Mehar Bhatia

Mehar Bhatia, Vered Shwartz

With the increasing integration of AI into everyday life, it's becoming crucial to design AI systems that serve users from diverse backgrounds by making them culturally aware. In this paper, we present GD-COMET, a geo-diverse version of the COMET commonsense inference model. GD-COMET goes beyond Western commonsense knowledge and is capable of generating inferences pertaining to a broad range of cultures. We demonstrate the effectiveness of GD-COMET through a comprehensive human evaluation across 5 diverse cultures, as well as extrinsic evaluation on a geo-diverse task. The evaluation shows that GD-COMET captures and generates culturally nuanced commonsense knowledge, demonstrating its potential to benefit NLP applications across the board and contribute to making NLP more inclusive.

Mehar Bhatia, Shravan Nayak, Gaurav Kamath et al.

As LLMs occupy an increasingly important role in society, they are more and more confronted with questions that require them not only to draw on their general knowledge but also to align with certain human value systems. Therefore, studying the alignment of LLMs with human values has become a crucial field of inquiry. Prior work, however, mostly focuses on evaluating the alignment of fully trained models, overlooking the training dynamics by which models learn to express human values. In this work, we investigate how and at which stage value alignment arises during the course of a model's post-training. Our analysis disentangles the effects of post-training algorithms and datasets, measuring both the magnitude and time of value drifts during training. Experimenting with Llama-3 and Qwen-3 models of different sizes and popular supervised fine-tuning (SFT) and preference optimization datasets and algorithms, we find that the SFT phase generally establishes a model's values, and subsequent preference optimization rarely re-aligns these values. Furthermore, using a synthetic preference dataset that enables controlled manipulation of values, we find that different preference optimization algorithms lead to different value alignment outcomes, even when preference data is held constant. Our findings provide actionable insights into how values are learned during post-training and help to inform data curation, as well as the selection of models and algorithms for preference optimization to improve model alignment to human values.

Sara Vera Marjanović, Arkil Patel, Vaibhav Adlakha et al. · eth-zurich, microsoft-research

Large Reasoning Models like DeepSeek-R1 mark a fundamental shift in how LLMs approach complex problems. Instead of directly producing an answer for a given input, DeepSeek-R1 creates detailed multi-step reasoning chains, seemingly "thinking" about a problem before providing an answer. This reasoning process is publicly available to the user, creating endless opportunities for studying the reasoning behaviour of the model and opening up the field of Thoughtology. Starting from a taxonomy of DeepSeek-R1's basic building blocks of reasoning, our analyses on DeepSeek-R1 investigate the impact and controllability of thought length, management of long or confusing contexts, cultural and safety concerns, and the status of DeepSeek-R1 vis-à-vis cognitive phenomena, such as human-like language processing and world modelling. Our findings paint a nuanced picture. Notably, we show DeepSeek-R1 has a 'sweet spot' of reasoning, where extra inference time can impair model performance. Furthermore, we find a tendency for DeepSeek-R1 to persistently ruminate on previously explored problem formulations, obstructing further exploration. We also note strong safety vulnerabilities of DeepSeek-R1 compared to its non-reasoning counterpart, which can also compromise safety-aligned LLMs.

Yu Ying Chiu, Liwei Jiang, Maria Antoniak et al. · cmu, uw

Frontier large language models (LLMs) are developed by researchers and practitioners with skewed cultural backgrounds and on datasets with skewed sources. However, LLMs' (lack of) multicultural knowledge cannot be effectively assessed with current methods for developing benchmarks. Existing multicultural evaluations primarily rely on expensive and restricted human annotations or potentially outdated internet resources. Thus, they struggle to capture the intricacy, dynamics, and diversity of cultural norms. LLM-generated benchmarks are promising, yet risk propagating the same biases they are meant to measure. To synergize the creativity and expert cultural knowledge of human annotators and the scalability and standardizability of LLM-based automation, we introduce CulturalTeaming, an interactive red-teaming system that leverages human-AI collaboration to build truly challenging evaluation dataset for assessing the multicultural knowledge of LLMs, while improving annotators' capabilities and experiences. Our study reveals that CulturalTeaming's various modes of AI assistance support annotators in creating cultural questions, that modern LLMs fail at, in a gamified manner. Importantly, the increased level of AI assistance (e.g., LLM-generated revision hints) empowers users to create more difficult questions with enhanced perceived creativity of themselves, shedding light on the promises of involving heavier AI assistance in modern evaluation dataset creation procedures. Through a series of 1-hour workshop sessions, we gather CULTURALBENCH-V0.1, a compact yet high-quality evaluation dataset with users' red-teaming attempts, that different families of modern LLMs perform with accuracy ranging from 37.7% to 72.2%, revealing a notable gap in LLMs' multicultural proficiency.

Shravan Nayak, Mehar Bhatia, Xiaofeng Zhang et al. · eth-zurich

The increasing ubiquity of text-to-image (T2I) models as tools for visual content generation raises concerns about their ability to accurately represent diverse cultural contexts -- where missed cues can stereotype communities and undermine usability. In this work, we present the first study to systematically quantify the alignment of T2I models and evaluation metrics with respect to both explicit (stated) as well as implicit (unstated, implied by the prompt's cultural context) cultural expectations. To this end, we introduce CulturalFrames, a novel benchmark designed for rigorous human evaluation of cultural representation in visual generations. Spanning 10 countries and 5 socio-cultural domains, CulturalFrames comprises 983 prompts, 3637 corresponding images generated by 4 state-of-the-art T2I models, and over 10k detailed human annotations. We find that across models and countries, cultural expectations are missed an average of 44% of the time. Among these failures, explicit expectations are missed at a surprisingly high average rate of 68%, while implicit expectation failures are also significant, averaging 49%. Furthermore, we show that existing T2I evaluation metrics correlate poorly with human judgments of cultural alignment, irrespective of their internal reasoning. Collectively, our findings expose critical gaps, provide a concrete testbed, and outline actionable directions for developing culturally informed T2I models and metrics that improve global usability.

Karolina Stańczak, Nicholas Meade, Mehar Bhatia et al. · eth-zurich

Recent progress in large language models (LLMs) has focused on producing responses that meet human expectations and align with shared values - a process coined alignment. However, aligning LLMs remains challenging due to the inherent disconnect between the complexity of human values and the narrow nature of the technological approaches designed to address them. Current alignment methods often lead to misspecified objectives, reflecting the broader issue of incomplete contracts, the impracticality of specifying a contract between a model developer, and the model that accounts for every scenario in LLM alignment. In this paper, we argue that improving LLM alignment requires incorporating insights from societal alignment frameworks, including social, economic, and contractual alignment, and discuss potential solutions drawn from these domains. Given the role of uncertainty within societal alignment frameworks, we then investigate how it manifests in LLM alignment. We end our discussion by offering an alternative view on LLM alignment, framing the underspecified nature of its objectives as an opportunity rather than perfect their specification. Beyond technical improvements in LLM alignment, we discuss the need for participatory alignment interface designs.

Zhuozhuo Joy Liu, Farhan Samir, Mehar Bhatia et al.

LLMs have been demonstrated to align with the values of Western or North American cultures. Prior work predominantly showed this effect through leveraging surveys that directly ask (originally people and now also LLMs) about their values. However, it is hard to believe that LLMs would consistently apply those values in real-world scenarios. To address that, we take a bottom-up approach, asking LLMs to reason about cultural norms in narratives from different cultures. We find that GPT-4 tends to generate norms that, while not necessarily incorrect, are significantly less culture-specific. In addition, while it avoids overtly generating stereotypes, the stereotypical representations of certain cultures are merely hidden rather than suppressed in the model, and such stereotypes can be easily recovered. Addressing these challenges is a crucial step towards developing LLMs that fairly serve their diverse user base.

Mehar Bhatia, Sahithya Ravi, Aditya Chinchure et al.

Despite recent advancements in vision-language models, their performance remains suboptimal on images from non-western cultures due to underrepresentation in training datasets. Various benchmarks have been proposed to test models' cultural inclusivity, but they have limited coverage of cultures and do not adequately assess cultural diversity across universal as well as culture-specific local concepts. To address these limitations, we introduce the GlobalRG benchmark, comprising two challenging tasks: retrieval across universals and cultural visual grounding. The former task entails retrieving culturally diverse images for universal concepts from 50 countries, while the latter aims at grounding culture-specific concepts within images from 15 countries. Our evaluation across a wide range of models reveals that the performance varies significantly across cultures -- underscoring the necessity for enhancing multicultural understanding in vision-language models.

Anuj Saraswat, Mehar Bhatia, Yaman Kumar Singla et al.

Recent studies in speech perception have been closely linked to fields of cognitive psychology, phonology, and phonetics in linguistics. During perceptual attunement, a critical and sensitive developmental trajectory has been examined in bilingual and monolingual infants where they can best discriminate common phonemes. In this paper, we compare and identify these cognitive aspects on deep neural-based visual lip-reading models. We conduct experiments on the two most extensive public visual speech recognition datasets for English and Mandarin. Through our experimental results, we observe a strong correlation between these theories in cognitive psychology and our unique modeling. We inspect how these computational models develop similar phases in speech perception and acquisitions.

Mehar Bhatia, Tenzin Singhay Bhotia, Akshat Agarwal et al.

This paper is a contribution to the Hate Speech and Offensive Content Identification in Indo-European Languages (HASOC) 2021 shared task. Social media today is a hotbed of toxic and hateful conversations, in various languages. Recent news reports have shown that current models struggle to automatically identify hate posted in minority languages. Therefore, efficiently curbing hate speech is a critical challenge and problem of interest. We present a multilingual architecture using state-of-the-art transformer language models to jointly learn hate and offensive speech detection across three languages namely, English, Hindi, and Marathi. On the provided testing corpora, we achieve Macro F1 scores of 0.7996, 0.7748, 0.8651 for sub-task 1A and 0.6268, 0.5603 during the fine-grained classification of sub-task 1B. These results show the efficacy of exploiting a multilingual training scheme.

Anubha Kabra, Mehar Bhatia, Yaman Kumar et al.

Automatic scoring engines have been used for scoring approximately fifteen million test-takers in just the last three years. This number is increasing further due to COVID-19 and the associated automation of education and testing. Despite such wide usage, the AI-based testing literature of these "intelligent" models is highly lacking. Most of the papers proposing new models rely only on quadratic weighted kappa (QWK) based agreement with human raters for showing model efficacy. However, this effectively ignores the highly multi-feature nature of essay scoring. Essay scoring depends on features like coherence, grammar, relevance, sufficiency and, vocabulary. To date, there has been no study testing Automated Essay Scoring: AES systems holistically on all these features. With this motivation, we propose a model agnostic adversarial evaluation scheme and associated metrics for AES systems to test their natural language understanding capabilities and overall robustness. We evaluate the current state-of-the-art AES models using the proposed scheme and report the results on five recent models. These models range from feature-engineering-based approaches to the latest deep learning algorithms. We find that AES models are highly overstable. Even heavy modifications(as much as 25%) with content unrelated to the topic of the questions do not decrease the score produced by the models. On the other hand, irrelevant content, on average, increases the scores, thus showing that the model evaluation strategy and rubrics should be reconsidered. We also ask 200 human raters to score both an original and adversarial response to seeing if humans can detect differences between the two and whether they agree with the scores assigned by auto scores.