Multimodal Large Language Models struggle with tabular reasoning on tables containing images as cells owing to their limited memory and the distractions posed by the surrounding irrelevant visual elements. Existing context-reduction methods reduce context size but degrade reasoning accuracy: text-only pruning neglects visual cells, token-level pruning discards valuable information, and retrieval cannot understand table structure. We propose CAPTR: Caption-based Context Pruning for Tabular Reasoning, a two-stage framework that uses image captions as a lightweight textual proxy to prune non-pertinent rows and columns and then restores images for reasoning. CAPTR can reduce the context length by 80% while achieving up to 8% accuracy improvement against competitive baselines. By analyzing the error cases, we find that entity confusion contributes to the majority of the remaining errors. CAPTR is the first multimodal context-reduction method to improve performance on multimodal tabular reasoning.

Generalization and robustness to input variation are core desiderata of machine learning research. Language varies along several axes, most importantly, language instance (e.g. French) and domain (e.g. news). While adapting NLP models to new languages within a single domain, or to new domains within a single language, is widely studied, research in joint adaptation is hampered by the lack of evaluation datasets. This prevents the transfer of NLP systems from well-resourced languages and domains to non-dominant language-domain combinations. To address this gap, we introduce M2QA, a multi-domain multilingual question answering benchmark.M2QA includes 13,500 SQuAD 2.0-style question-answer instances in German, Turkish, and Chinese for the domains of product reviews, news, and creative writing. We use M2QA to explore cross-lingual cross-domain performance of fine-tuned models and state-of-the-art LLMs and investigate modular approaches to domain and language adaptation.We witness 1) considerable performance variations across domain-language combinations within model classes and 2) considerable performance drops between source and target language-domain combinations across all model sizes. We demonstrate that M2QA is far from solved, and new methods to effectively transfer both linguistic and domain-specific information are necessary.

We introduce Adapters, an open-source library that unifies parameter-efficient and modular transfer learning in large language models. By integrating 10 diverse adapter methods into a unified interface, Adapters offers ease of use and flexible configuration. Our library allows researchers and practitioners to leverage adapter modularity through composition blocks, enabling the design of complex adapter setups. We demonstrate the library’s efficacy by evaluating its performance against full fine-tuning on various NLP tasks. Adapters provides a powerful tool for addressing the challenges of conventional fine-tuning paradigms and promoting more efficient and modular transfer learning. The library is available via https://adapterhub.ml/adapters.