We identify and analyse errors in the popular Massive Multitask Language Understanding (MMLU) benchmark. Even though MMLU is widely adopted, our analysis demonstrates numerous ground truth errors that obscure the true capabilities of LLMs. For example, we find that 57% of the analysed questions in the Virology subset contain errors. To address this issue, we introduce a comprehensive framework for identifying dataset errors using a novel error taxonomy. Then, we create MMLU-Redux, which is a subset of 3,000 manually re-annotated questions across 30 MMLU subjects. Using MMLU-Redux, we demonstrate significant discrepancies with the model performance metrics that were originally reported. Our results strongly advocate for revising MMLU's error-ridden questions to enhance its future utility and reliability as a benchmark.

Meta MMO is built on top of Neural MMO, a massively multiagent environment that has been the subject of two previous NeurIPS competitions. Our work expands Neural MMO with several computationally efficient minigames. We explore generalization across Meta MMO by learning to play several minigames with a single set of weights. We release the environment, baselines, and training code under the MIT license. We hope that Meta MMO will spur additional progress on Neural MMO and, more generally, will serve as a useful benchmark for many-agent generalization.

Although safely enhanced Large Language Models (LLMs) have achieved remarkable success in tackling various complex tasks in a zero-shot manner, they remain susceptible to jailbreak attacks, particularly the unknown jailbreak attack. To enhance LLMs' generalized defense capabilities, we propose a two-stage adversarial tuning framework, which generates adversarial prompts to explore worst-case scenarios by optimizing datasets containing pairs of adversarial prompts and their safe responses. In the first stage, we introduce the hierarchical meta-universal adversarial prompt learning to efficiently and effectively generate token-level adversarial prompts. In the second stage, we propose the automatic adversarial prompt learning to iteratively refine semantic-level adversarial prompts, further enhancing LLM's defense capabilities. We conducted comprehensive experiments on three widely used jailbreak datasets, comparing our framework with six defense baselines under five representative attack scenarios. The results underscore the superiority of our proposed methods. Furthermore, our adversarial tuning framework exhibits empirical generalizability across various attack strategies and target LLMs, highlighting its potential as a transferable defense mechanism.

Dexterous in-hand manipulation is a unique and valuable human skill requiring sophisticated sensorimotor interaction with the environment while respecting stability constraints. Satisfying these constraints with generated motions is essential for a robotic platform to achieve reliable in-hand manipulation skills. Explicitly modelling these constraints can be challenging, but they can be implicitly modelled and learned through experience or human demonstrations. We propose a learning and control approach based on dictionaries of motion primitives generated from human demonstrations. To achieve this, we defined an optimization process that combines motion primitives to generate robot fingertip trajectories for moving an object from an initial to a desired final pose. Based on our experiments, our approach allows a robotic hand to handle objects like humans, adhering to stability constraints without requiring explicit formalization. In other words, the proposed motion primitive dictionaries learn and implicitly embed the constraints crucial to the in-hand manipulation task.

Selective classification enables models to make predictions only when they are sufficiently confident, aiming to enhance safety and reliability, which is important in high-stakes scenarios. Previous methods mainly use deep neural networks and focus on modifying the architecture of classification layers to enable the model to estimate the confidence of its prediction. This work provides a generalization bound for selective classification, disclosing that optimizing feature layers helps improve the performance of selective classification. Inspired by this theory, we propose to explicitly improve the selective classification model at the feature level for the first time, leading to a novel Confidence-aware Contrastive Learning method for Selective Classification, CCL-SC, which similarizes the features of homogeneous instances and differentiates the features of heterogeneous instances, with the strength controlled by the model's confidence. The experimental results on typical datasets, i.e., CIFAR-10, CIFAR-100, CelebA, and ImageNet, show that CCL-SC achieves significantly lower selective risk than state-of-the-art methods, across almost all coverage degrees. Moreover, it can be combined with existing methods to bring further improvement.

Accurate estimates of Above Ground Biomass (AGB) are essential in addressing two of humanity's biggest challenges, climate change and biodiversity loss. Existing datasets for AGB estimation from satellite imagery are limited. Either they focus on specific, local regions at high resolution, or they offer global coverage at low resolution. There is a need for a machine learning-ready, globally representative, high-resolution benchmark. Our findings indicate significant variability in biomass estimates across different vegetation types, emphasizing the necessity for a dataset that accurately captures global diversity. To address these gaps, we introduce a comprehensive new dataset that is globally distributed, covers a range of vegetation types, and spans several years. This dataset combines AGB reference data from the GEDI mission with data from Sentinel-2 and PALSAR-2 imagery. Additionally, it includes pre-processed high-level features such as a dense canopy height map, an elevation map, and a land-cover classification map. We also produce a dense, high-resolution (10m) map of AGB predictions for the entire area covered by the dataset. Rigorously tested, our dataset is accompanied by several benchmark models and is publicly available. It can be easily accessed using a single line of code, offering a solid basis for efforts towards global AGB estimation. The GitHub repository github.com/ghjuliasialelli/AGBD serves as a one-stop shop for all code and data.

Millions of news articles published online daily can overwhelm readers. Headlines and entity (topic) tags are essential for guiding readers to decide if the content is worth their time. While headline generation has been extensively studied, tag generation remains largely unexplored, yet it offers readers better access to topics of interest. The need for conciseness in capturing readers' attention necessitates improved content selection strategies for identifying salient and relevant segments within lengthy articles, thereby guiding language models effectively. To address this, we propose to leverage auxiliary information such as images and captions embedded in the articles to retrieve relevant sentences and utilize instruction tuning with variations to generate both headlines and tags for news articles in a multilingual context. To make use of the auxiliary information, we have compiled a dataset named XL-HeadTags, which includes 20 languages across 6 diverse language families. Through extensive evaluation, we demonstrate the effectiveness of our plug-and-play multimodal-multilingual retrievers for both tasks. Additionally, we have developed a suite of tools for processing and evaluating multilingual texts, significantly contributing to the research community by enabling more accurate and efficient analysis across languages.

This paper explores the concept formation and alignment within the realm of language models (LMs). We propose a mechanism for identifying concepts and their hierarchical organization within the semantic representations learned by various LMs, encompassing a spectrum from early models like Glove to the transformer-based language models like ALBERT and T5. Our approach leverages the inherent structure present in the semantic embeddings generated by these models to extract a taxonomy of concepts and their hierarchical relationships. This investigation sheds light on how LMs develop conceptual understanding and opens doors to further research to improve their ability to reason and leverage real-world knowledge. We further conducted experiments and observed the possibility of isolating these extracted conceptual representations from the reasoning modules of the transformer-based LMs. The observed concept formation along with the isolation of conceptual representations from the reasoning modules can enable targeted token engineering to open the door for potential applications in knowledge transfer, explainable AI, and the development of more modular and conceptually grounded language models.

Pseudocode in a scholarly paper provides a concise way to express the algorithms implemented therein. Pseudocode can also be thought of as an intermediary representation that helps bridge the gap between programming languages and natural languages. Having access to a large collection of pseudocode can provide various benefits ranging from enhancing algorithmic understanding, facilitating further algorithmic design, to empowering NLP or computer vision based models for tasks such as automated code generation and optical character recognition (OCR). We have created a large pseudocode collection by extracting nearly 320,000 pseudocode examples from arXiv papers. This process involved scanning over $2.2$ million scholarly papers, with 1,000 of them being manually inspected and labeled. Our approach encompasses an extraction mechanism tailored to optimize the coverage and a validation mechanism based on random sampling to check its accuracy and reliability, given the inherent heterogeneity of the collection. In addition, we offer insights into common pseudocode structures, supported by clustering and statistical analyses. Notably, these analyses indicate an exponential-like growth in the usage of pseudocodes, highlighting their increasing significance.

Large Language Models (LLMs) have shown exceptional performance in text processing. Notably, LLMs can synthesize information from large datasets and explain their decisions similarly to human reasoning through a chain of thought (CoT). An emerging application of LLMs is the handling and interpreting of numerical data, where fine-tuning enhances their performance over basic inference methods. This paper proposes a novel approach to training LLMs using knowledge transfer from a random forest (RF) ensemble, leveraging its efficiency and accuracy. By converting RF decision paths into natural language statements, we generate outputs for LLM fine-tuning, enhancing the model's ability to classify and explain its decisions. Our method includes verifying these rules through established classification metrics, ensuring their correctness. We also examine the impact of preprocessing techniques on the representation of numerical data and their influence on classification accuracy and rule correctness