Annotating data is a time-consuming and costly task, but it is inherently required for supervised machine learning. Active Learning (AL) is an established method that minimizes human labeling effort by iteratively selecting the most informative unlabeled samples for expert annotation, thereby improving the overall classification performance. Even though AL has been known for decades [1], AL is still rarely used in real-world applications. As indicated in the two community web surveys among the NLP community about AL [2], [3], two main reasons continue to hold practitioners back from using AL: first, the complexity of setting AL up, and second, a lack of trust in its effectiveness. We hypothesize that both reasons share the same culprit: the large hyperparameter space of AL. This mostly unexplored hyperparameter space often leads to misleading and irreproducible glsAL experiment results. In this study, we first compiled a large hyperparameter grid of over 4.6 million hyperparameter combinations, second, recorded the performance of all combinations in the so-far biggest conducted AL study, and third, analyzed the impact of each hyperparameter in the experiment results. Rather than merely reporting correlations, we explicitly focus on distilling these results into practitioner-oriented rulesof-thumb for designing AL experiments under realistic resource constraints. In the end, we give recommendations about the influence of each hyperparameter, demonstrate the surprising influence of the concrete AL strategy implementation, and outline an experimental study design for reproducible AL experiments with minimal computational effort, thus contributing to more reproducible and trustworthy AL research in the future.

Do large language models (LLMs) solve reasoning tasks by learning robust generalizable algorithms, or do they memorize training data? To investigate this question, we use arithmetic reasoning as a representative task. Using causal analysis, we identify a subset of the model (a circuit) that explains most of the model's behavior for basic arithmetic logic and examine its functionality. By zooming in on the level of individual circuit neurons, we discover a sparse set of important neurons that implement simple heuristics. Each heuristic identifies a numerical input pattern and outputs corresponding answers. We hypothesize that the combination of these heuristic neurons is the mechanism used to produce correct arithmetic answers. To test this, we categorize each neuron into several heuristic types---such as neurons that activate when an operand falls within a certain range---and find that the unordered combination of these heuristic types is the mechanism that explains most of the model's accuracy on arithmetic prompts. Finally, we demonstrate that this mechanism appears as the main source of arithmetic accuracy early in training. Overall, our experimental results across several LLMs show that LLMs perform arithmetic using neither robust algorithms nor memorization; rather, they rely on a ``bag of heuristics''.

How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of lasting evaluation standards, we propose MIB, a Mechanistic Interpretability Benchmark, with two tracks spanning four tasks and five models. MIB favors methods that precisely and concisely recover relevant causal pathways or causal variables in neural language models. The circuit localization track compares methods that locate the model components---and connections between them---most important for performing a task (e.g., attribution patching or information flow routes). The causal variable track compares methods that featurize a hidden vector, e.g., sparse autoencoders (SAE) or distributed alignment search (DAS), and align those features to a task-relevant causal variable. Using MIB, we find that attribution and mask optimization methods perform best on circuit localization. For causal variable localization, we find that the supervised DAS method performs best, while SAEs features are not better than neurons, i.e., non-featurized hidden vectors. These findings illustrate that MIB enables meaningful comparisons, and increases our confidence that there has been real progress in the field.

Generative Information Retrieval (GenIR) is a novel paradigm in which a transformer encoder-decoder model predicts document rankings based on a query in an end-to-end fashion. These GenIR models have received significant attention due to their simple retrieval architecture while maintaining high retrieval effectiveness. However, in contrast to established retrieval architectures like cross-encoders or bi-encoders, their internal computations remain largely unknown. In this work, we investigate this retrieval mechanism and uncover the roles played by different model components (self-attention, cross-attention, MLPs) and their interaction to generate the document identifier. First, we show that the pre-trained encoder, which was not fine-tuned for retrieval, is sufficient for the retrieval process. Then, we find that the pass through the decoder can be divided into three stages: (I) the priming stage in which no component contributes query-specific information (II) the bridging stage where cross-attention transfers query information from the encoder to the decoder, and (III) the interaction stage where MLPs process this transferred information to predict the document identifier in the last layer. Our results indicate that document-specific information is only stored in a few components in the final stage of the retrieval process. We hope that our findings will motivate the development of more effective GenIR models and facilitate their improvements.

Mathematical formulas are an important tool to concisely communicate ideas in science and education, used to clarify descriptions, calculations or derivations. When searching in scientific literature, mathematical notation, which is often written using the LATEX notation, therefore plays a crucial role that should not be neglected. The task of mathematics-aware information retrieval is to retrieve relevant passages given a query or question, which both can include natural language and mathematical formulas. As in many domains that rely on Natural Language Understanding, transformer-based models are now dominating the field of information retrieval [3]. Apart from their size and the transformerencoder architecture, pre-training is considered to be a key factor for the high performance of these models. It has also been shown that domain-adaptive pre-training improves their performance on down-stream tasks even further [2] especially when the vocabulary overlap between pre-training and in-domain data is low. This is also the case for the domain of mathematical documents.

Mathematical Language Processing (MLP) deals with the automated processing and analysis of mathematical documents and relies heavily on good representations of mathematical symbols and texts. The aim of this work is to explore the modeling capabilities of state-of-the-art unsupervised deep learning methods to create such representations. Therefore, we pre-trained different instances of an ALBERT model on Mathematics StackExchange data and fine-tuned it on the task of Mathematical Answer Retrieval. Our evaluation shows that ALBERT outperforms all previous systems and is on par with current state-of-the-art systems for math retrieval indicating strong capabilities of modeling mathematical posts. This implies that our approach can also be beneficial to various other tasks in MLP such as automatic proof checking or summarization of scientific texts.