Mubashara Akhtar

Mubashara Akhtar, Anka Reuel, Prajna Soni et al. · meta-ai

Artificial Intelligence (AI) benchmarks play a central role in measuring progress in model development and guiding deployment decisions. However, many benchmarks quickly become saturated, meaning that they can no longer differentiate between the best-performing models, diminishing their long-term value. In this study, we analyze benchmark saturation across 60 Large Language Model (LLM) benchmarks selected from technical reports by major model developers. To identify factors driving saturation, we characterize benchmarks along 14 properties spanning task design, data construction, and evaluation format. We test five hypotheses examining how each property contributes to saturation rates. Our analysis reveals that nearly half of the benchmarks exhibit saturation, with rates increasing as benchmarks age. Notably, hiding test data (i.e., public vs. private) shows no protective effect, while expert-curated benchmarks resist saturation better than crowdsourced ones. Our findings highlight which design choices extend benchmark longevity and inform strategies for more durable evaluation.

Junling Wang, Boqi Chen, Heejin Do et al.

AI systems are increasingly used to support educational content creation, yet it remains unclear whether they can generate outputs that faithfully represent the pedagogical concepts they are intended to teach. Thus, we introduce equation-to-visual generation, a task that, in contrast to conventional image generation, requires producing pedagogically meaningful visuals from arithmetic equations while precisely preserving their numerical and relational structure. Informed by interviews with teachers and an analysis of educational materials, we construct E2V-Bench, a benchmark spanning four pedagogically grounded visual types, along with automatic metrics for evaluating visual correctness. Our evaluation reveals that recent text-to-image (T2I) models frequently fail on this task, with errors dominated by incorrect object counts and broken relational structure. Building on this, we explore benchmark-guided enhancement strategies. These strategies improve representative models, while the remaining gap calls for stronger numerical and relational grounding in future T2I models.

Mubashara Akhtar, Oana Cocarascu, Elena Simperl

Evidence data for automated fact-checking (AFC) can be in multiple modalities such as text, tables, images, audio, or video. While there is increasing interest in using images for AFC, previous works mostly focus on detecting manipulated or fake images. We propose a novel task, chart-based fact-checking, and introduce ChartBERT as the first model for AFC against chart evidence. ChartBERT leverages textual, structural and visual information of charts to determine the veracity of textual claims. For evaluation, we create ChartFC, a new dataset of 15, 886 charts. We systematically evaluate 75 different vision-language (VL) baselines and show that ChartBERT outperforms VL models, achieving 63.8% accuracy. Our results suggest that the task is complex yet feasible, with many challenges ahead.

Mubashara Akhtar, Abhilash Shankarampeta, Vivek Gupta et al.

Numbers are crucial for various real-world domains such as finance, economics, and science. Thus, understanding and reasoning with numbers are essential skills for language models to solve different tasks. While different numerical benchmarks have been introduced in recent years, they are limited to specific numerical aspects mostly. In this paper, we propose a hierarchical taxonomy for numerical reasoning skills with more than ten reasoning types across four levels: representation, number sense, manipulation, and complex reasoning. We conduct a comprehensive evaluation of state-of-the-art models to identify reasoning challenges specific to them. Henceforth, we develop a diverse set of numerical probes employing a semi-automated approach. We focus on the tabular Natural Language Inference (TNLI) task as a case study and measure models' performance shifts. Our results show that no model consistently excels across all numerical reasoning types. Among the probed models, FlanT5 (few-/zero-shot) and GPT-3.5 (few-shot) demonstrate strong overall numerical reasoning skills compared to other models. Label-flipping probes indicate that models often exploit dataset artifacts to predict the correct labels.

Mubashara Akhtar, Nikesh Subedi, Vivek Gupta et al.

Whilst fact verification has attracted substantial interest in the natural language processing community, verifying misinforming statements against data visualizations such as charts has so far been overlooked. Charts are commonly used in the real-world to summarize and communicate key information, but they can also be easily misused to spread misinformation and promote certain agendas. In this paper, we introduce ChartCheck, a novel, large-scale dataset for explainable fact-checking against real-world charts, consisting of 1.7k charts and 10.5k human-written claims and explanations. We systematically evaluate ChartCheck using vision-language and chart-to-table models, and propose a baseline to the community. Finally, we study chart reasoning types and visual attributes that pose a challenge to these models

Jingwei Ni, Ekaterina Fadeeva, Tianyi Wu et al.

Solving complex tasks usually requires LLMs to generate long multi-step reasoning chains. Previous work has shown that verifying the correctness of individual reasoning steps can further improve the performance and efficiency of LLMs on such tasks and enhance solution interpretability. However, existing verification approaches, such as Process Reward Models (PRMs), are either computationally expensive, limited to specific domains, or require large-scale human or model-generated annotations. Thus, we propose a lightweight alternative for step-level reasoning verification based on data-driven uncertainty scores. We train transformer-based uncertainty quantification heads (UHeads) that use the internal states of a frozen LLM to estimate the uncertainty of its reasoning steps during generation. The approach is fully automatic: target labels are generated either by another larger LLM (e.g., DeepSeek R1) or in a self-supervised manner by the original model itself. UHeads are both effective and lightweight, containing less than 10M parameters. Across multiple domains, including mathematics, planning, and general knowledge question answering, they match or even surpass the performance of PRMs that are up to 810x larger. Our findings suggest that the internal states of LLMs encode their uncertainty and can serve as reliable signals for reasoning verification, offering a promising direction toward scalable and generalizable introspective LLMs.

Anka Reuel, Avijit Ghosh, Jenny Chim et al.

Foundation models are increasingly central to high-stakes AI systems, and governance frameworks now depend on evaluations to assess their risks and capabilities. Although general capability evaluations are widespread, social impact assessments covering bias, fairness, privacy, environmental costs, and labor practices remain uneven across the AI ecosystem. To characterize this landscape, we conduct the first comprehensive analysis of both first-party and third-party social impact evaluation reporting across a wide range of model developers. Our study examines 186 first-party release reports and 183 post-release evaluation sources, and complements this quantitative analysis with interviews of model developers. We find a clear division of evaluation labor: first-party reporting is sparse, often superficial, and has declined over time in key areas such as environmental impact and bias, while third-party evaluators including academic researchers, nonprofits, and independent organizations provide broader and more rigorous coverage of bias, harmful content, and performance disparities. However, this complementarity has limits. Only model developers can authoritatively report on data provenance, content moderation labor, financial costs, and training infrastructure, yet interviews reveal that these disclosures are often deprioritized unless tied to product adoption or regulatory compliance. Our findings indicate that current evaluation practices leave major gaps in assessing AI's societal impacts, highlighting the urgent need for policies that promote developer transparency, strengthen independent evaluation ecosystems, and create shared infrastructure to aggregate and compare third-party evaluations in a consistent and accessible way.

Mubashara Akhtar, Michael Schlichtkrull, Andreas Vlachos

Current automated fact-checking (AFC) approaches typically evaluate evidence either implicitly via the predicted verdicts or through exact matches with predefined closed knowledge sources, such as Wikipedia. However, these methods are limited due to their reliance on evaluation metrics originally designed for other purposes and constraints from closed knowledge sources. In this work, we introduce \textbf{\textcolor{skyblue}{Ev\textsuperscript{2}}\textcolor{orangebrown}{R}} which combines the strengths of reference-based evaluation and verdict-level proxy scoring. Ev\textsuperscript{2}R jointly assesses how well the evidence aligns with the gold references and how reliably it supports the verdict, addressing the shortcomings of prior methods. We evaluate Ev\textsuperscript{2}R against three types of evidence evaluation approaches: reference-based, proxy-reference, and reference-less baselines. Assessments against human ratings and adversarial tests demonstrate that Ev\textsuperscript{2}R consistently outperforms existing scoring approaches in accuracy and robustness. It achieves stronger correlation with human judgments and greater robustness to adversarial perturbations, establishing it as a reliable metric for evidence evaluation in AFC.\footnote{Code is available at \href{https://github.com/mubasharaak/fc-evidence-evaluation}{https://github.com/mubasharaak/fc-evidence-evaluation}.}

Yu Fan, Jingwei Ni, Jakob Merane et al. · eth-zurich

Long-form legal reasoning remains a key challenge for large language models (LLMs) in spite of recent advances in test-time scaling. To address this, we introduce \textsc{LEXam}, a novel benchmark derived from 340 law exams spanning 116 law school courses across a range of subjects and degree levels. The dataset comprises 4,886 law exam questions in English and German, including 2,841 long-form, open-ended questions and 2,045 multiple-choice questions. Besides reference answers, the open questions are also accompanied by explicit guidance outlining the expected legal reasoning approach such as issue spotting, rule recall, or rule application. Our evaluation on both open-ended and multiple-choice questions present significant challenges for current LLMs; in particular, they notably struggle with open questions that require structured, multi-step legal reasoning. Moreover, our results underscore the effectiveness of the dataset in differentiating between models with varying capabilities. Deploying an ensemble LLM-as-a-Judge paradigm with rigorous human expert validation, we demonstrate how model-generated reasoning steps can be evaluated consistently and accurately, closely aligning with human expert assessments. Our evaluation setup provides a scalable method to assess legal reasoning quality beyond simple accuracy metrics. We have open-sourced our code on https://github.com/LEXam-Benchmark/LEXam and released our data on https://huggingface.co/datasets/LEXam-Benchmark/LEXam. Project page: https://lexam-benchmark.github.io.

Ziheng Chi, Yifan Hou, Chenxi Pang et al. · eth-zurich

Diagrams convey symbolic information in a visual format rather than a linear stream of words, making them especially challenging for AI models to process. While recent evaluations suggest that vision-language models (VLMs) perform well on diagram-related benchmarks, their reliance on knowledge, reasoning, or modality shortcuts raises concerns about whether they genuinely understand and reason over diagrams. To address this gap, we introduce Chimera, a comprehensive test suite comprising 7,500 high-quality diagrams sourced from Wikipedia; each diagram is annotated with its symbolic content represented by semantic triples along with multi-level questions designed to assess four fundamental aspects of diagram comprehension: entity recognition, relation understanding, knowledge grounding, and visual reasoning. We use Chimera to measure the presence of three types of shortcuts in visual question answering: (1) the visual-memorization shortcut, where VLMs rely on memorized visual patterns; (2) the knowledge-recall shortcut, where models leverage memorized factual knowledge instead of interpreting the diagram; and (3) the Clever-Hans shortcut, where models exploit superficial language patterns or priors without true comprehension. We evaluate 15 open-source VLMs from 7 model families on Chimera and find that their seemingly strong performance largely stems from shortcut behaviors: visual-memorization shortcuts have slight impact, knowledge-recall shortcuts play a moderate role, and Clever-Hans shortcuts contribute significantly. These findings expose critical limitations in current VLMs and underscore the need for more robust evaluation protocols that benchmark genuine comprehension of complex visual inputs (e.g., diagrams) rather than question-answering shortcuts.

Mubashara Akhtar, Omar Benjelloun, Costanza Conforti et al.

Data is a critical resource for machine learning (ML), yet working with data remains a key friction point. This paper introduces Croissant, a metadata format for datasets that creates a shared representation across ML tools, frameworks, and platforms. Croissant makes datasets more discoverable, portable, and interoperable, thereby addressing significant challenges in ML data management. Croissant is already supported by several popular dataset repositories, spanning hundreds of thousands of datasets, enabling easy loading into the most commonly-used ML frameworks, regardless of where the data is stored. Our initial evaluation by human raters shows that Croissant metadata is readable, understandable, complete, yet concise.

Michael Schlichtkrull, Yulong Chen, Chenxi Whitehouse et al. · amazon-science

The Automated Verification of Textual Claims (AVeriTeC) shared task asks participants to retrieve evidence and predict veracity for real-world claims checked by fact-checkers. Evidence can be found either via a search engine, or via a knowledge store provided by the organisers. Submissions are evaluated using AVeriTeC score, which considers a claim to be accurately verified if and only if both the verdict is correct and retrieved evidence is considered to meet a certain quality threshold. The shared task received 21 submissions, 18 of which surpassed our baseline. The winning team was TUDA_MAI with an AVeriTeC score of 63%. In this paper we describe the shared task, present the full results, and highlight key takeaways from the shared task.

Shaona Ghosh, Heather Frase, Adina Williams et al. · deepmind, stanford

The rapid advancement and deployment of AI systems have created an urgent need for standard safety-evaluation frameworks. This paper introduces AILuminate v1.0, the first comprehensive industry-standard benchmark for assessing AI-product risk and reliability. Its development employed an open process that included participants from multiple fields. The benchmark evaluates an AI system's resistance to prompts designed to elicit dangerous, illegal, or undesirable behavior in 12 hazard categories, including violent crimes, nonviolent crimes, sex-related crimes, child sexual exploitation, indiscriminate weapons, suicide and self-harm, intellectual property, privacy, defamation, hate, sexual content, and specialized advice (election, financial, health, legal). Our method incorporates a complete assessment standard, extensive prompt datasets, a novel evaluation framework, a grading and reporting system, and the technical as well as organizational infrastructure for long-term support and evolution. In particular, the benchmark employs an understandable five-tier grading scale (Poor to Excellent) and incorporates an innovative entropy-based system-response evaluation. In addition to unveiling the benchmark, this report also identifies limitations of our method and of building safety benchmarks generally, including evaluator uncertainty and the constraints of single-turn interactions. This work represents a crucial step toward establishing global standards for AI risk and reliability evaluation while acknowledging the need for continued development in areas such as multiturn interactions, multimodal understanding, coverage of additional languages, and emerging hazard categories. Our findings provide valuable insights for model developers, system integrators, and policymakers working to promote safer AI deployment.

Mubashara Akhtar, Chenxi Pang, Andreea Marzoca et al. · deepmind

Language models, potentially augmented with tool usage such as retrieval are becoming the go-to means of answering questions. Understanding and answering questions in real-world settings often requires retrieving information from different sources, processing and aggregating data to extract insights, and presenting complex findings in form of structured artifacts such as novel tables, charts, or infographics. In this paper, we introduce TANQ, the first open domain question answering dataset where the answers require building tables from information across multiple sources. We release the full source attribution for every cell in the resulting table and benchmark state-of-the-art language models in open, oracle, and closed book setups. Our best-performing baseline, Gemini Flash reaches an overall F1 score of 60.7, lagging behind human performance by 12.3 points. We analyse baselines' performance across different dataset attributes such as different skills required for this task, including multi-hop reasoning, math operations, and unit conversions. We further discuss common failures in model-generated answers, suggesting that TANQ is a complex task with many challenges ahead.

Yucheng Wang, Yifan Hou, Aydin Javadov et al. · eth-zurich

Multimodal large language models (MLLMs) promise enhanced reasoning by integrating diverse inputs such as text, vision, and audio. Yet cross-modal reasoning remains underexplored, with conflicting reports on whether added modalities help or harm performance. These inconsistencies stem from a lack of controlled evaluation frameworks and analysis of models' internals to isolate when and why modality interactions support or undermine reasoning. We address this gap through a logic-grounded evaluation framework that categorizes multimodal reasoning into six interaction patterns, varying how facts are distributed across modalities and logically combined. Empirically, additional modalities enhance reasoning only when they provide independent and sufficient reasoning paths, while redundant or chained entailment support often hurts performance. Moreover, reasoning degrades in three systematic ways: weaker modalities drag down overall performance, conflicts bias preference toward certain modalities, and joint signals from different modalities fail to be integrated effectively. Therefore, we identify two core failures: task-composition bottleneck, where recognition and reasoning cannot be jointly executed in one pass, and fusion bottleneck, where early integration introduces bias. For further investigation, we find that attention patterns fail to encode fact usefulness, but a simple two-step prompting (recognize then reason) restores performance, confirming the task-composition bottleneck. Moreover, modality identity remains recoverable in early layers, and softening attention in early fusion improves reasoning, highlighting biased fusion as another failure mode. Overall, our findings show that integration, not perception, is the main barrier to multimodal reasoning, suggesting composition-aware training and early fusion control as promising directions.

Nitisha Jain, Mubashara Akhtar, Joan Giner-Miguelez et al.

Data is critical to advancing AI technologies, yet its quality and documentation remain significant challenges, leading to adverse downstream effects (e.g., potential biases) in AI applications. This paper addresses these issues by introducing Croissant-RAI, a machine-readable metadata format designed to enhance the discoverability, interoperability, and trustworthiness of AI datasets. Croissant-RAI extends the Croissant metadata format and builds upon existing responsible AI (RAI) documentation frameworks, offering a standardized set of attributes and practices to facilitate community-wide adoption. Leveraging established web-publishing practices, such as Schema.org, Croissant-RAI enables dataset users to easily find and utilize RAI metadata regardless of the platform on which the datasets are published. Furthermore, it is seamlessly integrated into major data search engines, repositories, and machine learning frameworks, streamlining the reading and writing of responsible AI metadata within practitioners' existing workflows. Croissant-RAI was developed through a community-led effort. It has been designed to be adaptable to evolving documentation requirements and is supported by a Python library and a visual editor.

Mubashara Akhtar, Michael Schlichtkrull, Zhijiang Guo et al.

Misinformation is often conveyed in multiple modalities, e.g. a miscaptioned image. Multimodal misinformation is perceived as more credible by humans, and spreads faster than its text-only counterparts. While an increasing body of research investigates automated fact-checking (AFC), previous surveys mostly focus on text. In this survey, we conceptualise a framework for AFC including subtasks unique to multimodal misinformation. Furthermore, we discuss related terms used in different communities and map them to our framework. We focus on four modalities prevalent in real-world fact-checking: text, image, audio, and video. We survey benchmarks and models, and discuss limitations and promising directions for future research