Task decomposition is a fundamental mechanism in program synthesis, enabling complex problems to be broken down into manageable subtasks. ExeDec, a state-of-the-art program synthesis framework, employs this approach by combining a Subgoal Model for decomposition and a Synthesizer Model for program generation to facilitate compositional generalization. In this work, we develop REGISM, an adaptation of ExeDec that removes decomposition guidance and relies solely on iterative execution-driven synthesis. By comparing these two exemplary approaches-ExeDec, which leverages task decomposition, and REGISM, which does not-we investigate the interplay between task decomposition and program generation. Our findings indicate that ExeDec exhibits significant advantages in length generalization and concept composition tasks, likely due to its explicit decomposition strategies. At the same time, REGISM frequently matches or surpasses ExeDec's performance across various scenarios, with its solutions often aligning more closely with ground truth decompositions. These observations highlight the importance of repeated execution-guided synthesis in driving task-solving performance, even within frameworks that incorporate explicit decomposition strategies. Our analysis suggests that task decomposition approaches like ExeDec hold significant potential for advancing program synthesis, though further work is needed to clarify when and why these strategies are most effective.

Data repositories have accumulated a large number of tabular datasets from various domains. Machine Learning researchers are actively using these datasets to evaluate novel approaches. Consequently, data repositories have an important standing in tabular data research. They not only host datasets but also provide information on how to use them in supervised learning tasks. In this paper, we argue that, despite great achievements in usability, the unreflected usage of datasets from data repositories may have led to reduced research quality and scientific rigor. We present examples from prominent recent studies that illustrate the problematic use of datasets from OpenML, a large data repository for tabular data. Our illustrations help users of data repositories avoid falling into the traps of (1) using suboptimal model selection strategies, (2) overlooking strong baselines, and (3) inappropriate preprocessing. In response, we discuss possible solutions for how data repositories can prevent the inappropriate use of datasets and become the cornerstones for improved overall quality of empirical research studies. In tabular data research, the OpenML repository is used extensively (Gijsbers et al., 2019; Salinas & Erickson, 2024; Liu et al., 2024; Hollmann et al., 2025). A driving factor for tabular data repository usage is the recent increase in efforts to transfer the success of deep learning to the tabular domain. The development of novel neural network models (Arik & Pfister, 2021; Chang et al., 2021; Gorishniy et al., 2021; 2023; 2024), and more recently tabular foundation models (Gardner et al., 2024; Hollmann et al., 2025) dominates the tabular machine learning community. In response, recent comparative studies try to gather as many datasets as possible to facilitate a rigorous and comprehensive evaluation of novel approaches (Grinsztajn et al., 2022; McElfresh et al., 2023; Ye et al., 2024a). While McElfresh et al. (2023) used 196 datasets, a recent study scales up to 300 datasets from OpenML (Ye et al., 2024a). Similarly, studies evaluating foundation models seem to include as many datasets from these benchmarks as possible, apparently taking their quality and appropriateness for granted (Yan et al., 2024; Gardner et al., 2024). Different authors have recently criticized the intense focus on model development and the limited attention to data quality. Existing benchmarks often use outdated data (Kohli et al., 2024), ignore task-specific preprocessing (Tschalzev et al., 2024), or use inappropriate data splits (Rubachev et al., 2024).

Exploration is a crucial skill for self-improvement and open-ended problemsolving. However, it remains uncertain whether large language models can effectively explore the state-space. Existing evaluations predominantly focus on the trade-off between exploration and exploitation, often assessed in multi-armed bandit problems. In contrast, this work isolates exploration as the sole objective, tasking the agent with delivering information that enhances future returns. For the evaluation, we propose to decompose missing rewards into exploration and exploitation components by measuring the optimal achievable return for the states already explored. Our experiments with various LLMs reveal that most models struggle to sufficiently explore the state-space and that weak exploration is insufficient. We observe a positive correlation between model size and exploration performance, with larger models demonstrating superior capabilities. Furthermore, we show that our decomposition provides insights into differences in behaviors driven by agent instructions during prompt engineering, offering a valuable tool for refining LLM performance in exploratory tasks. Recently, large language models (LLMs) have demonstrated promising results in various decision making tasks such as web browsing (Yao et al., 2022; Shinn et al., 2024; Ma et al., 2023), game-playing (Paglieri et al., 2024), and tasks in simulated households (Yao et al., 2022; Shinn et al., 2024). This way, LLMs act as agents that observe states and take actions in different environments. Through their vast internal knowledge-base and autoregressive in-context reasoning capabilities, the models are supposed to quickly adapt to new tasks. However, previous work has shown that LLMs struggle with solving increasingly complex environments due to several limitations: For example, the ability to learn from mistakes is often limited (Huang et al., 2023) and LLMs have difficulties with planning over long horizons (Kambhampati et al., 2024). The examples emphasize that understanding LLM abilities is essential for their risk assessment in real life applications, and future development.

Human bilinguals often use similar brain regions to process multiple languages, depending on when they learned their second language and their proficiency. In large language models (LLMs), how are multiple languages learned and encoded? In this work, we explore the extent to which LLMs share representations of morphosyntactic concepts such as grammatical number, gender, and tense across languages. We train sparse autoencoders on Llama-3-8B and Aya-23-8B, and demonstrate that abstract grammatical concepts are often encoded in feature directions shared across many languages. We use causal interventions to verify the multilingual nature of these representations; specifically, we show that ablating only multilingual features decreases classifier performance to near-chance across languages. We then use these features to precisely modify model behavior in a machine translation task; this demonstrates both the generality and selectivity of these feature's roles in the network. Our findings suggest that even models trained predominantly on English data can develop robust, cross-lingual abstractions of morphosyntactic concepts.

In this technical report, we investigate the predictive performance differences of a rule-based approach and the GNN architectures NBFNet and A*Net with respect to knowledge graph completion. For the two most common benchmarks, we find that a substantial fraction of the performance difference can be explained by one unique negative pattern on each dataset that is hidden from the rule-based approach. Our findings add a unique perspective on the performance difference of different model classes for knowledge graph completion: Models can achieve a predictive performance advantage by penalizing scores of incorrect facts opposed to providing high scores for correct facts.

Tabular data is prevalent in real-world machine learning applications, and new models for supervised learning of tabular data are frequently proposed. Comparative studies assessing the performance of models typically consist of model-centric evaluation setups with overly standardized data preprocessing. This paper demonstrates that such model-centric evaluations are biased, as real-world modeling pipelines often require dataset-specific preprocessing and feature engineering. Therefore, we propose a data-centric evaluation framework. We select 10 relevant datasets from Kaggle competitions and implement expert-level preprocessing pipelines for each dataset. We conduct experiments with different preprocessing pipelines and hyperparameter optimization (HPO) regimes to quantify the impact of model selection, HPO, feature engineering, and test-time adaptation. Our main findings are: 1. After dataset-specific feature engineering, model rankings change considerably, performance differences decrease, and the importance of model selection reduces. 2. Recent models, despite their measurable progress, still significantly benefit from manual feature engineering. This holds true for both tree-based models and neural networks. 3. While tabular data is typically considered static, samples are often collected over time, and adapting to distribution shifts can be important even in supposedly static data. These insights suggest that research efforts should be directed toward a data-centric perspective, acknowledging that tabular data requires feature engineering and often exhibits temporal characteristics.

Neural networks often assume independence among input data samples, disregarding correlations arising from inherent clustering patterns in real-world datasets (e.g., due to different sites or repeated measurements). Recently, mixed effects neural networks (MENNs) which separate cluster-specific 'random effects' from cluster-invariant 'fixed effects' have been proposed to improve generalization and interpretability for clustered data. However, existing methods only allow for approximate quantification of cluster effects and are limited to regression and binary targets with only one clustering feature. We present MC-GMENN, a novel approach employing Monte Carlo methods to train Generalized Mixed Effects Neural Networks. We empirically demonstrate that MC-GMENN outperforms existing mixed effects deep learning models in terms of generalization performance, time complexity, and quantification of inter-cluster variance. Additionally, MC-GMENN is applicable to a wide range of datasets, including multi-class classification tasks with multiple high-cardinality categorical features. For these datasets, we show that MC-GMENN outperforms conventional encoding and embedding methods, simultaneously offering a principled methodology for interpreting the effects of clustering patterns.

Transformers demonstrate impressive performance on a range of reasoning benchmarks. To evaluate the degree to which these abilities are a result of actual reasoning, existing work has focused on developing sophisticated benchmarks for behavioral studies. However, these studies do not provide insights into the internal mechanisms driving the observed capabilities. To improve our understanding of the internal mechanisms of transformers, we present a comprehensive mechanistic analysis of a transformer trained on a synthetic reasoning task. We identify a set of interpretable mechanisms the model uses to solve the task, and validate our findings using correlational and causal evidence. Our results suggest that it implements a depth-bounded recurrent mechanisms that operates in parallel and stores intermediate results in selected token positions. We anticipate that the motifs we identified in our synthetic setting can provide valuable insights into the broader operating principles of transformers and thus provide a basis for understanding more complex models.

Explainable Artificial Intelligence is critical in unraveling decision-making processes in complex machine learning models. LIME (Local Interpretable Model-agnostic Explanations) is a well-known XAI framework for image analysis. It utilizes image segmentation to create features to identify relevant areas for classification. Consequently, poor segmentation can compromise the consistency of the explanation and undermine the importance of the segments, affecting the overall interpretability. Addressing these challenges, we introduce DSEG-LIME (Data-Driven Segmentation LIME), featuring: i) a data-driven segmentation for human-recognized feature generation, and ii) a hierarchical segmentation procedure through composition. We benchmark DSEG-LIME on pre-trained models with images from the ImageNet dataset - scenarios without domain-specific knowledge. The analysis includes a quantitative evaluation using established XAI metrics, complemented by a qualitative assessment through a user study. Our findings demonstrate that DSEG outperforms in most of the XAI metrics and enhances the alignment of explanations with human-recognized concepts, significantly improving interpretability. The code is available under: https://github. com/patrick-knab/DSEG-LIME

Goal recognition is an important problem in many application domains (e.g., pervasive computing, intrusion detection, computer games, etc.). In many application scenarios, it is important that goal recognition algorithms can recognize goals of an observed agent as fast as possible. However, many early approaches in the area of Plan Recognition As Planning, require quite large amounts of computation time to calculate a solution. Mainly to address this issue, recently, Pereira et al. [11] developed an approach that is based on planning landmarks and is much more computationally efficient than previous approaches. However, the approach, as proposed by Pereira et al., also uses trivial landmarks (i.e., facts that are part of the initial state and goal description are landmarks by definition). In this paper, we show that it does not provide any benefit to use landmarks that are part of the initial state in a planning landmark based goal recognition approach. The empirical results show that omitting initial state landmarks for goal recognition improves goal recognition performance.