likelihood core citations from vast candidates and then design an LLM agentic workflow to rank the retrieved papers through one-shot reasoning, revealing the implicit relationships among papers. With the pipeline, we can scale the candidate sets to 100K papers. We evaluate HLM-Cite across 19 scientific fields, demonstrating a 17.6% performance improvement comparing SOTA methods.