Samuel Stevens

Xiang Yue, Yuansheng Ni, Kai Zhang et al. · microsoft-research, princeton

We introduce MMMU: a new benchmark designed to evaluate multimodal models on massive multi-discipline tasks demanding college-level subject knowledge and deliberate reasoning. MMMU includes 11.5K meticulously collected multimodal questions from college exams, quizzes, and textbooks, covering six core disciplines: Art & Design, Business, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering. These questions span 30 subjects and 183 subfields, comprising 30 highly heterogeneous image types, such as charts, diagrams, maps, tables, music sheets, and chemical structures. Unlike existing benchmarks, MMMU focuses on advanced perception and reasoning with domain-specific knowledge, challenging models to perform tasks akin to those faced by experts. The evaluation of 14 open-source LMMs as well as the proprietary GPT-4V(ision) and Gemini highlights the substantial challenges posed by MMMU. Even the advanced GPT-4V and Gemini Ultra only achieve accuracies of 56% and 59% respectively, indicating significant room for improvement. We believe MMMU will stimulate the community to build next-generation multimodal foundation models towards expert artificial general intelligence.

Xiang Deng, Yu Gu, Boyuan Zheng et al. · microsoft-research

We introduce Mind2Web, the first dataset for developing and evaluating generalist agents for the web that can follow language instructions to complete complex tasks on any website. Existing datasets for web agents either use simulated websites or only cover a limited set of websites and tasks, thus not suitable for generalist web agents. With over 2,000 open-ended tasks collected from 137 websites spanning 31 domains and crowdsourced action sequences for the tasks, Mind2Web provides three necessary ingredients for building generalist web agents: 1) diverse domains, websites, and tasks, 2) use of real-world websites instead of simulated and simplified ones, and 3) a broad spectrum of user interaction patterns. Based on Mind2Web, we conduct an initial exploration of using large language models (LLMs) for building generalist web agents. While the raw HTML of real-world websites are often too large to be fed to LLMs, we show that first filtering it with a small LM significantly improves the effectiveness and efficiency of LLMs. Our solution demonstrates a decent level of performance, even on websites or entire domains the model has never seen before, but there is still a substantial room to improve towards truly generalizable agents. We open-source our dataset, model implementation, and trained models (https://osu-nlp-group.github.io/Mind2Web) to facilitate further research on building a generalist agent for the web.

Dipanjyoti Paul, Arpita Chowdhury, Xinqi Xiong et al. · microsoft-research

We present a novel usage of Transformers to make image classification interpretable. Unlike mainstream classifiers that wait until the last fully connected layer to incorporate class information to make predictions, we investigate a proactive approach, asking each class to search for itself in an image. We realize this idea via a Transformer encoder-decoder inspired by DEtection TRansformer (DETR). We learn "class-specific" queries (one for each class) as input to the decoder, enabling each class to localize its patterns in an image via cross-attention. We name our approach INterpretable TRansformer (INTR), which is fairly easy to implement and exhibits several compelling properties. We show that INTR intrinsically encourages each class to attend distinctively; the cross-attention weights thus provide a faithful interpretation of the prediction. Interestingly, via "multi-head" cross-attention, INTR could identify different "attributes" of a class, making it particularly suitable for fine-grained classification and analysis, which we demonstrate on eight datasets. Our code and pre-trained models are publicly accessible at the Imageomics Institute GitHub site: https://github.com/Imageomics/INTR.

Lingbo Mo, Shijie Chen, Ziru Chen et al. · microsoft-research

We introduce TacoBot, a user-centered task-oriented digital assistant designed to guide users through complex real-world tasks with multiple steps. Covering a wide range of cooking and how-to tasks, we aim to deliver a collaborative and engaging dialogue experience. Equipped with language understanding, dialogue management, and response generation components supported by a robust search engine, TacoBot ensures efficient task assistance. To enhance the dialogue experience, we explore a series of data augmentation strategies using LLMs to train advanced neural models continuously. TacoBot builds upon our successful participation in the inaugural Alexa Prize TaskBot Challenge, where our team secured third place among ten competing teams. We offer TacoBot as an open-source framework that serves as a practical example for deploying task-oriented dialogue systems.

Samuel Stevens, Jiaman Wu, Matthew J Thompson et al. · microsoft-research

Images of the natural world, collected by a variety of cameras, from drones to individual phones, are increasingly abundant sources of biological information. There is an explosion of computational methods and tools, particularly computer vision, for extracting biologically relevant information from images for science and conservation. Yet most of these are bespoke approaches designed for a specific task and are not easily adaptable or extendable to new questions, contexts, and datasets. A vision model for general organismal biology questions on images is of timely need. To approach this, we curate and release TreeOfLife-10M, the largest and most diverse ML-ready dataset of biology images. We then develop BioCLIP, a foundation model for the tree of life, leveraging the unique properties of biology captured by TreeOfLife-10M, namely the abundance and variety of images of plants, animals, and fungi, together with the availability of rich structured biological knowledge. We rigorously benchmark our approach on diverse fine-grained biology classification tasks and find that BioCLIP consistently and substantially outperforms existing baselines (by 16% to 17% absolute). Intrinsic evaluation reveals that BioCLIP has learned a hierarchical representation conforming to the tree of life, shedding light on its strong generalizability. https://imageomics.github.io/bioclip has models, data and code.

Chao Jiang, Wei Xu, Samuel Stevens · gatech

Scientific publications are the primary means to communicate research discoveries, where the writing quality is of crucial importance. However, prior work studying the human editing process in this domain mainly focused on the abstract or introduction sections, resulting in an incomplete picture. In this work, we provide a complete computational framework for studying text revision in scientific writing. We first introduce arXivEdits, a new annotated corpus of 751 full papers from arXiv with gold sentence alignment across their multiple versions of revision, as well as fine-grained span-level edits and their underlying intentions for 1,000 sentence pairs. It supports our data-driven analysis to unveil the common strategies practiced by researchers for revising their papers. To scale up the analysis, we also develop automatic methods to extract revision at document-, sentence-, and word-levels. A neural CRF sentence alignment model trained on our corpus achieves 93.8 F1, enabling the reliable matching of sentences between different versions. We formulate the edit extraction task as a span alignment problem, and our proposed method extracts more fine-grained and explainable edits, compared to the commonly used diff algorithm. An intention classifier trained on our dataset achieves 78.9 F1 on the fine-grained intent classification task. Our data and system are released at tiny.one/arxivedits.

Shijie Chen, Ziru Chen, Xiang Deng et al. · microsoft-research

We present TacoBot, a task-oriented dialogue system built for the inaugural Alexa Prize TaskBot Challenge, which assists users in completing multi-step cooking and home improvement tasks. TacoBot is designed with a user-centered principle and aspires to deliver a collaborative and accessible dialogue experience. Towards that end, it is equipped with accurate language understanding, flexible dialogue management, and engaging response generation. Furthermore, TacoBot is backed by a strong search engine and an automated end-to-end test suite. In bootstrapping the development of TacoBot, we explore a series of data augmentation strategies to train advanced neural language processing models and continuously improve the dialogue experience with collected real conversations. At the end of the semifinals, TacoBot achieved an average rating of 3.55/5.0.

Vardaan Pahuja, Samuel Stevens, Alyson East et al.

Morphological traits are physical characteristics of biological organisms that provide vital clues on how organisms interact with their environment. Yet extracting these traits remains a slow, expert-driven process, limiting their use in large-scale ecological studies. A major bottleneck is the absence of high-quality datasets linking biological images to trait-level annotations. In this work, we demonstrate that sparse autoencoders trained on foundation-model features yield monosemantic, spatially grounded neurons that consistently activate on meaningful morphological parts. Leveraging this property, we introduce a trait annotation pipeline that localizes salient regions and uses vision-language prompting to generate interpretable trait descriptions. Using this approach, we construct Bioscan-Traits, a dataset of 80K trait annotations spanning 19K insect images from BIOSCAN-5M. Human evaluation confirms the biological plausibility of the generated morphological descriptions. We assess design sensitivity through a comprehensive ablation study, systematically varying key design choices and measuring their impact on the quality of the resulting trait descriptions. By annotating traits with a modular pipeline rather than prohibitively expensive manual efforts, we offer a scalable way to inject biologically meaningful supervision into foundation models, enable large-scale morphological analyses, and bridge the gap between ecological relevance and machine-learning practicality.

Fangxun Liu, S M Rayeed, Samuel Stevens et al.

In entomology and ecology research, biologists often need to collect a large number of insects, among which beetles are the most common species. A common practice for biologists to organize beetles is to place them on trays and take a picture of each tray. Given the images of thousands of such trays, it is important to have an automated pipeline to process the large-scale data for further research. Therefore, we develop a 3-stage pipeline to detect all the beetles on each tray, sort and crop the image of each beetle, and do morphological segmentation on the cropped beetles. For detection, we design an iterative process utilizing a transformer-based open-vocabulary object detector and a vision-language model. For segmentation, we manually labeled 670 beetle images and fine-tuned two variants of a transformer-based segmentation model to achieve fine-grained segmentation of beetles with relatively high accuracy. The pipeline integrates multiple deep learning methods and is specialized for beetle image processing, which can greatly improve the efficiency to process large-scale beetle data and accelerate biological research.

Samuel Stevens

ImageNet-1K linear-probe transfer accuracy remains the default proxy for visual representation quality, yet it no longer predicts performance on scientific imagery. Across 46 modern vision model checkpoints, ImageNet top-1 accuracy explains only 34% of variance on ecology tasks and mis-ranks 30% of models above 75% accuracy. We present BioBench, an open ecology vision benchmark that captures what ImageNet misses. BioBench unifies 9 publicly released, application-driven tasks, 4 taxonomic kingdoms, and 6 acquisition modalities (drone RGB, web video, micrographs, in-situ and specimen photos, camera-trap frames), totaling 3.1M images. A single Python API downloads data, fits lightweight classifiers to frozen backbones, and reports class-balanced macro-F1 (plus domain metrics for FishNet and FungiCLEF); ViT-L models evaluate in 6 hours on an A6000 GPU. BioBench provides new signal for computer vision in ecology and a template recipe for building reliable AI-for-science benchmarks in any domain. Code and predictions are available at https://github.com/samuelstevens/biobench and results at https://samuelstevens.me/biobench.

Jenna Kline, Maksim Kholiavchenko, Samuel Stevens et al.

A comprehensive understanding of animal behavior ecology depends on scalable approaches to quantify and interpret complex, multidimensional behavioral patterns. Traditional field observations are often limited in scope, time-consuming, and labor-intensive, hindering the assessment of behavioral responses across landscapes. To address this, we present kabr-tools (Kenyan Animal Behavior Recognition Tools), an open-source package for automated multi-species behavioral monitoring. This framework integrates drone-based video with machine learning systems to extract behavioral, social, and spatial metrics from wildlife footage. Our pipeline leverages object detection, tracking, and behavioral classification systems to generate key metrics, including time budgets, behavioral transitions, social interactions, habitat associations, and group composition dynamics. Compared to ground-based methods, drone-based observations significantly improved behavioral granularity, reducing visibility loss by 15% and capturing more transitions with higher accuracy and continuity. We validate kabr-tools through three case studies, analyzing 969 behavioral sequences, surpassing the capacity of traditional methods for data capture and annotation. We found that, like Plains zebras, vigilance in Grevy's zebras decreases with herd size, but, unlike Plains zebras, habitat has a negligible impact. Plains and Grevy's zebras exhibit strong behavioral inertia, with rare transitions to alert behaviors and observed spatial segregation between Grevy's zebras, Plains zebras, and giraffes in mixed-species herds. By enabling automated behavioral monitoring at scale, kabr-tools offers a powerful tool for ecosystem-wide studies, advancing conservation, biodiversity research, and ecological monitoring.

Samuel Stevens, Yu Su

Over-parameterized neural language models (LMs) can memorize and recite long sequences of training data. While such memorization is normally associated with undesired properties such as overfitting and information leaking, our work casts memorization as an unexplored capability of LMs. We propose the first symmetric encryption algorithm with autoregressive language models (SELM). We show that autoregressive LMs can encode arbitrary data into a compact real-valued vector (i.e., encryption) and then losslessly decode the vector to the original message (i.e., decryption) via random subspace optimization and greedy decoding. While SELM is not amenable to conventional cryptanalysis, we investigate its security through a novel empirical variant of the classic IND-CPA (indistinguishability under chosen-plaintext attack) game and show promising results on security. Our code and datasets are available at https://github.com/OSU-NLP-Group/SELM.

Samuel Stevens, Yu Su

Pre-trained language models (PLMs) like BERT are being used for almost all language-related tasks, but interpreting their behavior still remains a significant challenge and many important questions remain largely unanswered. In this work, we re-purpose a sentence editing dataset, where faithful high-quality human rationales can be automatically extracted and compared with extracted model rationales, as a new testbed for interpretability. This enables us to conduct a systematic investigation on an array of questions regarding PLMs' interpretability, including the role of pre-training procedure, comparison of rationale extraction methods, and different layers in the PLM. The investigation generates new insights, for example, contrary to the common understanding, we find that attention weights correlate well with human rationales and work better than gradient-based saliency in extracting model rationales. Both the dataset and code are available at https://github.com/samuelstevens/sentence-editing-interpretability to facilitate future interpretability research.

Samuel Stevens, Wei-Lun Chao, Tanya Berger-Wolf et al. · microsoft-research

To truly understand vision models, we must not only interpret their learned features but also validate these interpretations through controlled experiments. Current approaches either provide interpretable features without the ability to test their causal influence, or enable model editing without interpretable controls. We present a unified framework using sparse autoencoders (SAEs) that bridges this gap, allowing us to discover human-interpretable visual features and precisely manipulate them to test hypotheses about model behavior. By applying our method to state-of-the-art vision models, we reveal key differences in the semantic abstractions learned by models with different pre-training objectives. We then demonstrate the practical usage of our framework through controlled interventions across multiple vision tasks. We show that SAEs can reliably identify and manipulate interpretable visual features without model re-training, providing a powerful tool for understanding and controlling vision model behavior. We provide code, demos and models on our project website: https://osu-nlp-group.github.io/SAE-V.

Samuel Stevens, Emily Wenger, Cathy Li et al.

Learning with Errors (LWE) is a hard math problem underlying recently standardized post-quantum cryptography (PQC) systems for key exchange and digital signatures. Prior work proposed new machine learning (ML)-based attacks on LWE problems with small, sparse secrets, but these attacks require millions of LWE samples to train on and take days to recover secrets. We propose three key methods -- better preprocessing, angular embeddings and model pre-training -- to improve these attacks, speeding up preprocessing by $25\times$ and improving model sample efficiency by $10\times$. We demonstrate for the first time that pre-training improves and reduces the cost of ML attacks on LWE. Our architecture improvements enable scaling to larger-dimension LWE problems: this work is the first instance of ML attacks recovering sparse binary secrets in dimension $n=1024$, the smallest dimension used in practice for homomorphic encryption applications of LWE where sparse binary secrets are proposed.

Jianyang Gu, Samuel Stevens, Elizabeth G Campolongo et al. · microsoft-research

Foundation models trained at scale exhibit remarkable emergent behaviors, learning new capabilities beyond their initial training objectives. We find such emergent behaviors in biological vision models via large-scale contrastive vision-language training. To achieve this, we first curate TreeOfLife-200M, comprising 214 million images of living organisms, the largest and most diverse biological organism image dataset to date. We then train BioCLIP 2 on TreeOfLife-200M to distinguish different species. Despite the narrow training objective, BioCLIP 2 yields extraordinary accuracy when applied to various biological visual tasks such as habitat classification and trait prediction. We identify emergent properties in the learned embedding space of BioCLIP 2. At the inter-species level, the embedding distribution of different species aligns closely with functional and ecological meanings (e.g., beak sizes and habitats). At the intra-species level, instead of being diminished, the intra-species variations (e.g., life stages and sexes) are preserved and better separated in subspaces orthogonal to inter-species distinctions. We provide formal proof and analyses to explain why hierarchical supervision and contrastive objectives encourage these emergent properties. Crucially, our results reveal that these properties become increasingly significant with larger-scale training data, leading to a biologically meaningful embedding space.

Alyson East, Elizabeth G. Campolongo, Luke Meyers et al.

1) Biological collections house millions of specimens with digital images increasingly available through open-access platforms. However, most imaging protocols were developed for human interpretation without considering automated analysis requirements. As computer vision applications revolutionize taxonomic identification and trait extraction, a critical gap exists between current digitization practices and computational analysis needs. This review provides the first comprehensive practical framework for optimizing biological specimen imaging for computer vision applications. 2) Through interdisciplinary collaboration between taxonomists, collection managers, ecologists, and computer scientists, we synthesized evidence-based recommendations addressing fundamental computer vision concepts and practical imaging considerations. We provide immediately actionable implementation guidance while identifying critical areas requiring community standards development. 3) Our framework encompasses ten interconnected considerations for optimizing image capture for computer vision-powered taxonomic identification and trait extraction. We translate these into practical implementation checklists, equipment selection guidelines, and a roadmap for community standards development including filename conventions, pixel density requirements, and cross-institutional protocols. 4)By bridging biological and computational disciplines, this approach unlocks automated analysis potential for millions of existing specimens and guides future digitization efforts toward unprecedented analytical capabilities.

Samuel Stevens, S M Rayeed, Jenna Kline

The practical application of AI tools for specific computer vision tasks relies on the "small-data regime" of hundreds to thousands of labeled samples. This small-data regime is vital for applications requiring expensive expert annotations, such as ecological monitoring, medical diagnostics or industrial quality control. We find, however, that computer vision research has ignored the small data regime as evaluations increasingly focus on zero- and few-shot learning. We use the Natural World Tasks (NeWT) benchmark to compare multi-modal large language models (MLLMs) and vision-only methods across varying training set sizes. MLLMs exhibit early performance plateaus, while vision-only methods improve throughout the small-data regime, with performance gaps widening beyond 10 training examples. We provide the first comprehensive comparison between these approaches in small-data contexts and advocate for explicit small-data evaluations in AI research to better bridge theoretical advances with practical deployments.

S M Rayeed, Alyson East, Samuel Stevens et al.

Ground beetles are a highly sensitive and speciose biological indicator, making them vital for monitoring biodiversity. However, they are currently an underutilized resource due to the manual effort required by taxonomic experts to perform challenging species differentiations based on subtle morphological differences, precluding widespread applications. In this paper, we evaluate 12 vision models on taxonomic classification across four diverse, long-tailed datasets spanning over 230 genera and 1769 species, with images ranging from controlled laboratory settings to challenging field-collected (in-situ) photographs. We further explore taxonomic classification in two important real-world contexts: sample efficiency and domain adaptation. Our results show that the Vision and Language Transformer combined with an MLP head is the best performing model, with 97% accuracy at genus and 94% at species level. Sample efficiency analysis shows that we can reduce train data requirements by up to 50% with minimal compromise in performance. The domain adaptation experiments reveal significant challenges when transferring models from lab to in-situ images, highlighting a critical domain gap. Overall, our study lays a foundation for large-scale automated taxonomic classification of beetles, and beyond that, advances sample-efficient learning and cross-domain adaptation for diverse long-tailed ecological datasets.

S M Rayeed, Mridul Khurana, Alyson East et al.

Despite the ecological significance of invertebrates, global trait databases remain heavily biased toward vertebrates and plants, limiting comprehensive ecological analyses of high-diversity groups like ground beetles. Ground beetles (Coleoptera: Carabidae) serve as critical bioindicators of ecosystem health, providing valuable insights into biodiversity shifts driven by environmental changes. While the National Ecological Observatory Network (NEON) maintains an extensive collection of carabid specimens from across the United States, these primarily exist as physical collections, restricting widespread research access and large-scale analysis. To address these gaps, we present a multimodal dataset digitizing over 13,200 NEON carabids from 30 sites spanning the continental US and Hawaii through high-resolution imaging, enabling broader access and computational analysis. The dataset includes digitally measured elytra length and width of each specimen, establishing a foundation for automated trait extraction using AI. Validated against manual measurements, our digital trait extraction achieves sub-millimeter precision, ensuring reliability for ecological and computational studies. By addressing invertebrate under-representation in trait databases, this work supports AI-driven tools for automated species identification and trait-based research, fostering advancements in biodiversity monitoring and conservation.

Samuel Stevens, Jacob Beattie, Tanya Berger-Wolf et al.

Scientific archives now contain hundreds of petabytes of data across genomics, ecology, climate, and molecular biology that could reveal undiscovered patterns if systematically analyzed at scale. Large-scale, weakly-supervised datasets in language and vision have driven the development of foundation models whose internal representations encode structure (patterns, co-occurrences and statistical regularities) beyond their training objectives. Most existing methods extract structure only for pre-specified targets; they excel at confirmation but do not support open-ended discovery of unknown patterns. We ask whether sparse autoencoders (SAEs) can enable open-ended feature discovery from foundation model representations. We evaluate this question in controlled rediscovery studies, where the learned SAE features are tested for alignment with semantic concepts on a standard segmentation benchmark and compared against strong label-free alternatives on concept-alignment metrics. Applied to ecological imagery, the same procedure surfaces fine-grained anatomical structure without access to segmentation or part labels, providing a scientific case study with ground-truth validation. While our experiments focus on vision with an ecology case study, the method is domain-agnostic and applicable to models in other sciences (e.g., proteins, genomics, weather). Our results indicate that sparse decomposition provides a practical instrument for exploring what scientific foundation models have learned, an important prerequisite for moving from confirmation to genuine discovery.

Ziheng Zhang, Xinyue Ma, Arpita Chowdhury et al.

This work investigates descriptive captions as an additional source of supervision for biological multimodal foundation models. Images and captions can be viewed as complementary samples from the latent morphospace of a species, each capturing certain biological traits. Incorporating captions during training encourages alignment with this shared latent structure, emphasizing potentially diagnostic characters while suppressing spurious correlations. The main challenge, however, lies in obtaining faithful, instance-specific captions at scale. This requirement has limited the utilization of natural language supervision in organismal biology compared with many other scientific domains. We complement this gap by generating synthetic captions with multimodal large language models (MLLMs), guided by Wikipedia-derived visual information and taxon-tailored format examples. These domain-specific contexts help reduce hallucination and yield accurate, instance-based descriptive captions. Using these captions, we train BioCAP (i.e., BioCLIP with Captions), a biological foundation model that captures rich semantics and achieves strong performance in species classification and text-image retrieval. These results demonstrate the value of descriptive captions beyond labels in bridging biological images with multimodal foundation models.

Jenna Kline, Samuel Stevens, Guy Maalouf et al.

Real-time wildlife detection in drone imagery supports critical ecological and conservation monitoring. However, standard detection models like YOLO often fail to generalize across locations and struggle with rare species, limiting their use in automated drone deployments. We present MMLA, a novel multi-environment, multi-species, low-altitude drone dataset collected across three sites (Ol Pejeta Conservancy and Mpala Research Centre in Kenya, and The Wilds in Ohio), featuring six species (zebras, giraffes, onagers, and African wild dogs). The dataset contains 811K annotations from 37 high-resolution videos. Baseline YOLO models show performance disparities across locations while fine-tuning YOLOv11m on MMLA improves mAP50 to 82%, a 52-point gain over baseline. Our results underscore the need for diverse training data to enable robust animal detection in autonomous drone systems.