Kabir Ahuja, Harshita Diddee, Rishav Hada et al. · microsoft-research
Generative AI models have shown impressive performance on many Natural Language Processing tasks such as language understanding, reasoning, and language generation. An important question being asked by the AI community today is about the capabilities and limits of these models, and it is clear that evaluating generative AI is very challenging. Most studies on generative LLMs have been restricted to English and it is unclear how capable these models are at understanding and generating text in other languages. We present the first comprehensive benchmarking of generative LLMs - MEGA, which evaluates models on standard NLP benchmarks, covering 16 NLP datasets across 70 typologically diverse languages. We compare the performance of generative LLMs including Chat-GPT and GPT-4 to State of the Art (SOTA) non-autoregressive models on these tasks to determine how well generative models perform compared to the previous generation of LLMs. We present a thorough analysis of the performance of models across languages and tasks and discuss challenges in improving the performance of generative LLMs on low-resource languages. We create a framework for evaluating generative LLMs in the multilingual setting and provide directions for future progress in the field.
Nirmit Arora, Sathvik Joel, Ishan Kavathekar et al.
LLM-based agents are increasingly deployed in multi-agent systems (MAS). As these systems move toward real-world applications, their security becomes paramount. Existing research largely evaluates single-agent security, leaving a critical gap in understanding the vulnerabilities introduced by multi-agent design. However, existing systems fall short due to lack of unified frameworks and metrics focusing on unique rejection modes in MAS. We present SafeAgents, a unified and extensible framework for fine-grained security assessment of MAS. SafeAgents systematically exposes how design choices such as plan construction strategies, inter-agent context sharing, and fallback behaviors affect susceptibility to adversarial prompting. We introduce Dharma, a diagnostic measure that helps identify weak links within multi-agent pipelines. Using SafeAgents, we conduct a comprehensive study across five widely adopted multi-agent architectures (centralized, decentralized, and hybrid variants) on four datasets spanning web tasks, tool use, and code generation. Our findings reveal that common design patterns carry significant vulnerabilities. For example, centralized systems that delegate only atomic instructions to sub-agents obscure harmful objectives, reducing robustness. Our results highlight the need for security-aware design in MAS. Link to code is https://github.com/microsoft/SafeAgents
Harshita Diddee, Sandipan Dandapat, Monojit Choudhury et al.
Leveraging shared learning through Massively Multilingual Models, state-of-the-art machine translation models are often able to adapt to the paucity of data for low-resource languages. However, this performance comes at the cost of significantly bloated models which are not practically deployable. Knowledge Distillation is one popular technique to develop competitive, lightweight models: In this work, we first evaluate its use to compress MT models focusing on languages with extremely limited training data. Through our analysis across 8 languages, we find that the variance in the performance of the distilled models due to their dependence on priors including the amount of synthetic data used for distillation, the student architecture, training hyperparameters and confidence of the teacher models, makes distillation a brittle compression mechanism. To mitigate this, we explore the use of post-training quantization for the compression of these models. Here, we find that while distillation provides gains across some low-resource languages, quantization provides more consistent performance trends for the entire range of languages, especially the lowest-resource languages in our target set.
Visual cues such as structure, emphasis, and icons play an important role in efficient information foraging by sighted individuals and make for a pleasurable reading experience. Blind, low-vision and other print-disabled individuals miss out on these cues since current OCR and text-to-speech software ignore them, resulting in a tedious reading experience. We identify four semantic goals for an enjoyable listening experience, and identify syntactic visual cues that help make progress towards these goals. Empirically, we find that preserving even one or two visual cues in aural form significantly enhances the experience for listening to print content.
We present very early results on using GPT-3 to perform question answering on tabular data. We find that stock pre-trained GPT-3 is able to zero-shot learn the table structure from a serialized JSON array-of-arrays representation, and able to answer lookup queries and simple comparison questions in natural language without any fine-tuning. We further find that simple prompt engineering to include few-shot static Q&A examples significantly improves accuracy. Lastly, we find that intermixing passage text improves accuracy even further on heterogeneous data. We apply our approach on a novel dataset of simple tables in newspaper infographics with promising results. Overall, we find much cause for optimism in this basic approach.
Infographics are often an integral component of scientific documents for reporting qualitative or quantitative findings as they make it much simpler to comprehend the underlying complex information. However, their interpretation continues to be a challenge for the blind, low-vision, and other print-impaired (BLV) individuals. In this paper, we propose ChartParser, a fully automated pipeline that leverages deep learning, OCR, and image processing techniques to extract all figures from a research paper, classify them into various chart categories (bar chart, line chart, etc.) and obtain relevant information from them, specifically bar charts (including horizontal, vertical, stacked horizontal and stacked vertical charts) which already have several exciting challenges. Finally, we present the retrieved content in a tabular format that is screen-reader friendly and accessible to the BLV users. We present a thorough evaluation of our approach by applying our pipeline to sample real-world annotated bar charts from research papers.
Accessing daily news content still remains a big challenge for people with print-impairment including blind and low-vision due to opacity of printed content and hindrance from online sources. In this paper, we present our approach for digitization of print newspaper into an accessible file format such as HTML. We use an ensemble of instance segmentation and detection framework for newspaper layout analysis and then OCR to recognize text elements such as headline and article text. Additionally, we propose EdgeMask loss function for Mask-RCNN framework to improve segmentation mask boundary and hence accuracy of downstream OCR task. Empirically, we show that our proposed loss function reduces the Word Error Rate (WER) of news article text by 32.5 %.
Printed documents continue to be a challenge for blind, low-vision, and other print-disabled (BLV) individuals. In this paper, we focus on the specific problem of (in-)accessibility of internal references to citations, footnotes, figures, tables and equations. While sighted users can flip to the referenced content and flip back in seconds, linear audio narration that BLV individuals rely on makes following these references extremely hard. We propose a vision based technique to locate the referenced content and extract metadata needed to (in subsequent work) inline a content summary into the audio narration. We apply our technique to citations in scientific documents and find it works well both on born-digital as well as scanned documents.
Multimodal Large Language Models (MLLMs) show reasoning promise, yet their visual perception is a critical bottleneck. Strikingly, MLLMs can produce correct answers even while misinterpreting crucial visual elements, masking these underlying failures. Our preliminary study on a joint perception-reasoning dataset revealed that for one leading MLLM, 29% of its correct answers to reasoning questions still exhibited visual perception errors. To systematically address this, we introduce "Do You See Me", a scalable benchmark with 1,758 images and 2,612 questions. It spans seven human-psychology inspired subtasks in 2D and 3D, featuring controllable complexity to rigorously evaluate MLLM visual skills. Our findings on 3 leading closed-source and 5 major open-source models reveal a stark deficit: humans achieve 96.49% accuracy, while top MLLMs average below 50%. This performance gap widens rapidly with increased task complexity (e.g., from 12% to 45% in the visual form constancy subtask). Further analysis into the root causes suggests that failures stem from challenges like misallocated visual attention and the instability of internal representations for fine-grained details, especially at or below encoder patch resolution. This underscores an urgent need for MLLMs with truly robust visual perception. The benchmark dataset, source code and evaluation scripts are available at https://github.com/microsoft/Do-You-See-Me.
Spatial representation learning (SRL) aims at learning general-purpose neural network representations from various types of spatial data (e.g., points, polylines, polygons, networks, images, etc.) in their native formats. Learning good spatial representations is a fundamental problem for various downstream applications such as species distribution modeling, weather forecasting, trajectory generation, geographic question answering, etc. Even though SRL has become the foundation of almost all geospatial artificial intelligence (GeoAI) research, we have not yet seen significant efforts to develop an extensive deep learning framework and benchmark to support SRL model development and evaluation. To fill this gap, we propose TorchSpatial, a learning framework and benchmark for location (point) encoding, which is one of the most fundamental data types of spatial representation learning. TorchSpatial contains three key components: 1) a unified location encoding framework that consolidates 15 commonly recognized location encoders, ensuring scalability and reproducibility of the implementations; 2) the LocBench benchmark tasks encompassing 7 geo-aware image classification and 10 geo-aware image regression datasets; 3) a comprehensive suite of evaluation metrics to quantify geo-aware model's overall performance as well as their geographic bias, with a novel Geo-Bias Score metric. Finally, we provide a detailed analysis and insights into the model performance and geographic bias of different location encoders. We believe TorchSpatial will foster future advancement of spatial representation learning and spatial fairness in GeoAI research. The TorchSpatial model framework and LocBench benchmark are available at https://github.com/seai-lab/TorchSpatial, and the Geo-Bias Score evaluation framework is available at https://github.com/seai-lab/PyGBS.
Multimodal Large Language Models (MLLMs) have shown significant promise in various applications, leading to broad interest from researchers and practitioners alike. However, a comprehensive evaluation of their long-context capabilities remains underexplored. To address these gaps, we introduce the MultiModal Needle-in-a-haystack (MMNeedle) benchmark, specifically designed to assess the long-context capabilities of MLLMs. Besides multi-image input, we employ image stitching to further increase the input context length, and develop a protocol to automatically generate labels for sub-image level retrieval. Essentially, MMNeedle evaluates MLLMs by stress-testing their capability to locate a target sub-image (needle) within a set of images (haystack) based on textual instructions and descriptions of image contents. This setup necessitates an advanced understanding of extensive visual contexts and effective information retrieval within long-context image inputs. With this benchmark, we evaluate state-of-the-art MLLMs, encompassing both API-based and open-source models. The findings reveal that GPT-4o consistently surpasses other models in long-context scenarios, but suffers from hallucination problems in negative samples, i.e., when needles are not in the haystacks. Our comprehensive long-context evaluation of MLLMs also sheds lights on the considerable performance gap between API-based and open-source models. All the code, data, and instructions required to reproduce the main results are available at https://github.com/Wang-ML-Lab/multimodal-needle-in-a-haystack.
Mercy Ranjit, Gopinath Ganapathy, Shaury Srivastav et al.
Small Language Models (SLMs) have shown remarkable performance in general domain language understanding, reasoning and coding tasks, but their capabilities in the medical domain, particularly concerning radiology text, is less explored. In this study, we investigate the application of SLMs for general radiology knowledge specifically question answering related to understanding of symptoms, radiological appearances of findings, differential diagnosis, assessing prognosis, and suggesting treatments w.r.t diseases pertaining to different organ systems. Additionally, we explore the utility of SLMs in handling text-related tasks with respect to radiology reports within AI-driven radiology workflows. We fine-tune Phi-2, a SLM with 2.7 billion parameters using high-quality educational content from Radiopaedia, a collaborative online radiology resource. The resulting language model, RadPhi-2-Base, exhibits the ability to address general radiology queries across various systems (e.g., chest, cardiac). Furthermore, we investigate Phi-2 for instruction tuning, enabling it to perform specific tasks. By fine-tuning Phi-2 on both general domain tasks and radiology-specific tasks related to chest X-ray reports, we create Rad-Phi2. Our empirical results reveal that Rad-Phi2 Base and Rad-Phi2 perform comparably or even outperform larger models such as Mistral-7B-Instruct-v0.2 and GPT-4 providing concise and precise answers. In summary, our work demonstrates the feasibility and effectiveness of utilizing SLMs in radiology workflows both for knowledge related queries as well as for performing specific tasks related to radiology reports thereby opening up new avenues for enhancing the quality and efficiency of radiology practice.
LLM based copilot assistants are useful in everyday tasks. There is a proliferation in the exploration of AI assistant use cases to support radiology workflows in a reliable manner. In this work, we present RadPhi-3, a Small Language Model instruction tuned from Phi-3-mini-4k-instruct with 3.8B parameters to assist with various tasks in radiology workflows. While impression summary generation has been the primary task which has been explored in prior works w.r.t radiology reports of Chest X-rays, we also explore other useful tasks like change summary generation comparing the current radiology report and its prior report, section extraction from radiology reports, tagging the reports with various pathologies and tubes, lines or devices present in them etc. In-addition, instruction tuning RadPhi-3 involved learning from a credible knowledge source used by radiologists, Radiopaedia.org. RadPhi-3 can be used both to give reliable answers for radiology related queries as well as perform useful tasks related to radiology reports. RadPhi-3 achieves SOTA results on the RaLEs radiology report generation benchmark.
Somnath Kumar, Vaibhav Balloli, Mercy Ranjit et al.
Large language models (LLMs) have revolutionized various domains but still struggle with non-Latin scripts and low-resource languages. This paper addresses the critical challenge of improving multilingual performance without extensive fine-tuning. We introduce a novel dynamic learning approach that optimizes prompt strategy, embedding model, and LLM per query at runtime. By adapting configurations dynamically, our method achieves significant improvements over static, best and random baselines. It operates efficiently in both offline and online settings, generalizing seamlessly across new languages and datasets. Leveraging Retrieval-Augmented Generation (RAG) with state-of-the-art multilingual embeddings, we achieve superior task performance across diverse linguistic contexts. Through systematic investigation and evaluation across 18 diverse languages using popular question-answering (QA) datasets we show our approach results in 10-15% improvements in multilingual performance over pre-trained models and 4x gains compared to fine-tuned, language-specific models.
Mercy Ranjit, Gopinath Ganapathy, Ranjit Manuel et al.
We propose Retrieval Augmented Generation (RAG) as an approach for automated radiology report writing that leverages multimodally aligned embeddings from a contrastively pretrained vision language model for retrieval of relevant candidate radiology text for an input radiology image and a general domain generative model like OpenAI text-davinci-003, gpt-3.5-turbo and gpt-4 for report generation using the relevant radiology text retrieved. This approach keeps hallucinated generations under check and provides capabilities to generate report content in the format we desire leveraging the instruction following capabilities of these generative models. Our approach achieves better clinical metrics with a BERTScore of 0.2865 (Δ+ 25.88%) and Semb score of 0.4026 (Δ+ 6.31%). Our approach can be broadly relevant for different clinical settings as it allows to augment the automated radiology report generation process with content relevant for that setting while also having the ability to inject user intents and requirements in the prompts as part of the report generation process to modulate the content and format of the generated reports as applicable for that clinical setting.
Anirudh Srinivasan, Sunayana Sitaram, Tanuja Ganu et al.
Recent advancements in NLP have given us models like mBERT and XLMR that can serve over 100 languages. The languages that these models are evaluated on, however, are very few in number, and it is unlikely that evaluation datasets will cover all the languages that these models support. Potential solutions to the costly problem of dataset creation are to translate datasets to new languages or use template-filling based techniques for creation. This paper proposes an alternate solution for evaluating a model across languages which make use of the existing performance scores of the model on languages that a particular task has test sets for. We train a predictor on these performance scores and use this predictor to predict the model's performance in different evaluation settings. Our results show that our method is effective in filling the gaps in the evaluation for an existing set of languages, but might require additional improvements if we want it to generalize to unseen languages.
Anurag Ghosh, Vaibhav Balloli, Akshay Nambi et al.
Real-time perception requires planned resource utilization. Computational planning in real-time perception is governed by two considerations -- accuracy and latency. There exist run-time decisions (e.g. choice of input resolution) that induce tradeoffs affecting performance on a given hardware, arising from intrinsic (content, e.g. scene clutter) and extrinsic (system, e.g. resource contention) characteristics. Earlier runtime execution frameworks employed rule-based decision algorithms and operated with a fixed algorithm latency budget to balance these concerns, which is sub-optimal and inflexible. We propose Chanakya, a learned approximate execution framework that naturally derives from the streaming perception paradigm, to automatically learn decisions induced by these tradeoffs instead. Chanakya is trained via novel rewards balancing accuracy and latency implicitly, without approximating either objectives. Chanakya simultaneously considers intrinsic and extrinsic context, and predicts decisions in a flexible manner. Chanakya, designed with low overhead in mind, outperforms state-of-the-art static and dynamic execution policies on public datasets on both server GPUs and edge devices.
In the context of the ongoing Covid-19 pandemic, several reports and studies have attempted to model and predict the spread of the disease. There is also intense debate about policies for limiting the damage, both to health and to the economy. On the one hand, the health and safety of the population is the principal consideration for most countries. On the other hand, we cannot ignore the potential for long-term economic damage caused by strict nation-wide lockdowns. In this working paper, we present a quantitative way to compute lockdown decisions for individual cities or regions, while balancing health and economic considerations. Furthermore, these policies are learnt automatically by the proposed algorithm, as a function of disease parameters (infectiousness, gestation period, duration of symptoms, probability of death) and population characteristics (density, movement propensity). We account for realistic considerations such as imperfect lockdowns, and show that the policy obtained using reinforcement learning is a viable quantitative approach towards lockdowns.