Filip Ilievski

Peifeng Wang, Aaron Chan, Filip Ilievski et al. · meta-ai

Neural language models (LMs) have achieved impressive results on various language-based reasoning tasks by utilizing latent knowledge encoded in their own pretrained parameters. To make this reasoning process more explicit, recent works retrieve a rationalizing LM's internal knowledge by training or prompting it to generate free-text rationales, which can be used to guide task predictions made by either the same LM or a separate reasoning LM. However, rationalizing LMs require expensive rationale annotation and/or computation, without any assurance that their generated rationales improve LM task performance or faithfully reflect LM decision-making. In this paper, we propose PINTO, an LM pipeline that rationalizes via prompt-based learning, and learns to faithfully reason over rationales via counterfactual regularization. First, PINTO maps out a suitable reasoning process for the task input by prompting a frozen rationalizing LM to generate a free-text rationale. Second, PINTO's reasoning LM is fine-tuned to solve the task using the generated rationale as context, while regularized to output less confident predictions when the rationale is perturbed. Across four datasets, we show that PINTO significantly improves the generalization ability of the reasoning LM, yielding higher performance on both in-distribution and out-of-distribution test sets. Also, we find that PINTO's rationales are more faithful to its task predictions than those generated by competitive baselines.

Kaixin Ma, Filip Ilievski, Jonathan Francis et al. · cmu

Procedural text understanding is a challenging language reasoning task that requires models to track entity states across the development of a narrative. A complete procedural understanding solution should combine three core aspects: local and global views of the inputs, and global view of outputs. Prior methods considered a subset of these aspects, resulting in either low precision or low recall. In this paper, we propose Coalescing Global and Local Information (CGLI), a new model that builds entity- and timestep-aware input representations (local input) considering the whole context (global input), and we jointly model the entity states with a structured prediction objective (global output). Thus, CGLI simultaneously optimizes for both precision and recall. We extend CGLI with additional output layers and integrate it into a story reasoning framework. Extensive experiments on a popular procedural text understanding dataset show that our model achieves state-of-the-art results; experiments on a story reasoning benchmark show the positive impact of our model on downstream reasoning.

Jiarui Zhang, Filip Ilievski, Kaixin Ma et al. · cmu

Intelligent Traffic Monitoring (ITMo) technologies hold the potential for improving road safety/security and for enabling smart city infrastructure. Understanding traffic situations requires a complex fusion of perceptual information with domain-specific and causal commonsense knowledge. Whereas prior work has provided benchmarks and methods for traffic monitoring, it remains unclear whether models can effectively align these information sources and reason in novel scenarios. To address this assessment gap, we devise three novel text-based tasks for situational reasoning in the traffic domain: i) BDD-QA, which evaluates the ability of Language Models (LMs) to perform situational decision-making, ii) TV-QA, which assesses LMs' abilities to reason about complex event causality, and iii) HDT-QA, which evaluates the ability of models to solve human driving exams. We adopt four knowledge-enhanced methods that have shown generalization capability across language reasoning tasks in prior work, based on natural language inference, commonsense knowledge-graph self-supervision, multi-QA joint training, and dense retrieval of domain information. We associate each method with a relevant knowledge source, including knowledge graphs, relevant benchmarks, and driving manuals. In extensive experiments, we benchmark various knowledge-aware methods against the three datasets, under zero-shot evaluation; we provide in-depth analyses of model performance on data partitions and examine model predictions categorically, to yield useful insights on traffic understanding, given different background knowledge and reasoning strategies.

Jiarui Zhang, Filip Ilievski, Kaixin Ma et al. · cmu

Self-supervision based on the information extracted from large knowledge graphs has been shown to improve the generalization of language models, in zero-shot evaluation on various downstream language reasoning tasks. Since these improvements are reported in aggregate, however, little is known about (i) how to select the appropriate knowledge for solid performance across tasks, (ii) how to combine this knowledge with neural language models, and (iii) how these pairings affect granular task performance. In this paper, we study the effect of knowledge sampling strategies and sizes that can be used to generate synthetic data for adapting language models. We study the effect of different synthetic datasets on language models with various architectures and sizes. The resulting models are evaluated against four task properties: domain overlap, answer similarity, vocabulary overlap, and answer length. Our experiments show that encoder-decoder models benefit from more data to learn from, whereas sampling strategies that balance across different aspects yield best performance. Most of the improvement occurs on questions with short answers and dissimilar answer candidates, which corresponds to the characteristics of the data used for pre-training.

Zhivar Sourati, Vishnu Priya Prasanna Venkatesh, Darshan Deshpande et al.

The spread of misinformation, propaganda, and flawed argumentation has been amplified in the Internet era. Given the volume of data and the subtlety of identifying violations of argumentation norms, supporting information analytics tasks, like content moderation, with trustworthy methods that can identify logical fallacies is essential. In this paper, we formalize prior theoretical work on logical fallacies into a comprehensive three-stage evaluation framework of detection, coarse-grained, and fine-grained classification. We adapt existing evaluation datasets for each stage of the evaluation. We employ three families of robust and explainable methods based on prototype reasoning, instance-based reasoning, and knowledge injection. The methods combine language models with background knowledge and explainable mechanisms. Moreover, we address data sparsity with strategies for data augmentation and curriculum learning. Our three-stage framework natively consolidates prior datasets and methods from existing tasks, like propaganda detection, serving as an overarching evaluation testbed. We extensively evaluate these methods on our datasets, focusing on their robustness and explainability. Our results provide insight into the strengths and weaknesses of the methods on different components and fallacy classes, indicating that fallacy identification is a challenging task that may require specialized forms of reasoning to capture various classes. We share our open-source code and data on GitHub to support further work on logical fallacy identification.

Jiarui Zhang, Filip Ilievski, Aravinda Kollaa et al. · cmu

Understanding novel situations in the traffic domain requires an intricate combination of domain-specific and causal commonsense knowledge. Prior work has provided sufficient perception-based modalities for traffic monitoring, in this paper, we focus on a complementary research aspect of Intelligent Transportation: traffic understanding. We scope our study to text-based methods and datasets given the abundant commonsense knowledge that can be extracted using language models from large corpus and knowledge graphs. We adopt three knowledge-driven approaches for zero-shot QA over traffic situations, based on prior natural language inference methods, commonsense models with knowledge graph self-supervision, and dense retriever-based models. We constructed two text-based multiple-choice question answering sets: BDD-QA for evaluating causal reasoning in the traffic domain and HDT-QA for measuring the possession of domain knowledge akin to human driving license tests. Among the methods, Unified-QA reaches the best performance on the BDD-QA dataset with the adaptation of multiple formats of question answers. Language models trained with inference information and commonsense knowledge are also good at predicting the cause and effect in the traffic domain but perform badly at answering human-driving QA sets. For such sets, DPR+Unified-QA performs the best due to its efficient knowledge extraction.

Prateek Chhikara, Dhiraj Chaurasia, Yifan Jiang et al.

Food computing has emerged as a prominent multidisciplinary field of research in recent years. An ambitious goal of food computing is to develop end-to-end intelligent systems capable of autonomously producing recipe information for a food image. Current image-to-recipe methods are retrieval-based and their success depends heavily on the dataset size and diversity, as well as the quality of learned embeddings. Meanwhile, the emergence of powerful attention-based vision and language models presents a promising avenue for accurate and generalizable recipe generation, which has yet to be extensively explored. This paper proposes FIRE, a novel multimodal methodology tailored to recipe generation in the food computing domain, which generates the food title, ingredients, and cooking instructions based on input food images. FIRE leverages the BLIP model to generate titles, utilizes a Vision Transformer with a decoder for ingredient extraction, and employs the T5 model to generate recipes incorporating titles and ingredients as inputs. We showcase two practical applications that can benefit from integrating FIRE with large language model prompting: recipe customization to fit recipes to user preferences and recipe-to-code transformation to enable automated cooking processes. Our experimental findings validate the efficacy of our proposed approach, underscoring its potential for future advancements and widespread adoption in food computing.

Thiloshon Nagarajah, Filip Ilievski, Jay Pujara

Analogical reasoning is a powerful qualitative reasoning tool that enables humans to connect two situations, and to generalize their knowledge from familiar to novel situations. Cognitive Science research provides valuable insights into the richness and complexity of analogical reasoning, together with implementations of expressive analogical reasoners with limited scalability. Modern scalable AI techniques with the potential to reason by analogy have been only applied to the special case of proportional analogy, and not to understanding higher-order analogies. In this paper, we aim to bridge the gap by: 1) formalizing six dimensions of analogy based on mature insights from Cognitive Science research, 2) annotating a corpus of fables with each of these dimensions, and 3) defining four tasks with increasing complexity that enable scalable evaluation of AI techniques. Experiments with language models and neuro-symbolic AI reasoners on these tasks reveal that state-of-the-art methods can be applied to reason by analogy with a limited success, motivating the need for further research towards comprehensive and scalable analogical reasoning by AI. We make all our code and data available.

Jiang Wang, Filip Ilievski, Pedro Szekely et al.

Embedding methods have demonstrated robust performance on the task of link prediction in knowledge graphs, by mostly encoding entity relationships. Recent methods propose to enhance the loss function with a literal-aware term. In this paper, we propose KGA: a knowledge graph augmentation method that incorporates literals in an embedding model without modifying its loss function. KGA discretizes quantity and year values into bins, and chains these bins both horizontally, modeling neighboring values, and vertically, modeling multiple levels of granularity. KGA is scalable and can be used as a pre-processing step for any existing knowledge graph embedding model. Experiments on legacy benchmarks and a new large benchmark, DWD, show that augmenting the knowledge graph with quantities and years is beneficial for predicting both entities and numbers, as KGA outperforms the vanilla models and other relevant baselines. Our ablation studies confirm that both quantities and years contribute to KGA's performance, and that its performance depends on the discretization and binning settings. We make the code, models, and the DWD benchmark publicly available to facilitate reproducibility and future research.

Yifan Jiang, Filip Ilievski, Kaixin Ma et al.

The success of language models has inspired the NLP community to attend to tasks that require implicit and complex reasoning, relying on human-like commonsense mechanisms. While such vertical thinking tasks have been relatively popular, lateral thinking puzzles have received little attention. To bridge this gap, we devise BRAINTEASER: a multiple-choice Question Answering task designed to test the model's ability to exhibit lateral thinking and defy default commonsense associations. We design a three-step procedure for creating the first lateral thinking benchmark, consisting of data collection, distractor generation, and generation of adversarial examples, leading to 1,100 puzzles with high-quality annotations. To assess the consistency of lateral reasoning by models, we enrich BRAINTEASER based on a semantic and contextual reconstruction of its questions. Our experiments with state-of-the-art instruction- and commonsense language models reveal a significant gap between human and model performance, which is further widened when consistency across adversarial formats is considered. We make all of our code and data available to stimulate work on developing and evaluating lateral thinking models.

Filip Ilievski, Jay Pujara, Kartik Shenoy

Analogical reasoning methods have been built over various resources, including commonsense knowledge bases, lexical resources, language models, or their combination. While the wide coverage of knowledge about entities and events make Wikidata a promising resource for analogical reasoning across situations and domains, Wikidata has not been employed for this task yet. In this paper, we investigate whether the knowledge in Wikidata supports analogical reasoning. Specifically, we study whether relational knowledge is modeled consistently in Wikidata, observing that relevant relational information is typically missing or modeled in an inconsistent way. Our further experiments show that Wikidata can be used to create data for analogy classification, but this requires much manual effort. To facilitate future work that can support analogies, we discuss key desiderata, and devise a set of metrics to guide an automatic method for extracting analogies from Wikidata.

Koen Kraaijveld, Yifan Jiang, Kaixin Ma et al.

While visual question-answering (VQA) benchmarks have catalyzed the development of reasoning techniques, they have focused on vertical thinking. Effective problem-solving also necessitates lateral thinking, which remains understudied in AI and has not been used to test visual perception systems. To bridge this gap, we formulate visual lateral thinking as a multiple-choice question-answering task and describe a three-step taxonomy-driven methodology for instantiating task examples. Then, we develop COLUMBUS, a synthetic benchmark that applies the task pipeline to create QA sets with text and icon rebus puzzles based on publicly available collections of compounds and common phrases. COLUMBUS comprises over 1,000 puzzles, each with four answer candidates. While the SotA vision-language models (VLMs) achieve decent performance, our evaluation demonstrates a substantial gap between humans and models. VLMs benefit from human-curated descriptions but struggle to self-generate such representations at the right level of abstraction.

Zhivar Sourati, Filip Ilievski, Hông-Ân Sandlin et al.

The ease and speed of spreading misinformation and propaganda on the Web motivate the need to develop trustworthy technology for detecting fallacies in natural language arguments. However, state-of-the-art language modeling methods exhibit a lack of robustness on tasks like logical fallacy classification that require complex reasoning. In this paper, we propose a Case-Based Reasoning method that classifies new cases of logical fallacy by language-modeling-driven retrieval and adaptation of historical cases. We design four complementary strategies to enrich input representation for our model, based on external information about goals, explanations, counterarguments, and argument structure. Our experiments in in-domain and out-of-domain settings indicate that Case-Based Reasoning improves the accuracy and generalizability of language models. Our ablation studies suggest that representations of similar cases have a strong impact on the model performance, that models perform well with fewer retrieved cases, and that the size of the case database has a negligible effect on the performance. Finally, we dive deeper into the relationship between the properties of the retrieved cases and the model performance.

Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara et al.

Multimodal Large Language Models (MLLMs) have recently achieved promising zero-shot accuracy on visual question answering (VQA) -- a fundamental task affecting various downstream applications and domains. Given the great potential for the broad use of these models, it is important to investigate their limitations in dealing with different image and question properties. In this work, we investigate whether MLLMs can perceive small details as well as large details in images. In particular, we show that their zero-shot accuracy in answering visual questions is very sensitive to the size of the visual subject of the question, declining up to 46% with size. Furthermore, we show that this effect is causal by observing that human visual cropping can significantly mitigate their sensitivity to size. Inspired by the usefulness of human cropping, we then propose five automatic visual cropping methods -- leveraging either external localization models or the decision process of the given MLLM itself -- as inference time mechanisms to improve the zero-shot performance of MLLMs. We study their effectiveness on four popular VQA datasets, and a subset of the VQAv2 dataset tailored towards fine visual details. Our findings suggest that MLLMs should be used with caution in detail-sensitive VQA applications, and that visual cropping is a promising direction to improve their zero-shot performance. To facilitate further investigation of MLLMs' behaviors, our code and data are publicly released.

Zhivar Sourati, Darshan Deshpande, Filip Ilievski et al.

Downstream applications often require text classification models to be accurate and robust. While the accuracy of the state-of-the-art Language Models (LMs) approximates human performance, they often exhibit a drop in performance on noisy data found in the real world. This lack of robustness can be concerning, as even small perturbations in the text, irrelevant to the target task, can cause classifiers to incorrectly change their predictions. A potential solution can be the family of Prototype-Based Networks (PBNs) that classifies examples based on their similarity to prototypical examples of a class (prototypes) and has been shown to be robust to noise for computer vision tasks. In this paper, we study whether the robustness properties of PBNs transfer to text classification tasks under both targeted and static adversarial attack settings. Our results show that PBNs, as a mere architectural variation of vanilla LMs, offer more robustness compared to vanilla LMs under both targeted and static settings. We showcase how PBNs' interpretability can help us to understand PBNs' robustness properties. Finally, our ablation studies reveal the sensitivity of PBNs' robustness to how strictly clustering is done in the training phase, as tighter clustering results in less robust PBNs.

Yifan Jiang, Filip Ilievski, Kaixin Ma

Stories about everyday situations are an essential part of human communication, motivating the need to develop AI agents that can reliably understand these stories. Despite the long list of supervised methods for story completion and procedural understanding, current AI has no mechanisms to automatically track and explain procedures in unseen stories. To bridge this gap, we study the ability of AI models to transfer procedural knowledge to novel narrative tasks in a transparent manner. We design LEAP: a comprehensive framework that integrates state-of-the-art modeling architectures, training regimes, and augmentation strategies based on both natural and synthetic stories. To address the lack of densely annotated training data, we devise a robust automatic labeler based on few-shot prompting to enhance the augmented data. Our experiments with in- and out-of-domain tasks reveal insights into the interplay of different architectures, training regimes, and augmentation strategies. LEAP's labeler has a clear positive impact on out-of-domain datasets, while the resulting dense annotation provides native explainability.

Abhinav Kumar Thakur, Filip Ilievski, Hông-Ân Sandlin et al.

In the current context where online platforms have been effectively weaponized in a variety of geo-political events and social issues, Internet memes make fair content moderation at scale even more difficult. Existing work on meme classification and tracking has focused on black-box methods that do not explicitly consider the semantics of the memes or the context of their creation. In this paper, we pursue a modular and explainable architecture for Internet meme understanding. We design and implement multimodal classification methods that perform example- and prototype-based reasoning over training cases, while leveraging both textual and visual SOTA models to represent the individual cases. We study the relevance of our modular and explainable models in detecting harmful memes on two existing tasks: Hate Speech Detection and Misogyny Classification. We compare the performance between example- and prototype-based methods, and between text, vision, and multimodal models, across different categories of harmfulness (e.g., stereotype and objectification). We devise a user-friendly interface that facilitates the comparative analysis of examples retrieved by all of our models for any given meme, informing the community about the strengths and limitations of these explainable methods.

Zhivar Sourati, Filip Ilievski, Pia Sommerauer et al.

As a core cognitive skill that enables the transferability of information across domains, analogical reasoning has been extensively studied for both humans and computational models. However, while cognitive theories of analogy often focus on narratives and study the distinction between surface, relational, and system similarities, existing work in natural language processing has a narrower focus as far as relational analogies between word pairs. This gap brings a natural question: can state-of-the-art large language models (LLMs) detect system analogies between narratives? To gain insight into this question and extend word-based relational analogies to relational system analogies, we devise a comprehensive computational framework that operationalizes dominant theories of analogy, using narrative elements to create surface and system mappings. Leveraging the interplay between these mappings, we create a binary task and benchmark for Analogical Reasoning on Narratives (ARN), covering four categories of far (cross-domain)/near (within-domain) analogies and disanalogies. We show that while all LLMs can largely recognize near analogies, even the largest ones struggle with far analogies in a zero-shot setting, with GPT4.0 scoring below random. Guiding the models through solved examples and chain-of-thought reasoning enhances their analogical reasoning ability. Yet, since even in the few-shot setting, the best model only performs halfway between random and humans, ARN opens exciting directions for computational analogical reasoners.

Bohui Zhang, Filip Ilievski, Pedro Szekely

Large public knowledge graphs, like Wikidata, contain billions of statements about tens of millions of entities, thus inspiring various use cases to exploit such knowledge graphs. However, practice shows that much of the relevant information that fits users' needs is still missing in Wikidata, while current linked open data (LOD) tools are not suitable to enrich large graphs like Wikidata. In this paper, we investigate the potential of enriching Wikidata with structured data sources from the LOD cloud. We present a novel workflow that includes gap detection, source selection, schema alignment, and semantic validation. We evaluate our enrichment method with two complementary LOD sources: a noisy source with broad coverage, DBpedia, and a manually curated source with a narrow focus on the art domain, Getty. Our experiments show that our workflow can enrich Wikidata with millions of novel statements from external LOD sources with high quality. Property alignment and data quality are key challenges, whereas entity alignment and source selection are well-supported by existing Wikidata mechanisms. We make our code and data available to support future work.

Bradley P. Allen, Filip Ilievski, Saurav Joshi

Knowledge engineering is the process of creating and maintaining knowledge-producing systems. Throughout the history of computer science and AI, knowledge engineering workflows have been widely used because high-quality knowledge is assumed to be crucial for reliable intelligent agents. However, the landscape of knowledge engineering has changed, presenting four challenges: unaddressed stakeholder requirements, mismatched technologies, adoption barriers for new organizations, and misalignment with software engineering practices. In this paper, we propose to address these challenges by developing a reference architecture using a mainstream software methodology. By studying the requirements of different stakeholders and eras, we identify 23 essential quality attributes for evaluating reference architectures. We assess three candidate architectures from recent literature based on these attributes. Finally, we discuss the next steps towards a comprehensive reference architecture, including prioritizing quality attributes, integrating components with complementary strengths, and supporting missing socio-technical requirements. As this endeavor requires a collaborative effort, we invite all knowledge engineering researchers and practitioners to join us.

Luca Vogelgesang, Ahmed Mehdi Soltani, Mohammadhossein Khojasteh et al.

Assistive robots have growing potential to support physical wellbeing in home and healthcare settings, for example, by guiding users through stretching or rehabilitation routines. However, existing systems remain largely scripted, which limits their ability to adapt to user state, environmental context, and interaction dynamics. In this work, we present StretchBot, a hybrid neuro-symbolic robotic coach for adaptive assistive guidance. The system combines multimodal perception with knowledge-graph-grounded large language model reasoning to support context-aware adjustments during short stretching sessions while maintaining a structured routine. To complement the system description, we report an exploratory pilot comparison between scripted and adaptive guidance with three participants. The pilot findings suggest that the adaptive condition improved perceived adaptability and contextual relevance, while scripted guidance remained competitive in smoothness and predictability. These results provide preliminary evidence that structured actionable knowledge can help ground language-model-based adaptation in embodied assistive interaction, while also highlighting the need for larger, longitudinal studies to evaluate robustness, generalizability, and long-term user experience.

Aravinda Kolla, Filip Ilievski, Hông-Ân Sandlin et al.

Considering the large amount of content created online by the minute, slang-aware automatic tools are critically needed to promote social good, and assist policymakers and moderators in restricting the spread of offensive language, abuse, and hate speech. Despite the success of large language models and the spontaneous emergence of slang dictionaries, it is unclear how far their combination goes in terms of slang understanding for downstream social good tasks. In this paper, we provide a framework to study different combinations of representation learning models and knowledge resources for a variety of downstream tasks that rely on slang understanding. Our experiments show the superiority of models that have been pre-trained on social media data, while the impact of dictionaries is positive only for static word embeddings. Our error analysis identifies core challenges for slang representation learning, including out-of-vocabulary words, polysemy, variance, and annotation disagreements, which can be traced to characteristics of slang as a quickly evolving and highly subjective language.

Jiarui Zhang, Jinyi Hu, Mahyar Khayatkhoei et al. · tsinghua

Multimodal Large Language Models (MLLMs) have recently shown remarkable perceptual capability in answering visual questions, however, little is known about the limits of their perception. In particular, while prior works have provided anecdotal evidence of MLLMs' sensitivity to object size, this phenomenon and its underlying causes have not been explored comprehensively. In this work, we quantitatively study the perception of small visual objects in several state-of-the-art MLLMs and reveal a pervasive limitation in answering questions about small objects in images. Next, we identify four independent factors that can contribute to this limitation -- object quality, size, distractors, and location -- and conduct controlled intervention studies to measure the effect of each factor on MLLMs' perception. In particular, we find that lower object quality and smaller object size can both independently reduce MLLMs' ability to answer visual questions. More surprisingly, we find that the location of the object in the image and the presence of visual distractors can also significantly reduce MLLMs' question answering accuracy. Our study provides a better understanding of the perceptual limitation of MLLMs and contributes new evaluation protocols for analyzing the perception of future MLLMs. To facilitate further investigations, we release our code and data.

Saurav Joshi, Filip Ilievski, Luca Luceri

Internet memes have emerged as a novel format for communication and expressing ideas on the web. Their fluidity and creative nature are reflected in their widespread use, often across platforms and occasionally for unethical or harmful purposes. While computational work has already analyzed their high-level virality over time and developed specialized classifiers for hate speech detection, there have been no efforts to date that aim to holistically track, identify, and map internet memes posted on social media. To bridge this gap, we investigate whether internet memes across social media platforms can be contextualized by using a semantic repository of knowledge, namely, a knowledge graph. We collect thousands of potential internet meme posts from two social media platforms, namely Reddit and Discord, and develop an extract-transform-load procedure to create a data lake with candidate meme posts. By using vision transformer-based similarity, we match these candidates against the memes cataloged in IMKG -- a recently released knowledge graph of internet memes. We leverage this grounding to highlight the potential of our proposed framework to study the prevalence of memes on different platforms, map them to IMKG, and provide context about memes on social media.

Abhishek Kolari, Mohammadhossein Khojasteh, Yifan Jiang et al.

While vision-language models (VLMs) have made remarkable progress on many popular visual question answering (VQA) benchmarks, it remains unclear whether they abstract and reason over depicted objects. Inspired by human object categorisation, object property reasoning involves identifying and recognising low-level details and higher-level abstractions. While current VQA benchmarks consider a limited set of object property attributes like size, they typically blend perception and reasoning, and lack representativeness in terms of reasoning and image categories. To this end, we introduce a systematic evaluation framework with images of three representative types, three reasoning levels of increasing complexity, and four object property dimensions driven by prior work on commonsense reasoning. We develop a procedure to instantiate this benchmark into ORBIT, a multi-level reasoning VQA benchmark for object properties comprising 360 images paired with a total of 1,080 count-based questions. Experiments with 12 state-of-the-art VLMs in zero-shot settings reveal significant limitations compared to humans, with the best-performing model only reaching 40\% accuracy. VLMs struggle particularly with realistic (photographic) images, counterfactual reasoning about physical and functional properties, and higher counts. ORBIT points to the need to develop methods for scalable benchmarking, generalize annotation guidelines, and explore additional reasoning VLMs. We make the ORBIT benchmark and the experimental code available to support such endeavors.

Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara et al.

Visual Question Answering is a challenging task, as it requires seamless interaction between perceptual, linguistic, and background knowledge systems. While the recent progress of visual and natural language models like BLIP has led to improved performance on this task, we lack understanding of the ability of such models to perform on different kinds of questions and reasoning types. As our initial analysis of BLIP-family models revealed difficulty with answering fine-detail questions, we investigate the following question: Can visual cropping be employed to improve the performance of state-of-the-art visual question answering models on fine-detail questions? Given the recent success of the BLIP-family models, we study a zero-shot and a fine-tuned BLIP model. We define three controlled subsets of the popular VQA-v2 benchmark to measure whether cropping can help model performance. Besides human cropping, we devise two automatic cropping strategies based on multi-modal embedding by CLIP and BLIP visual QA model gradients. Our experiments demonstrate that the performance of BLIP model variants can be significantly improved through human cropping, and automatic cropping methods can produce comparable benefits. A deeper dive into our findings indicates that the performance enhancement is more pronounced in zero-shot models than in fine-tuned models and more salient with smaller bounding boxes than larger ones. We perform case studies to connect quantitative differences with qualitative observations across question types and datasets. Finally, we see that the cropping enhancement is robust, as we gain an improvement of 4.59% (absolute) in the general VQA-random task by simply inputting a concatenation of the original and gradient-based cropped images. We make our code available to facilitate further innovation on visual cropping methods for question answering.

Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara et al.

Multimodal Large Language Models (MLLMs) have experienced rapid progress in visual recognition tasks in recent years. Given their potential integration into many critical applications, it is important to understand the limitations of their visual perception. In this work, we study whether MLLMs can perceive small visual details as effectively as large ones when answering questions about images. We observe that their performance is very sensitive to the size of the visual subject of the question, and further show that this effect is in fact causal by conducting an intervention study. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then propose training-free visual intervention methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to enhance its perception of small visual details. We evaluate our proposed methods on two widely-used MLLMs and seven visual question answering benchmarks and show that they can significantly improve MLLMs' accuracy without requiring any training. Our results elucidate the risk of applying MLLMs to visual recognition tasks concerning small details and indicate that visual intervention using the model's internal state is a promising direction to mitigate this risk.

Yifan Jiang, Filip Ilievski, Kaixin Ma

While vertical thinking relies on logical and commonsense reasoning, lateral thinking requires systems to defy commonsense associations and overwrite them through unconventional thinking. Lateral thinking has been shown to be challenging for current models but has received little attention. A recent benchmark, BRAINTEASER, aims to evaluate current models' lateral thinking ability in a zero-shot setting. In this paper, we split the original benchmark to also support fine-tuning setting and present SemEval Task 9: BRAIN-TEASER(S), the first task at this competition designed to test the system's reasoning and lateral thinking ability. As a popular task, BRAINTEASER(S)'s two subtasks receive 483 team submissions from 182 participants during the competition. This paper provides a fine-grained system analysis of the competition results, together with a reflection on what this means for the ability of the systems to reason laterally. We hope that the BRAINTEASER(S) subtasks and findings in this paper can stimulate future work on lateral thinking and robust reasoning by computational models.

Yifan Jiang, Jiarui Zhang, Kexuan Sun et al.

While multi-modal large language models (MLLMs) have shown significant progress on many popular visual reasoning benchmarks, whether they possess abstract visual reasoning abilities remains an open question. Similar to the Sudoku puzzles, abstract visual reasoning (AVR) problems require finding high-level patterns (e.g., repetition constraints) that control the input shapes (e.g., digits) in a specific task configuration (e.g., matrix). However, existing AVR benchmarks only considered a limited set of patterns (addition, conjunction), input shapes (rectangle, square), and task configurations (3 by 3 matrices). To evaluate MLLMs' reasoning abilities comprehensively, we introduce MARVEL, a multidimensional AVR benchmark with 770 puzzles composed of six core knowledge patterns, geometric and abstract shapes, and five different task configurations. To inspect whether the model accuracy is grounded in perception and reasoning, MARVEL complements the general AVR question with perception questions in a hierarchical evaluation framework. We conduct comprehensive experiments on MARVEL with nine representative MLLMs in zero-shot and few-shot settings. Our experiments reveal that all models show near-random performance on the AVR question, with significant performance gaps (40%) compared to humans across all patterns and task configurations. Further analysis of perception questions reveals that MLLMs struggle to comprehend the visual features (near-random performance) and even count the panels in the puzzle ( <45%), hindering their ability for abstract reasoning. We release our entire code and dataset.

Fabian Hoppe, Filip Ilievski, Jan-Christoph Kalo

Large Language Models (LLMs) have been shown to achieve impressive results for many reasoning-based NLP tasks, suggesting a degree of deductive reasoning capability. However, it remains unclear to which extent LLMs, in both informal and autoformalisation methods, are robust on logical deduction tasks. Moreover, while many LLM-based deduction methods have been proposed, a systematic study that analyses the impact of their design components is lacking. Addressing these two challenges, we propose the first study of the robustness of formal and informal LLM-based deductive reasoning methods. We devise a framework with two families of perturbations: adversarial noise and counterfactual statements, which jointly generate seven perturbed datasets. We organize the landscape of LLM reasoners according to their reasoning format, formalisation syntax, and feedback for error recovery. The results show that adversarial noise affects autoformalisation, while counterfactual statements influence all approaches. Detailed feedback does not improve overall accuracy despite reducing syntax errors, pointing to the challenge of LLM-based methods to self-correct effectively.

Huabin Liu, Filip Ilievski, Cees G. M. Snoek

This paper proposes the first video-grounded entailment tree reasoning method for commonsense video question answering (VQA). Despite the remarkable progress of large visual-language models (VLMs), there are growing concerns that they learn spurious correlations between videos and likely answers, reinforced by their black-box nature and remaining benchmarking biases. Our method explicitly grounds VQA tasks to video fragments in four steps: entailment tree construction, video-language entailment verification, tree reasoning, and dynamic tree expansion. A vital benefit of the method is its generalizability to current video and image-based VLMs across reasoning types. To support fair evaluation, we devise a de-biasing procedure based on large-language models that rewrites VQA benchmark answer sets to enforce model reasoning. Systematic experiments on existing and de-biased benchmarks highlight the impact of our method components across benchmarks, VLMs, and reasoning types.

Bradley P. Allen, Filip Ilievski

Knowledge engineering is the process of creating and maintaining knowledge-producing systems. Throughout the history of computer science and AI, knowledge engineering workflows have been widely used given the importance of high-quality knowledge for reliable intelligent agents. Meanwhile, the scope of knowledge engineering, as apparent from its target tasks and use cases, has been shifting, together with its paradigms such as expert systems, semantic web, and language modeling. The intended use cases and supported user requirements between these paradigms have not been analyzed globally, as new paradigms often satisfy prior pain points while possibly introducing new ones. The recent abstraction of systemic patterns into a boxology provides an opening for aligning the requirements and use cases of knowledge engineering with the systems, components, and software that can satisfy them best. This paper proposes a vision of harmonizing the best practices in the field of knowledge engineering by leveraging the software engineering methodology of creating reference architectures. We describe how a reference architecture can be iteratively designed and implemented to associate user needs with recurring systemic patterns, building on top of existing knowledge engineering workflows and boxologies. We provide a six-step roadmap that can enable the development of such an architecture, providing an initial design and outcome of the definition of architectural scope, selection of information sources, and analysis. We expect that following through on this vision will lead to well-grounded reference architectures for knowledge engineering, will advance the ongoing initiatives of organizing the neurosymbolic knowledge engineering space, and will build new links to the software architectures and data science communities.

Filip Ilievski, Barbara Hammer, Frank van Harmelen et al.

Recent advances in AI -- including generative approaches -- have resulted in technology that can support humans in scientific discovery and forming decisions, but may also disrupt democracies and target individuals. The responsible use of AI and its participation in human-AI teams increasingly shows the need for AI alignment, that is, to make AI systems act according to our preferences. A crucial yet often overlooked aspect of these interactions is the different ways in which humans and machines generalise. In cognitive science, human generalisation commonly involves abstraction and concept learning. In contrast, AI generalisation encompasses out-of-domain generalisation in machine learning, rule-based reasoning in symbolic AI, and abstraction in neurosymbolic AI. In this perspective paper, we combine insights from AI and cognitive science to identify key commonalities and differences across three dimensions: notions of, methods for, and evaluation of generalisation. We map the different conceptualisations of generalisation in AI and cognitive science along these three dimensions and consider their role for alignment in human-AI teaming. This results in interdisciplinary challenges across AI and cognitive science that must be tackled to provide a foundation for effective and cognitively supported alignment in human-AI teaming scenarios.

Saurav Joshi, Filip Ilievski, Jay Pujara

According to WWF, 1.1 billion people lack access to water, and 2.7 billion experience water scarcity at least one month a year. By 2025, two-thirds of the world's population may be facing water shortages. This highlights the urgency of managing water usage efficiently, especially in water-intensive sectors like food. This paper proposes a recommendation engine, powered by knowledge graphs, aiming to facilitate sustainable and healthy food consumption. The engine recommends ingredient substitutes in user recipes that improve nutritional value and reduce environmental impact, particularly water footprint. The system architecture includes source identification, information extraction, schema alignment, knowledge graph construction, and user interface development. The research offers a promising tool for promoting healthier eating habits and contributing to water conservation efforts.

Mohammadhossein Khojasteh, Yifan Jiang, Stefano De Giorgis et al.

Analogical reasoning is a key driver of human generalization in problem-solving and argumentation. Yet, analogies between narrative structures remain challenging for machines. Cognitive engines for structural mapping are not directly applicable, as they assume pre-extracted entities, whereas LLMs' performance is sensitive to prompt format and the degree of surface similarity between narratives. This gap motivates a key question: What is the impact of enhancing structural mapping with LLM-derived abstractions on their analogical reasoning ability in narratives? To that end, we propose a modular framework named YARN (Yielding Abstractions for Reasoning in Narratives), which uses LLMs to decompose narratives into units, abstract these units, and then passes them to a mapping component that aligns elements across stories to perform analogical reasoning. We define and operationalize four levels of abstraction that capture both the general meaning of units and their roles in the story, grounded in prior work on framing. Our experiments reveal that abstractions consistently improve model performance, resulting in competitive or better performance than end-to-end LLM baselines. Closer error analysis reveals the remaining challenges in abstraction at the right level, in incorporating implicit causality, and an emerging categorization of analogical patterns in narratives. YARN enables systematic variation of experimental settings to analyze component contributions, and to support future work, we make the code for YARN openly available.

Mustafa Kaan Aslan, Reinout Heijungs, Filip Ilievski

Environmental sustainability, particularly in relation to climate change, is a key concern for consumers, producers, and policymakers. The carbon footprint, based on greenhouse gas emissions, is a standard metric for quantifying the contribution to climate change of activities and is often assessed using life cycle assessment (LCA). However, conducting LCA is complex due to opaque and global supply chains, as well as fragmented data. This paper presents a methodology that combines advances in LCA and publicly available databases with knowledge-augmented AI techniques, including retrieval-augmented generation, to estimate cradle-to-gate carbon footprints of food products. We introduce a chatbot interface that allows users to interactively explore the carbon impact of composite meals and relate the results to familiar activities. A live web demonstration showcases our proof-of-concept system with arbitrary food items and follow-up questions, highlighting both the potential and limitations - such as database uncertainties and AI misinterpretations - of delivering LCA insights in an accessible format.

Bradley P. Allen, Prateek Chhikara, Thomas Macaulay Ferguson et al.

Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but they exhibit problems with logical consistency in the output they generate. How can we harness LLMs' broad-coverage parametric knowledge in formal reasoning despite their inconsistency? We present a method for directly integrating an LLM into the interpretation function of the formal semantics for a paraconsistent logic. We provide experimental evidence for the feasibility of the method by evaluating the function using datasets created from several short-form factuality benchmarks. Unlike prior work, our method offers a theoretical framework for neurosymbolic reasoning that leverages an LLM's knowledge while preserving the underlying logic's soundness and completeness properties.

Tygo Bloem, Filip Ilievski

Meme clustering is critical for toxicity detection, virality modeling, and typing, but it has received little attention in previous research. Clustering similar Internet memes is challenging due to their multimodality, cultural context, and adaptability. Existing approaches rely on databases, overlook semantics, and struggle to handle diverse dimensions of similarity. This paper introduces a novel method that uses template-based matching with multi-dimensional similarity features, thus eliminating the need for predefined databases and supporting adaptive matching. Memes are clustered using local and global features across similarity categories such as form, visual content, text, and identity. Our combined approach outperforms existing clustering methods, producing more consistent and coherent clusters, while similarity-based feature sets enable adaptability and align with human intuition. We make all supporting code publicly available to support subsequent research.

Darshan Deshpande, Zhivar Sourati, Filip Ilievski et al.

Automatic assessment of the quality of arguments has been recognized as a challenging task with significant implications for misinformation and targeted speech. While real-world arguments are tightly anchored in context, existing computational methods analyze their quality in isolation, which affects their accuracy and generalizability. We propose SPARK: a novel method for scoring argument quality based on contextualization via relevant knowledge. We devise four augmentations that leverage large language models to provide feedback, infer hidden assumptions, supply a similar-quality argument, or give a counter-argument. SPARK uses a dual-encoder Transformer architecture to enable the original argument and its augmentation to be considered jointly. Our experiments in both in-domain and zero-shot setups show that SPARK consistently outperforms existing techniques across multiple metrics.

Prateek Chhikara, Jiarui Zhang, Filip Ilievski et al.

Communication via natural language is a key aspect of machine intelligence, and it requires computational models to learn and reason about world concepts, with varying levels of supervision. Significant progress has been made on fully-supervised non-interactive tasks, such as question-answering and procedural text understanding. Yet, various sequential interactive tasks, as in text-based games, have revealed limitations of existing approaches in terms of coherence, contextual awareness, and their ability to learn effectively from the environment. In this paper, we propose a knowledge-injection framework for improved functional grounding of agents in text-based games. Specifically, we consider two forms of domain knowledge that we inject into learning-based agents: memory of previous correct actions and affordances of relevant objects in the environment. Our framework supports two representative model classes: reinforcement learning agents and language model agents. Furthermore, we devise multiple injection strategies for the above domain knowledge types and agent architectures, including injection via knowledge graphs and augmentation of the existing input encoding strategies. We experiment with four models on the 10 tasks in the ScienceWorld text-based game environment, to illustrate the impact of knowledge injection on various model configurations and challenging task settings. Our findings provide crucial insights into the interplay between task properties, model architectures, and domain knowledge for interactive contexts.

Alessandro Oltramari, Jonathan Francis, Filip Ilievski et al.

This chapter illustrates how suitable neuro-symbolic models for language understanding can enable domain generalizability and robustness in downstream tasks. Different methods for integrating neural language models and knowledge graphs are discussed. The situations in which this combination is most appropriate are characterized, including quantitative evaluation and qualitative error analysis on a variety of commonsense question answering benchmark datasets.

PeiFeng Wang, Jonathan Zamora, Junfeng Liu et al.

Humans use natural language to compose common concepts from their environment into plausible, day-to-day scene descriptions. However, such generative commonsense reasoning (GCSR) skills are lacking in state-of-the-art text generation methods. Descriptive sentences about arbitrary concepts generated by neural text generation models (e.g., pre-trained text-to-text Transformers) are often grammatically fluent but may not correspond to human common sense, largely due to their lack of mechanisms to capture concept relations, to identify implicit concepts, and to perform generalizable reasoning about unseen concept compositions. In this paper, we propose an Imagine-and-Verbalize (I&V) method, which learns to imagine a relational scene knowledge graph (SKG) with relations between the input concepts, and leverage the SKG as a constraint when generating a plausible scene description. We collect and harmonize a set of knowledge resources from different domains and modalities, providing a rich auxiliary supervision signal for I&V. The experiments demonstrate the effectiveness of I&V in improving language models on both concept-to-sentence and concept-to-story generation tasks, while enabling the model to learn well from fewer task examples and generate SKGs that make common sense to human annotators.

Kaixin Ma, Filip Ilievski, Jonathan Francis et al.

Commonsense reasoning benchmarks have been largely solved by fine-tuning language models. The downside is that fine-tuning may cause models to overfit to task-specific data and thereby forget their knowledge gained during pre-training. Recent works only propose lightweight model updates as models may already possess useful knowledge from past experience, but a challenge remains in understanding what parts and to what extent models should be refined for a given task. In this paper, we investigate what models learn from commonsense reasoning datasets. We measure the impact of three different adaptation methods on the generalization and accuracy of models. Our experiments with two models show that fine-tuning performs best, by learning both the content and the structure of the task, but suffers from overfitting and limited generalization to novel answers. We observe that alternative adaptation methods like prefix-tuning have comparable accuracy, but generalize better to unseen answers and are more robust to adversarial splits.

Zaina Shaik, Filip Ilievski, Fred Morstatter

As an open and collaborative knowledge graph created by users and bots, it is possible that the knowledge in Wikidata is biased in regards to multiple factors such as gender, race, and country of citizenship. Previous work has mostly studied the representativeness of Wikidata knowledge in terms of genders of people. In this paper, we examine the race and citizenship bias in general and in regards to STEM representation for scientists, software developers, and engineers. By comparing Wikidata queries to real-world datasets, we identify the differences in representation to characterize the biases present in Wikidata. Through this analysis, we discovered that there is an overrepresentation of white individuals and those with citizenship in Europe and North America; the rest of the groups are generally underrepresented. Based on these findings, we have found and linked to Wikidata additional data about STEM scientists from the minorities. This data is ready to be inserted into Wikidata with a bot. Increasing representation of minority race and country of citizenship groups can create a more accurate portrayal of individuals in STEM.

Filip Ilievski, Pedro Szekely, Gleb Satyukov et al.

While the similarity between two concept words has been evaluated and studied for decades, much less attention has been devoted to algorithms that can compute the similarity of nodes in very large knowledge graphs, like Wikidata. To facilitate investigations and head-to-head comparisons of similarity algorithms on Wikidata, we present a user-friendly interface that allows flexible computation of similarity between Qnodes in Wikidata. At present, the similarity interface supports four algorithms, based on: graph embeddings (TransE, ComplEx), text embeddings (BERT), and class-based similarity. We demonstrate the behavior of the algorithms on representative examples about semantically similar, related, and entirely unrelated entity pairs. To support anticipated applications that require efficient similarity computations, like entity linking and recommendation, we also provide a REST API that can compute most similar neighbors for any Qnode in Wikidata.

Hans Chalupsky, Pedro Szekely, Filip Ilievski et al.

Application developers today have three choices for exploiting the knowledge present in Wikidata: they can download the Wikidata dumps in JSON or RDF format, they can use the Wikidata API to get data about individual entities, or they can use the Wikidata SPARQL endpoint. None of these methods can support complex, yet common, query use cases, such as retrieval of large amounts of data or aggregations over large fractions of Wikidata. This paper introduces KGTK Kypher, a query language and processor that allows users to create personalized variants of Wikidata on a laptop. We present several use cases that illustrate the types of analyses that Kypher enables users to run on the full Wikidata KG on a laptop, combining data from external resources such as DBpedia. The Kypher queries for these use cases run much faster on a laptop than the equivalent SPARQL queries on a Wikidata clone running on a powerful server with 24h time-out limits.

Kartik Shenoy, Filip Ilievski, Daniel Garijo et al.

Wikidata has been increasingly adopted by many communities for a wide variety of applications, which demand high-quality knowledge to deliver successful results. In this paper, we develop a framework to detect and analyze low-quality statements in Wikidata by shedding light on the current practices exercised by the community. We explore three indicators of data quality in Wikidata, based on: 1) community consensus on the currently recorded knowledge, assuming that statements that have been removed and not added back are implicitly agreed to be of low quality; 2) statements that have been deprecated; and 3) constraint violations in the data. We combine these indicators to detect low-quality statements, revealing challenges with duplicate entities, missing triples, violated type rules, and taxonomic distinctions. Our findings complement ongoing efforts by the Wikidata community to improve data quality, aiming to make it easier for users and editors to find and correct mistakes.

Peifeng Wang, Filip Ilievski, Muhao Chen et al.

Inspired by evidence that pretrained language models (LMs) encode commonsense knowledge, recent work has applied LMs to automatically populate commonsense knowledge graphs (CKGs). However, there is a lack of understanding on their generalization to multiple CKGs, unseen relations, and novel entities. This paper analyzes the ability of LMs to perform generalizable commonsense inference, in terms of knowledge capacity, transferability, and induction. Our experiments with these three aspects show that: (1) LMs can adapt to different schemas defined by multiple CKGs but fail to reuse the knowledge to generalize to new relations. (2) Adapted LMs generalize well to unseen subjects, but less so on novel objects. Future work should investigate how to improve the transferability and induction of commonsense mining from LMs.

Ehsan Qasemi, Filip Ilievski, Muhao Chen et al.

Humans can seamlessly reason with circumstantial preconditions of commonsense knowledge. We understand that a glass is used for drinking water, unless the glass is broken or the water is toxic. Despite state-of-the-art (SOTA) language models' (LMs) impressive performance on inferring commonsense knowledge, it is unclear whether they understand the circumstantial preconditions. To address this gap, we propose a novel challenge of reasoning with circumstantial preconditions. We collect a dataset, called PaCo, consisting of 12.4 thousand preconditions of commonsense statements expressed in natural language. Based on this dataset, we create three canonical evaluation tasks and use them to examine the capability of existing LMs to understand situational preconditions. Our results reveal a 10-30% gap between machine and human performance on our tasks, which shows that reasoning with preconditions is an open challenge.

Avijit Thawani, Jay Pujara, Pedro A. Szekely et al.

NLP systems rarely give special consideration to numbers found in text. This starkly contrasts with the consensus in neuroscience that, in the brain, numbers are represented differently from words. We arrange recent NLP work on numeracy into a comprehensive taxonomy of tasks and methods. We break down the subjective notion of numeracy into 7 subtasks, arranged along two dimensions: granularity (exact vs approximate) and units (abstract vs grounded). We analyze the myriad representational choices made by 18 previously published number encoders and decoders. We synthesize best practices for representing numbers in text and articulate a vision for holistic numeracy in NLP, comprised of design trade-offs and a unified evaluation.