Justin Lee

Tan H. Nguyen, Dinkar Juyal, Jin Li et al.

Differences in staining and imaging procedures can cause significant color variations in histopathology images, leading to poor generalization when deploying deep-learning models trained from a different data source. Various color augmentation methods have been proposed to generate synthetic images during training to make models more robust, eliminating the need for stain normalization during test time. Many color augmentation methods leverage domain labels to generate synthetic images. This approach causes three significant challenges to scaling such a model. Firstly, incorporating data from a new domain into deep-learning models trained on existing domain labels is not straightforward. Secondly, dependency on domain labels prevents the use of pathology images without domain labels to improve model performance. Finally, implementation of these methods becomes complicated when multiple domain labels (e.g., patient identification, medical center, etc) are associated with a single image. We introduce ContriMix, a novel domain label free stain color augmentation method based on DRIT++, a style-transfer method. Contrimix leverages sample stain color variation within a training minibatch and random mixing to extract content and attribute information from pathology images. This information can be used by a trained ContriMix model to create synthetic images to improve the performance of existing classifiers. ContriMix outperforms competing methods on the Camelyon17-WILDS dataset. Its performance is consistent across different slides in the test set while being robust to the color variation from rare substances in pathology images. We make our code and trained ContriMix models available for research use. The code for ContriMix can be found at https://gitlab.com/huutan86/contrimix

Justin Lee, Sowmya Vajjala

Automatic Readability Assessment (ARA), the task of assigning a reading level to a text, is traditionally treated as a classification problem in NLP research. In this paper, we propose the first neural, pairwise ranking approach to ARA and compare it with existing classification, regression, and (non-neural) ranking methods. We establish the performance of our model by conducting experiments with three English, one French and one Spanish datasets. We demonstrate that our approach performs well in monolingual single/cross corpus testing scenarios and achieves a zero-shot cross-lingual ranking accuracy of over 80% for both French and Spanish when trained on English data. Additionally, we also release a new parallel bilingual readability dataset in English and French. To our knowledge, this paper proposes the first neural pairwise ranking model for ARA, and shows the first results of cross-lingual, zero-shot evaluation of ARA with neural models.

Justin Lee, Tuomas Oikarinen, Arjun Chatha et al.

Recent advances have greatly increased the capabilities of large language models (LLMs), but our understanding of the models and their safety has not progressed as fast. In this paper we aim to understand LLMs deeper by studying their individual neurons. We build upon previous work showing large language models such as GPT-4 can be useful in explaining what each neuron in a language model does. Specifically, we analyze the effect of the prompt used to generate explanations and show that reformatting the explanation prompt in a more natural way can significantly improve neuron explanation quality and greatly reduce computational cost. We demonstrate the effects of our new prompts in three different ways, incorporating both automated and human evaluations.

Jihyung Kil, Zheda Mai, Justin Lee et al.

The ability to compare objects, scenes, or situations is crucial for effective decision-making and problem-solving in everyday life. For instance, comparing the freshness of apples enables better choices during grocery shopping while comparing sofa designs helps optimize the aesthetics of our living space. Despite its significance, the comparative capability is largely unexplored in artificial general intelligence (AGI). In this paper, we introduce MLLM-CompBench, a benchmark designed to evaluate the comparative reasoning capability of multimodal large language models (MLLMs). MLLM-CompBench mines and pairs images through visually oriented questions covering eight dimensions of relative comparison: visual attribute, existence, state, emotion, temporality, spatiality, quantity, and quality. We curate a collection of around 40K image pairs using metadata from diverse vision datasets and CLIP similarity scores. These image pairs span a broad array of visual domains, including animals, fashion, sports, and both outdoor and indoor scenes. The questions are carefully crafted to discern relative characteristics between two images and are labeled by human annotators for accuracy and relevance. We use MLLM-CompBench to evaluate recent MLLMs, including GPT-4V(ision), Gemini-Pro, and LLaVA-1.6. Our results reveal notable shortcomings in their comparative abilities. We believe MLLM-COMPBENCH not only sheds light on these limitations but also establishes a solid foundation for future enhancements in the comparative capability of MLLMs.

Mridul Khurana, Amin Karimi Monsefi, Justin Lee et al.

Accurately generating images across the Tree of Life is difficult: there are over 10M distinct species on Earth, many of which differ only by subtle visual traits. Despite the remarkable progress in text-to-image synthesis, existing models often fail to capture the fine-grained visual cues that define species identity, even when their outputs appear photo-realistic. To this end, we propose TaxaAdapter, a simple and lightweight approach that incorporates Vision Taxonomy Models (VTMs) such as BioCLIP to guide fine-grained species generation. Our method injects VTM embeddings into a frozen text-to-image diffusion model, improving species-level fidelity while preserving flexible text control over attributes such as pose, style, and background. Extensive experiments demonstrate that TaxaAdapter consistently improves morphology fidelity and species-identity accuracy over strong baselines, with a cleaner architecture and training recipe. To better evaluate these improvements, we also introduce a multimodal Large Language Model-based metric that summarizes trait-level descriptions from generated and real images, providing a more interpretable measure of morphological consistency. Beyond this, we observe that TaxaAdapter exhibits strong generalization capabilities, enabling species synthesis in challenging regimes such as few-shot species with only a handful of training images and even species unseen during training. Overall, our results highlight that VTMs are a key ingredient for scalable, fine-grained species generation.

Mustafa Eyceoz, Justin Lee, Siddharth Pittie et al.

Most state-of-the-art spoken language identification models are closed-set; in other words, they can only output a language label from the set of classes they were trained on. Open-set spoken language identification systems, however, gain the ability to detect when an input exhibits none of the original languages. In this paper, we implement a novel approach to open-set spoken language identification that uses MFCC and pitch features, a TDNN model to extract meaningful feature embeddings, confidence thresholding on softmax outputs, and LDA and pLDA for learning to classify new unknown languages. We present a spoken language identification system that achieves 91.76% accuracy on trained languages and has the capability to adapt to unknown languages on the fly. To that end, we also built the CU MultiLang Dataset, a large and diverse multilingual speech corpus which was used to train and evaluate our system.

Mustafa Eyceoz, Justin Lee, Homayoon Beigi

While most modern speech Language Identification methods are closed-set, we want to see if they can be modified and adapted for the open-set problem. When switching to the open-set problem, the solution gains the ability to reject an audio input when it fails to match any of our known language options. We tackle the open-set task by adapting two modern-day state-of-the-art approaches to closed-set language identification: the first using a CRNN with attention and the second using a TDNN. In addition to enhancing our input feature embeddings using MFCCs, log spectral features, and pitch, we will be attempting two approaches to out-of-set language detection: one using thresholds, and the other essentially performing a verification task. We will compare both the performance of the TDNN and the CRNN, as well as our detection approaches.

Justin Lee, Zheda Mai, Jinsu Yoo et al.

Machine unlearning--the ability to remove designated concepts from a pre-trained model--has advanced rapidly, particularly for text-to-image diffusion models. However, existing methods typically assume that unlearning requests arrive all at once, whereas in practice they often arrive sequentially. We present the first systematic study of continual unlearning in text-to-image diffusion models and show that popular unlearning methods suffer from rapid utility collapse: after only a few requests, models forget retained knowledge and generate degraded images. We trace this failure to cumulative parameter drift from the pre-training weights and argue that regularization is crucial to addressing it. To this end, we study a suite of add-on regularizers that (1) mitigate drift and (2) remain compatible with existing unlearning methods. Beyond generic regularizers, we show that semantic awareness is essential for preserving concepts close to the unlearning target, and propose a gradient-projection method that constrains parameter drift orthogonal to their subspace. This substantially improves continual unlearning performance and is complementary to other regularizers for further gains. Taken together, our study establishes continual unlearning as a fundamental challenge in text-to-image generation and provides insights, baselines, and open directions for advancing safe and accountable generative AI.

Team Cohere, Aakanksha, Arash Ahmadian et al. · mila

In this report we describe the development of Command A, a powerful large language model purpose-built to excel at real-world enterprise use cases. Command A is an agent-optimised and multilingual-capable model, with support for 23 languages of global business, and a novel hybrid architecture balancing efficiency with top of the range performance. It offers best-in-class Retrieval Augmented Generation (RAG) capabilities with grounding and tool use to automate sophisticated business processes. These abilities are achieved through a decentralised training approach, including self-refinement algorithms and model merging techniques. We also include results for Command R7B which shares capability and architectural similarities to Command A. Weights for both models have been released for research purposes. This technical report details our original training pipeline and presents an extensive evaluation of our models across a suite of enterprise-relevant tasks and public benchmarks, demonstrating excellent performance and efficiency.

Justin Lee, Behnaz Moradijamei, Heman Shakeri

Modeling the evolution of high-dimensional systems from limited snapshot observations at irregular time points poses a significant challenge in quantitative biology and related fields. Traditional approaches often rely on dimensionality reduction techniques, which can oversimplify the dynamics and fail to capture critical transient behaviors in non-equilibrium systems. We present Multi-Marginal Stochastic Flow Matching (MMSFM), a novel extension of simulation-free score and flow matching methods to the multi-marginal setting, enabling the alignment of high-dimensional data measured at non-equidistant time points without reducing dimensionality. The use of measure-valued splines enhances robustness to irregular snapshot timing, and score matching prevents overfitting in high-dimensional spaces. We validate our framework on several synthetic and benchmark datasets, including gene expression data collected at uneven time points and an image progression task, demonstrating the method's versatility.

Yazeed Alnumay, Alexandre Barbet, Anna Bialas et al.

Building high-quality large language models (LLMs) for enterprise Arabic applications remains challenging due to the limited availability of digitized Arabic data. In this work, we present a data synthesis and refinement strategy to help address this problem, namely, by leveraging synthetic data generation and human-in-the-loop annotation to expand our Arabic training corpus. We further present our iterative post training recipe that is essential to achieving state-of-the-art performance in aligning the model with human preferences, a critical aspect to enterprise use cases. The culmination of this effort is the release of a small, 7B, open-weight model that outperforms similarly sized peers in head-to-head comparisons and on Arabic-focused benchmarks covering cultural knowledge, instruction following, RAG, and contextual faithfulness.

Yuanchen Wu, Saurabh Verma, Justin Lee et al.

Large language models (LLMs) are highly sensitive to their input prompts, making prompt design a central challenge. While automatic prompt optimization (APO) reduces manual engineering, most approaches assume access to ground-truth references such as labeled validation data. In practice, however, collecting high-quality labels is costly and slow. We propose the Prompt Duel Optimizer (PDO), a sample-efficient framework for label-free prompt optimization. PDO formulates the problem as a dueling-bandit setting, where supervision signal comes from pairwise preference feedback provided by an LLM judge. The framework combines Double Thompson Sampling (D-TS), which prioritizes informative prompt comparisons, with Top-Performer Guided Mutation, which expands the candidate pool by mutating high-performing prompts. PDO naturally operates in label-free settings and can also incorporate partial labels to mitigate judge noise. Experiments on BIG-bench Hard (BBH) and MS MARCO show that PDO consistently outperforms baseline methods. Ablation studies further demonstrate the effectiveness of both D-TS and prompt mutation.

Mahla Abdolahnejad, Justin Lee, Hannah Chan et al.

Burn injuries can result from mechanisms such as thermal, chemical, and electrical insults. A prompt and accurate assessment of burns is essential for deciding definitive clinical treatments. Currently, the primary approach for burn assessments, via visual and tactile observations, is approximately 60%-80% accurate. The gold standard is biopsy and a close second would be non-invasive methods like Laser Doppler Imaging (LDI) assessments, which have up to 97% accuracy in predicting burn severity and the required healing time. In this paper, we introduce a machine learning pipeline for assessing burn severities and segmenting the regions of skin that are affected by burn. Segmenting 2D colour images of burns allows for the injured versus non-injured skin to be delineated, clearly marking the extent and boundaries of the localized burn/region-of-interest, even during remote monitoring of a burn patient. We trained a convolutional neural network (CNN) to classify four severities of burns. We built a saliency mapping method, Boundary Attention Mapping (BAM), that utilises this trained CNN for the purpose of accurately localizing and segmenting the burn regions from skin burn images. We demonstrated the effectiveness of our proposed pipeline through extensive experiments and evaluations using two datasets; 1) A larger skin burn image dataset consisting of 1684 skin burn images of four burn severities, 2) An LDI dataset that consists of a total of 184 skin burn images with their associated LDI scans. The CNN trained using the first dataset achieved an average F1-Score of 78% and micro/macro- average ROC of 85% in classifying the four burn severities. Moreover, a comparison between the BAM results and LDI results for measuring injury boundary showed that the segmentations generated by our method achieved 91.60% accuracy, 78.17% sensitivity, and 93.37% specificity.

Sai Chowdary Gullapally, Yibo Zhang, Nitin Kumar Mittal et al.

Machine learning algorithms have the potential to improve patient outcomes in digital pathology. However, generalization of these tools is currently limited by sensitivity to variations in tissue preparation, staining procedures and scanning equipment that lead to domain shift in digitized slides. To overcome this limitation and improve model generalization, we studied the effectiveness of two Synthetic DOmain-Targeted Augmentation (S-DOTA) methods, namely CycleGAN-enabled Scanner Transform (ST) and targeted Stain Vector Augmentation (SVA), and compared them against the International Color Consortium (ICC) profile-based color calibration (ICC Cal) method and a baseline method using traditional brightness, color and noise augmentations. We evaluated the ability of these techniques to improve model generalization to various tasks and settings: four models, two model types (tissue segmentation and cell classification), two loss functions, six labs, six scanners, and three indications (hepatocellular carcinoma (HCC), nonalcoholic steatohepatitis (NASH), prostate adenocarcinoma). We compared these methods based on the macro-averaged F1 scores on in-distribution (ID) and out-of-distribution (OOD) test sets across multiple domains, and found that S-DOTA methods (i.e., ST and SVA) led to significant improvements over ICC Cal and baseline on OOD data while maintaining comparable performance on ID data. Thus, we demonstrate that S-DOTA may help address generalization due to domain shift in real world applications.

Derek Howard, Marta Maslej, Justin Lee et al.

Mental illness affects a significant portion of the worldwide population. Online mental health forums can provide a supportive environment for those afflicted and also generate a large amount of data which can be mined to predict mental health states using machine learning methods. We benchmark multiple methods of text feature representation for social media posts and compare their downstream use with automated machine learning (AutoML) tools to triage content for moderator attention. We used 1588 labeled posts from the CLPsych 2017 shared task collected from the Reachout.com forum (Milne et al., 2019). Posts were represented using lexicon based tools including VADER, Empath, LIWC and also used pre-trained artificial neural network models including DeepMoji, Universal Sentence Encoder, and GPT-1. We used TPOT and auto-sklearn as AutoML tools to generate classifiers to triage the posts. The top-performing system used features derived from the GPT-1 model, which was finetuned on over 150,000 unlabeled posts from Reachout.com. Our top system had a macro averaged F1 score of 0.572, providing a new state-of-the-art result on the CLPsych 2017 task. This was achieved without additional information from meta-data or preceding posts. Error analyses revealed that this top system often misses expressions of hopelessness. We additionally present visualizations that aid understanding of the learned classifiers. We show that transfer learning is an effective strategy for predicting risk with relatively little labeled data. We note that finetuning of pretrained language models provides further gains when large amounts of unlabeled text is available.

Ayan Sinha, Justin Lee, Shuai Li et al.

Deep learning has been proven to yield reliably generalizable answers to numerous classification and decision tasks. Here, we demonstrate for the first time, to our knowledge, that deep neural networks (DNNs) can be trained to solve inverse problems in computational imaging. We experimentally demonstrate a lens-less imaging system where a DNN was trained to recover a phase object given a raw intensity image recorded some distance away.