Parminder Bhatia

Ben Athiwaratkun, Sanjay Krishna Gouda, Zijian Wang et al. · amazon-science, ibm-research

We present new benchmarks on evaluation code generation models: MBXP and Multilingual HumanEval, and MathQA-X. These datasets cover over 10 programming languages and are generated using a scalable conversion framework that transpiles prompts and test cases from the original Python datasets into the corresponding data in the target language. Using these benchmarks, we are able to assess the performance of code generation models in a multi-lingual fashion, and discovered generalization ability of language models on out-of-domain languages, advantages of multi-lingual models over mono-lingual, the ability of few-shot prompting to teach the model new languages, and zero-shot translation abilities even on mono-lingual settings. Furthermore, we use our code generation model to perform large-scale bootstrapping to obtain synthetic canonical solutions in several languages, which can be used for other code-related evaluations such as code insertion, robustness, or summarization tasks. Overall, our benchmarks represents a significant step towards a deeper understanding of language models' code generation abilities. We publicly release our code and datasets at https://github.com/amazon-research/mxeval.

Hantian Ding, Varun Kumar, Yuchen Tian et al. · amazon-science, stanford

Large language models trained on code have shown great potential to increase productivity of software developers. Several execution-based benchmarks have been proposed to evaluate functional correctness of model-generated code on simple programming problems. Nevertheless, it is expensive to perform the same evaluation on complex real-world projects considering the execution cost. On the contrary, static analysis tools such as linters, which can detect errors without running the program, haven't been well explored for evaluating code generation models. In this work, we propose a static evaluation framework to quantify static errors in Python code completions, by leveraging Abstract Syntax Trees. Compared with execution-based evaluation, our method is not only more efficient, but also applicable to code in the wild. For experiments, we collect code context from open source repos to generate one million function bodies using public models. Our static analysis reveals that Undefined Name and Unused Variable are the most common errors among others made by language models. Through extensive studies, we also show the impact of sampling temperature, model size, and context on static errors in code completions.

Prateek Yadav, Qing Sun, Hantian Ding et al. · amazon-science

Large-scale code generation models such as Codex and CodeT5 have achieved impressive performance. However, libraries are upgraded or deprecated very frequently and re-training large-scale language models is computationally expensive. Therefore, Continual Learning (CL) is an important aspect that remains underexplored in the code domain. In this paper, we introduce a benchmark called CodeTask-CL that covers a wide range of tasks, including code generation, translation, summarization, and refinement, with different input and output programming languages. Next, on our CodeTask-CL benchmark, we compare popular CL techniques from NLP and Vision domains. We find that effective methods like Prompt Pooling (PP) suffer from catastrophic forgetting due to the unstable training of the prompt selection mechanism caused by stark distribution shifts in coding tasks. We address this issue with our proposed method, Prompt Pooling with Teacher Forcing (PP-TF), that stabilizes training by enforcing constraints on the prompt selection mechanism and leads to a 21.54% improvement over Prompt Pooling. Along with the benchmark, we establish a training pipeline that can be used for CL on code models, which we believe can motivate further development of CL methods for code models. Our code is available at https://github.com/amazon-science/codetaskcl-pptf

Shiqi Wang, Zheng Li, Haifeng Qian et al. · amazon-science, ibm-research

Code generation models have achieved impressive performance. However, they tend to be brittle as slight edits to a prompt could lead to very different generations; these robustness properties, critical for user experience when deployed in real-life applications, are not well understood. Most existing works on robustness in text or code tasks have focused on classification, while robustness in generation tasks is an uncharted area and to date there is no comprehensive benchmark for robustness in code generation. In this paper, we propose ReCode, a comprehensive robustness evaluation benchmark for code generation models. We customize over 30 transformations specifically for code on docstrings, function and variable names, code syntax, and code format. They are carefully designed to be natural in real-life coding practice, preserve the original semantic meaning, and thus provide multifaceted assessments of a model's robustness performance. With human annotators, we verified that over 90% of the perturbed prompts do not alter the semantic meaning of the original prompt. In addition, we define robustness metrics for code generation models considering the worst-case behavior under each type of perturbation, taking advantage of the fact that executing the generated code can serve as objective evaluation. We demonstrate ReCode on SOTA models using HumanEval, MBPP, as well as function completion tasks derived from them. Interesting observations include: better robustness for CodeGen over InCoder and GPT-J; models are most sensitive to syntax perturbations; more challenging robustness evaluation on MBPP over HumanEval.

Yangruibo Ding, Zijian Wang, Wasi Uddin Ahmad et al. · stanford

While pre-trained language models (LM) for code have achieved great success in code completion, they generate code conditioned only on the contents within the file, i.e., in-file context, but ignore the rich semantics in other files within the same project, i.e., cross-file context, a critical source of information that is especially useful in modern modular software development. Such overlooking constrains code language models' capacity in code completion, leading to unexpected behaviors such as generating hallucinated class member functions or function calls with unexpected arguments. In this work, we develop a cross-file context finder tool, CCFINDER, that effectively locates and retrieves the most relevant cross-file context. We propose CoCoMIC, a framework that incorporates cross-file context to learn the in-file and cross-file context jointly on top of pretrained code LMs. CoCoMIC successfully improves the existing code LM with a 33.94% relative increase in exact match and a 28.69% relative increase in identifier matching for code completion when the cross-file context is provided.

Zheng Li, Zijian Wang, Ming Tan et al. · stanford

Large-scale pre-trained sequence-to-sequence models like BART and T5 achieve state-of-the-art performance on many generative NLP tasks. However, such models pose a great challenge in resource-constrained scenarios owing to their large memory requirements and high latency. To alleviate this issue, we propose to jointly distill and quantize the model, where knowledge is transferred from the full-precision teacher model to the quantized and distilled low-precision student model. Empirical analyses show that, despite the challenging nature of generative tasks, we were able to achieve a 16.5x model footprint compression ratio with little performance drop relative to the full-precision counterparts on multiple summarization and QA datasets. We further pushed the limit of compression ratio to 27.7x and presented the performance-efficiency trade-off for generative tasks using pre-trained models. To the best of our knowledge, this is the first work aiming to effectively distill and quantize sequence-to-sequence pre-trained models for language generation tasks.

Yangruibo Ding, Zijian Wang, Wasi Uddin Ahmad et al.

Code completion models have made significant progress in recent years, yet current popular evaluation datasets, such as HumanEval and MBPP, predominantly focus on code completion tasks within a single file. This over-simplified setting falls short of representing the real-world software development scenario where repositories span multiple files with numerous cross-file dependencies, and accessing and understanding cross-file context is often required to complete the code correctly. To fill in this gap, we propose CrossCodeEval, a diverse and multilingual code completion benchmark that necessitates an in-depth cross-file contextual understanding to complete the code accurately. CrossCodeEval is built on a diverse set of real-world, open-sourced, permissively-licensed repositories in four popular programming languages: Python, Java, TypeScript, and C#. To create examples that strictly require cross-file context for accurate completion, we propose a straightforward yet efficient static-analysis-based approach to pinpoint the use of cross-file context within the current file. Extensive experiments on state-of-the-art code language models like CodeGen and StarCoder demonstrate that CrossCodeEval is extremely challenging when the relevant cross-file context is absent, and we see clear improvements when adding these context into the prompt. However, despite such improvements, the pinnacle of performance remains notably unattained even with the highest-performing model, indicating that CrossCodeEval is also capable of assessing model's capability in leveraging extensive context to make better code completion. Finally, we benchmarked various methods in retrieving cross-file context, and show that CrossCodeEval can also be used to measure the capability of code retrievers.

Nihal Jain, Dejiao Zhang, Wasi Uddin Ahmad et al. · amazon-science, stanford

Despite exciting progress in causal language models, the expressiveness of the representations is largely limited due to poor discrimination ability. To remedy this issue, we present ContraCLM, a novel contrastive learning framework at both token-level and sequence-level. We assess ContraCLM on a variety of downstream tasks. We show that ContraCLM enhances discrimination of the representations and bridges the gap with the encoder-only models, which makes causal language models better suited for tasks beyond language generation. Specifically, we attain $44\%$ relative improvement on the Semantic Textual Similarity tasks and $34\%$ on Code-to-Code Search tasks. Furthermore, by improving the expressiveness of the representations, ContraCLM also boosts the source code generation capability with $9\%$ relative improvement on execution accuracy on the HumanEval benchmark.

Xiaokai Wei, Sujan Gonugondla, Wasi Ahmad et al. · amazon-science, ibm-research

ML-powered code generation aims to assist developers to write code in a more productive manner, by intelligently generating code blocks based on natural language prompts. Recently, large pretrained deep learning models have substantially pushed the boundary of code generation and achieved impressive performance. Despite their great power, the huge number of model parameters poses a significant threat to adapting them in a regular software development environment, where a developer might use a standard laptop or mid-size server to develop her code. Such large models incur significant resource usage (in terms of memory, latency, and dollars) as well as carbon footprint. Model compression is a promising approach to address these challenges. Several techniques are proposed to compress large pretrained models typically used for vision or textual data. Out of many available compression techniques, we identified that quantization is mostly applicable for code generation task as it does not require significant retraining cost. As quantization represents model parameters with lower-bit integer (e.g., int8), the model size and runtime latency would both benefit from such int representation. We extensively study the impact of quantized model on code generation tasks across different dimension: (i) resource usage and carbon footprint, (ii) accuracy, and (iii) robustness. To this end, through systematic experiments we find a recipe of quantization technique that could run even a $6$B model in a regular laptop without significant accuracy or robustness degradation. We further found the recipe is readily applicable to code summarization task as well.

Yuantong Li, Xiaokai Wei, Zijian Wang et al. · amazon-science, stanford

People frequently interact with information retrieval (IR) systems, however, IR models exhibit biases and discrimination towards various demographics. The in-processing fair ranking methods provide a trade-offs between accuracy and fairness through adding a fairness-related regularization term in the loss function. However, there haven't been intuitive objective functions that depend on the click probability and user engagement to directly optimize towards this. In this work, we propose the In-Batch Balancing Regularization (IBBR) to mitigate the ranking disparity among subgroups. In particular, we develop a differentiable \textit{normed Pairwise Ranking Fairness} (nPRF) and leverage the T-statistics on top of nPRF over subgroups as a regularization to improve fairness. Empirical results with the BERT-based neural rankers on the MS MARCO Passage Retrieval dataset with the human-annotated non-gendered queries benchmark \citep{rekabsaz2020neural} show that our IBBR method with nPRF achieves significantly less bias with minimal degradation in ranking performance compared with the baseline.

Pengcheng Jiang, Cao Xiao, Tianfan Fu et al.

Molecular representation learning is vital for various downstream applications, including the analysis and prediction of molecular properties and side effects. While Graph Neural Networks (GNNs) have been a popular framework for modeling molecular data, they often struggle to capture the full complexity of molecular representations. In this paper, we introduce a novel method called GODE, which accounts for the dual-level structure inherent in molecules. Molecules possess an intrinsic graph structure and simultaneously function as nodes within a broader molecular knowledge graph. GODE integrates individual molecular graph representations with multi-domain biochemical data from knowledge graphs. By pre-training two GNNs on different graph structures and employing contrastive learning, GODE effectively fuses molecular structures with their corresponding knowledge graph substructures. This fusion yields a more robust and informative representation, enhancing molecular property predictions by leveraging both chemical and biological information. When fine-tuned across 11 chemical property tasks, our model significantly outperforms existing benchmarks, achieving an average ROC-AUC improvement of 12.7% for classification tasks and an average RMSE/MAE improvement of 34.4% for regression tasks. Notably, GODE surpasses the current leading model in property prediction, with advancements of 2.2% in classification and 7.2% in regression tasks.

Matthew Trager, Pramuditha Perera, Luca Zancato et al.

We investigate compositional structures in data embeddings from pre-trained vision-language models (VLMs). Traditionally, compositionality has been associated with algebraic operations on embeddings of words from a pre-existing vocabulary. In contrast, we seek to approximate representations from an encoder as combinations of a smaller set of vectors in the embedding space. These vectors can be seen as "ideal words" for generating concepts directly within the embedding space of the model. We first present a framework for understanding compositional structures from a geometric perspective. We then explain what these compositional structures entail probabilistically in the case of VLM embeddings, providing intuitions for why they arise in practice. Finally, we empirically explore these structures in CLIP's embeddings and we evaluate their usefulness for solving different vision-language tasks such as classification, debiasing, and retrieval. Our results show that simple linear algebraic operations on embedding vectors can be used as compositional and interpretable methods for regulating the behavior of VLMs.

Deepa Anand, Gurunath Reddy M, Vanika Singhal et al.

Recent advances in Vision Transformers (ViT) and Stable Diffusion (SD) models with their ability to capture rich semantic features of the image have been used for image correspondence tasks on natural images. In this paper, we examine the ability of a variety of pre-trained ViT (DINO, DINOv2, SAM, CLIP) and SD models, trained exclusively on natural images, for solving the correspondence problems on medical images. While many works have made a case for in-domain training, we show that the models trained on natural images can offer good performance on medical images across different modalities (CT,MR,Ultrasound) sourced from various manufacturers, over multiple anatomical regions (brain, thorax, abdomen, extremities), and on wide variety of tasks. Further, we leverage the correspondence with respect to a template image to prompt a Segment Anything (SAM) model to arrive at single shot segmentation, achieving dice range of 62%-90% across tasks, using just one image as reference. We also show that our single-shot method outperforms the recently proposed few-shot segmentation method - UniverSeg (Dice range 47%-80%) on most of the semantic segmentation tasks(six out of seven) across medical imaging modalities.

Alex Boyd, Andrew Warrington, Taha Kass-Hout et al.

Foundational marked temporal point process (MTPP) models, such as the Hawkes process, often use inexpressive model families in order to offer interpretable parameterizations of event data. On the other hand, neural MTPPs models forego this interpretability in favor of absolute predictive performance. In this work, we present a new family MTPP models: the hyper Hawkes process (HHP), which aims to be as flexible and performant as neural MTPPs, while retaining interpretable aspects. To achieve this, the HHP extends the classical Hawkes process to increase its expressivity by first expanding the dimension of the process into a latent space, and then introducing a hypernetwork to allow time- and data-dependent dynamics. These extensions define a highly performant MTPP family, achieving state-of-the-art performance across a range of benchmark tasks and metrics. Furthermore, by retaining the linearity of the recurrence, albeit now piecewise and conditionally linear, the HHP also retains much of the structure of the original Hawkes process, which we exploit to create direct probes into how the model creates predictions. HHP models therefore offer both state-of-the-art predictions, while also providing an opportunity to ``open the box'' and inspect how predictions were generated.

Hariharan Ravishankar, Rohan Patil, Vikram Melapudi et al.

In this paper, we present SonoSAMTrack - that combines a promptable foundational model for segmenting objects of interest on ultrasound images called SonoSAM, with a state-of-the art contour tracking model to propagate segmentations on 2D+t and 3D ultrasound datasets. Fine-tuned and tested exclusively on a rich, diverse set of objects from $\approx200$k ultrasound image-mask pairs, SonoSAM demonstrates state-of-the-art performance on 7 unseen ultrasound data-sets, outperforming competing methods by a significant margin. We also extend SonoSAM to 2-D +t applications and demonstrate superior performance making it a valuable tool for generating dense annotations and segmentation of anatomical structures in clinical workflows. Further, to increase practical utility of the work, we propose a two-step process of fine-tuning followed by knowledge distillation to a smaller footprint model without comprising the performance. We present detailed qualitative and quantitative comparisons of SonoSAM with state-of-the-art methods showcasing efficacy of the method. This is followed by demonstrating the reduction in number of clicks in a dense video annotation problem of adult cardiac ultrasound chamber segmentation using SonoSAMTrack.

Parminder Bhatia, Busra Celikkaya, Mohammed Khalilia et al.

Comprehend Medical is a stateless and Health Insurance Portability and Accountability Act (HIPAA) eligible Named Entity Recognition (NER) and Relationship Extraction (RE) service launched under Amazon Web Services (AWS) trained using state-of-the-art deep learning models. Contrary to many existing open source tools, Comprehend Medical is scalable and does not require steep learning curve, dependencies, pipeline configurations, or installations. Currently, Comprehend Medical performs NER in five medical categories: Anatomy, Medical Condition, Medications, Protected Health Information (PHI) and Treatment, Test and Procedure (TTP). Additionally, the service provides relationship extraction for the detected entities as well as contextual information such as negation and temporality in the form of traits. Comprehend Medical provides two Application Programming Interfaces (API): 1) the NERe API which returns all the extracted named entities, their traits and the relationships between them and 2) the PHId API which returns just the protected health information contained in the text. Furthermore, Comprehend Medical is accessible through AWS Console, Java and Python Software Development Kit (SDK), making it easier for non-developers and developers to use.

Xiyao Wang, Jiuhai Chen, Zhaoyang Wang et al.

Large vision-language models (LVLMs) have achieved impressive results in visual question-answering and reasoning tasks through vision instruction tuning on specific datasets. However, there remains significant room for improvement in aligning visual and language modalities. Existing methods often depend on external models or data, leading to uncontrollable and unstable alignment results. In this paper, we propose SIMA, a self-improvement framework that enhances visual and language modality alignment without external dependencies. SIMA leverages existing vision instruction tuning datasets to self-generate responses, incorporating an in-context self-critic mechanism that constructs preference pairs for tuning. Crucially, our approach allows LVLMs to act as critics by designing effective critic prompts, eliminating the need for additional fine-tuning with external instruction data. We introduce three novel visual metrics within the self-critic process to guide judgment, significantly improving the accuracy of self-critic. Through extensive experiments across 14 hallucination and comprehensive benchmarks, we demonstrate that SIMA significantly improves LVLM's performance and outperforms previous approaches, achieving superior modality alignment.

Pengcheng Jiang, Cao Xiao, Zifeng Wang et al.

The advent of large language models (LLMs) has significantly advanced natural language processing tasks like text summarization. However, their large size and computational demands, coupled with privacy concerns in data transmission, limit their use in resource-constrained and privacy-centric settings. To overcome this, we introduce TriSum, a framework for distilling LLMs' text summarization abilities into a compact, local model. Initially, LLMs extract a set of aspect-triple rationales and summaries, which are refined using a dual-scoring method for quality. Next, a smaller local model is trained with these tasks, employing a curriculum learning strategy that evolves from simple to complex tasks. Our method enhances local model performance on various benchmarks (CNN/DailyMail, XSum, and ClinicalTrial), outperforming baselines by 4.5%, 8.5%, and 7.4%, respectively. It also improves interpretability by providing insights into the summarization rationale.

Ben Athiwaratkun, Sujan Kumar Gonugondla, Sanjay Krishna Gouda et al. · amazon-science

This study introduces bifurcated attention, a method designed to enhance language model inference in shared-context batch decoding scenarios. Our approach addresses the challenge of redundant memory IO costs, a critical factor contributing to latency in high batch sizes and extended context lengths. Bifurcated attention achieves this by strategically dividing the attention mechanism during incremental decoding into two separate GEMM operations: one focusing on the KV cache from prefill, and another on the decoding process itself. While maintaining the computational load (FLOPs) of standard attention mechanisms, bifurcated attention ensures precise computation with significantly reduced memory IO. Our empirical results show over 2.1$\times$ speedup when sampling 16 output sequences and more than 6.2$\times$ speedup when sampling 32 sequences at context lengths exceeding 8k tokens on a 7B model that uses multi-head attention. The efficiency gains from bifurcated attention translate into lower latency, making it particularly suitable for real-time applications. For instance, it enables massively parallel answer generation without substantially increasing latency, thus enhancing performance when integrated with post-processing techniques such as re-ranking.

Aishik Konwer, Zhijian Yang, Erhan Bas et al.

Foundational models such as the Segment Anything Model (SAM) are gaining traction in medical imaging segmentation, supporting multiple downstream tasks. However, such models are supervised in nature, still relying on large annotated datasets or prompts supplied by experts. Conventional techniques such as active learning to alleviate such limitations are limited in scope and still necessitate continuous human involvement and complex domain knowledge for label refinement or establishing reward ground truth. To address these challenges, we propose an enhanced Segment Anything Model (SAM) framework that utilizes annotation-efficient prompts generated in a fully unsupervised fashion, while still capturing essential semantic, location, and shape information through contrastive language-image pretraining and visual question answering. We adopt the direct preference optimization technique to design an optimal policy that enables the model to generate high-fidelity segmentations with simple ratings or rankings provided by a virtual annotator simulating the human annotation process. State-of-the-art performance of our framework in tasks such as lung segmentation, breast tumor segmentation, and organ segmentation across various modalities, including X-ray, ultrasound, and abdominal CT, justifies its effectiveness in low-annotation data scenarios.

Aofei Chang, Le Huang, Parminder Bhatia et al.

Large Vision Language Models (LVLMs) are becoming increasingly important in the medical domain, yet Medical LVLMs (Med-LVLMs) frequently generate hallucinations due to limited expertise and the complexity of medical applications. Existing benchmarks fail to effectively evaluate hallucinations based on their underlying causes and lack assessments of mitigation strategies. To address this gap, we introduce MedHEval, a novel benchmark that systematically evaluates hallucinations and mitigation strategies in Med-LVLMs by categorizing them into three underlying causes: visual misinterpretation, knowledge deficiency, and context misalignment. We construct a diverse set of close- and open-ended medical VQA datasets with comprehensive evaluation metrics to assess these hallucination types. We conduct extensive experiments across 11 popular (Med)-LVLMs and evaluate 7 state-of-the-art hallucination mitigation techniques. Results reveal that Med-LVLMs struggle with hallucinations arising from different causes while existing mitigation methods show limited effectiveness, especially for knowledge- and context-based errors. These findings underscore the need for improved alignment training and specialized mitigation strategies to enhance Med-LVLMs' reliability. MedHEval establishes a standardized framework for evaluating and mitigating medical hallucinations, guiding the development of more trustworthy Med-LVLMs.

Shuyang Yu, Runxue Bao, Parminder Bhatia et al.

Large language models (LLMs) can learn vast amounts of knowledge from diverse domains during pre-training. However, long-tail knowledge from specialized domains is often scarce and underrepresented, rarely appearing in the models' memorization. Prior work has shown that in-context learning (ICL) with retriever augmentation can help LLMs better capture long-tail knowledge, reducing their reliance on pre-trained data. Despite these advances, we observe that LLM predictions for long-tail questions remain uncertain to variations in retrieved samples. To take advantage of the uncertainty in ICL for guiding LLM predictions toward correct answers on long-tail samples, we propose a reinforcement learning-based dynamic uncertainty ranking method for ICL that accounts for the varying impact of each retrieved sample on LLM predictions. Our approach prioritizes more informative and stable samples while demoting misleading ones, updating rankings based on the feedback from the LLM w.r.t. each retrieved sample. To enhance training efficiency and reduce query costs, we introduce a learnable dynamic ranking threshold, adjusted when the model encounters negative prediction shifts. Experimental results on various question-answering datasets from different domains show that our method outperforms the best baseline by $2.76\%$, with a notable $5.96\%$ boost in accuracy on long-tail questions that elude zero-shot inference.

Yuxin Chang, Alex Boyd, Cao Xiao et al.

Marked temporal point processes (MTPPs) model sequences of events occurring at irregular time intervals, with wide-ranging applications in fields such as healthcare, finance and social networks. We propose the state-space point process (S2P2) model, a novel and performant model that leverages techniques derived for modern deep state-space models (SSMs) to overcome limitations of existing MTPP models, while simultaneously imbuing strong inductive biases for continuous-time event sequences that other discrete sequence models (i.e., RNNs, transformers) do not capture. Inspired by the classical linear Hawkes processes, we propose an architecture that interleaves stochastic jump differential equations with nonlinearities to create a highly expressive intensity-based MTPP model, without the need for restrictive parametric assumptions for the intensity. Our approach enables efficient training and inference with a parallel scan, bringing linear complexity and sublinear scaling while retaining expressivity to MTPPs. Empirically, S2P2 achieves state-of-the-art predictive likelihoods across eight real-world datasets, delivering an average improvement of 33% over the best existing approaches.

Aofei Chang, Le Huang, Alex James Boyd et al.

Medical Large Vision-Language Models (Med-LVLMs) often exhibit suboptimal attention distribution on visual inputs, leading to hallucinated or inaccurate outputs. Existing mitigation methods primarily rely on inference-time interventions, which are limited in attention adaptation or require additional supervision. To address this, we propose A$^3$Tune, a novel fine-tuning framework for Automatic Attention Alignment Tuning. A$^3$Tune leverages zero-shot weak labels from SAM, refines them into prompt-aware labels using BioMedCLIP, and then selectively modifies visually-critical attention heads to improve alignment while minimizing interference. Additionally, we introduce a A$^3$MoE module, enabling adaptive parameter selection for attention tuning across diverse prompts and images. Extensive experiments on medical VQA and report generation benchmarks show that A$^3$Tune outperforms state-of-the-art baselines, achieving enhanced attention distributions and performance in Med-LVLMs.

Haoyan Yang, Runxue Bao, Cao Xiao et al.

LLM-as-a-Judge has emerged as a promising tool for automatically evaluating generated outputs, but its reliability is often undermined by potential biases in judgment. Existing efforts to mitigate these biases face key limitations: in-context learning-based methods fail to address rooted biases due to the evaluator's limited capacity for self-reflection, whereas fine-tuning is not applicable to all evaluator types, especially closed-source models. To address this challenge, we introduce the Reasoning-based Bias Detector (RBD), which is a plug-in module that identifies biased evaluations and generates structured reasoning to guide evaluator self-correction. Rather than modifying the evaluator itself, RBD operates externally and engages in an iterative process of bias detection and feedback-driven revision. To support its development, we design a complete pipeline consisting of biased dataset construction, supervision collection, distilled reasoning-based fine-tuning of RBD, and integration with LLM evaluators. We fine-tune four sizes of RBD models, ranging from 1.5B to 14B, and observe consistent performance improvements across all scales. Experimental results on 4 bias types--verbosity, position, bandwagon, and sentiment--evaluated using 8 LLM evaluators demonstrate RBD's strong effectiveness. For example, the RBD-8B model improves evaluation accuracy by an average of 18.5% and consistency by 10.9%, and surpasses prompting-based baselines and fine-tuned judges by 12.8% and 17.2%, respectively. These results highlight RBD's effectiveness and scalability. Additional experiments further demonstrate its strong generalization across biases and domains, as well as its efficiency.

Chenliang Li, Adel Elmahdy, Alex Boyd et al.

Reinforcement learning (RL) algorithms such as PPO and GRPO are widely used to train large language models (LLMs) for multi-turn agentic tasks. However, in off-policy training pipelines, these methods often exhibit unstable optimization dynamics and are prone to performance collapse. Through empirical analysis, we identify two fundamental sources of instability in this setting: (1)~a granularity mismatch between token-level policy optimization and turn-structured interactions, and (2) high-variance and unreliable gradient updates induced by off-policy importance sampling and inaccurate advantage estimation. To address these challenges, we propose SORL, \underline{S}tabilizing \underline{O}ff-Policy \underline{R}einforcement \underline{L}earning for Long-Horizon Agent Training. SORL introduces principled mechanisms that align policy optimization with the structure of multi-turn interactions and adaptively suppress unreliable off-policy updates, yielding more conservative and robust learning dynamics. Within this framework, we instantiate two stabilized algorithms: SO-PPO and SO-GRPO. Both algorithms are designed to mitigate gradient variance and prevent optimization collapse without requiring careful early stopping or heuristic tuning. We evaluate SO-PPO and SO-GRPO on a range of multi-turn search benchmarks, including general question answering, multi-hop question answering, and medical multiple-choice QA tasks. Experimental results show that both methods consistently prevent training instabilities and performance collapses observed in standard PPO and GRPO, maintain lower clipping ratios and more stable optimization trajectories, and achieve superior or comparable task performance. These results demonstrate that the proposed algorithm provides a practical, scalable, and general framework for stabilizing reinforcement learning in multi-turn LLM agent training.

Sicheng Zhou, Lei Wu, Cao Xiao et al.

Self-supervised learning (SSL) has transformed vision encoder training in general domains but remains underutilized in medical imaging due to limited data and domain specific biases. We present MammoDINO, a novel SSL framework for mammography, pretrained on 1.4 million mammographic images. To capture clinically meaningful features, we introduce a breast tissue aware data augmentation sampler for both image-level and patch-level supervision and a cross-slice contrastive learning objective that leverages 3D digital breast tomosynthesis (DBT) structure into 2D pretraining. MammoDINO achieves state-of-the-art performance on multiple breast cancer screening tasks and generalizes well across five benchmark datasets. It offers a scalable, annotation-free foundation for multipurpose computer-aided diagnosis (CAD) tools for mammogram, helping reduce radiologists' workload and improve diagnostic efficiency in breast cancer screening.

Zhijian Yang, Noel DSouza, Istvan Megyeri et al.

Magnetic Resonance Imaging (MRI) is a critical medical imaging modality in clinical diagnosis and research, yet its complexity and heterogeneity pose challenges for automated analysis, particularly in scalable and generalizable machine learning applications. While foundation models have revolutionized natural language and vision tasks, their application to MRI remains limited due to data scarcity and narrow anatomical focus. In this work, we present Decipher-MR, a 3D MRI-specific vision-language foundation model trained on a large-scale dataset comprising 200,000 MRI series from over 22,000 studies spanning diverse anatomical regions, sequences, and pathologies. Decipher-MR integrates self-supervised vision learning with report-guided text supervision to build robust, generalizable representations, enabling effective adaptation across broad applications. To enable robust and diverse clinical tasks with minimal computational overhead, Decipher-MR supports a modular design that enables tuning of lightweight, task-specific decoders attached to a frozen pretrained encoder. Following this setting, we evaluate Decipher-MR across diverse benchmarks including disease classification, demographic prediction, anatomical localization, and cross-modal retrieval, demonstrating consistent performance gains over existing foundation models and task-specific approaches. Our results establish Decipher-MR as a scalable and versatile foundation for MRI-based AI, facilitating efficient development across clinical and research domains.

Denis Jered McInerney, Luyang Kong, Kristjan Arumae et al.

Multi-task learning is useful in NLP because it is often practically desirable to have a single model that works across a range of tasks. In the medical domain, sequential training on tasks may sometimes be the only way to train models, either because access to the original (potentially sensitive) data is no longer available, or simply owing to the computational costs inherent to joint retraining. A major issue inherent to sequential learning, however, is catastrophic forgetting, i.e., a substantial drop in accuracy on prior tasks when a model is updated for a new task. Elastic Weight Consolidation is a recently proposed method to address this issue, but scaling this approach to the modern large models used in practice requires making strong independence assumptions about model parameters, limiting its effectiveness. In this work, we apply Kronecker Factorization--a recent approach that relaxes independence assumptions--to prevent catastrophic forgetting in convolutional and Transformer-based neural networks at scale. We show the effectiveness of this technique on the important and illustrative task of medical entity linking across three datasets, demonstrating the capability of the technique to be used to make efficient updates to existing methods as new medical data becomes available. On average, the proposed method reduces catastrophic forgetting by 51% when using a BERT-based model, compared to a 27% reduction using standard Elastic Weight Consolidation, while maintaining spatial complexity proportional to the number of model parameters.

Xiaokai Wei, Shen Wang, Dejiao Zhang et al.

Pretrained Language Models (PLM) have established a new paradigm through learning informative contextualized representations on large-scale text corpus. This new paradigm has revolutionized the entire field of natural language processing, and set the new state-of-the-art performance for a wide variety of NLP tasks. However, though PLMs could store certain knowledge/facts from training corpus, their knowledge awareness is still far from satisfactory. To address this issue, integrating knowledge into PLMs have recently become a very active research area and a variety of approaches have been developed. In this paper, we provide a comprehensive survey of the literature on this emerging and fast-growing field - Knowledge Enhanced Pretrained Language Models (KE-PLMs). We introduce three taxonomies to categorize existing work. Besides, we also survey the various NLU and NLG applications on which KE-PLM has demonstrated superior performance over vanilla PLMs. Finally, we discuss challenges that face KE-PLMs and also promising directions for future research.

Fred Qin, Vivek Madan, Ujjwal Ratan et al.

Sepsis is a life-threatening disease with high morbidity, mortality and healthcare costs. The early prediction and administration of antibiotics and intravenous fluids is considered crucial for the treatment of sepsis and can save potentially millions of lives and billions in health care costs. Professional clinical care practitioners have proposed clinical criterion which aid in early detection of sepsis; however, performance of these criterion is often limited. Clinical text provides essential information to estimate the severity of the sepsis in addition to structured clinical data. In this study, we explore how clinical text can complement structured data towards early sepsis prediction task. In this paper, we propose multi modal model which incorporates both structured data in the form of patient measurements as well as textual notes on the patient. We employ state-of-the-art NLP models such as BERT and a highly specialized NLP model in Amazon Comprehend Medical to represent the text. On the MIMIC-III dataset containing records of ICU admissions, we show that by using these notes, one achieves an improvement of 6.07 points in a standard utility score for Sepsis prediction and 2.89% in AUROC score. Our methods significantly outperforms a clinical criteria suggested by experts, qSOFA, as well as the winning model of the PhysioNet Computing in Cardiology Challenge for predicting Sepsis.

Qing Sun, Parminder Bhatia

Incorporating external knowledge into Named Entity Recognition (NER) systems has been widely studied in the generic domain. In this paper, we focus on clinical domain where only limited data is accessible and interpretability is important. Recent advancement in technology and the acceleration of clinical trials has resulted in the discovery of new drugs, procedures as well as medical conditions. These factors motivate towards building robust zero-shot NER systems which can quickly adapt to new medical terminology. We propose an auxiliary gazetteer model and fuse it with an NER system, which results in better robustness and interpretability across different clinical datasets. Our gazetteer based fusion model is data efficient, achieving +1.7 micro-F1 gains on the i2b2 dataset using 20% training data, and brings + 4.7 micro-F1 gains on novel entity mentions never presented during training. Moreover, our fusion model is able to quickly adapt to new mentions in gazetteers without re-training and the gains from the proposed fusion model are transferable to related datasets.

Luyang Kong, Christopher Winestock, Parminder Bhatia

Medical entity retrieval is an integral component for understanding and communicating information across various health systems. Current approaches tend to work well on specific medical domains but generalize poorly to unseen sub-specialties. This is of increasing concern under a public health crisis as new medical conditions and drug treatments come to light frequently. Zero-shot retrieval is challenging due to the high degree of ambiguity and variability in medical corpora, making it difficult to build an accurate similarity measure between mentions and concepts. Medical knowledge graphs (KG), however, contain rich semantics including large numbers of synonyms as well as its curated graphical structures. To take advantage of this valuable information, we propose a suite of learning tasks designed for training efficient zero-shot entity retrieval models. Without requiring any human annotation, our knowledge graph enriched architecture significantly outperforms common zero-shot benchmarks including BM25 and Clinical BERT with 7% to 30% higher recall across multiple major medical ontologies, such as UMLS, SNOMED, and ICD-10.

Han-Chin Shing, Chaitanya Shivade, Nima Pourdamghani et al.

The records of a clinical encounter can be extensive and complex, thus placing a premium on tools that can extract and summarize relevant information. This paper introduces the task of generating discharge summaries for a clinical encounter. Summaries in this setting need to be faithful, traceable, and scale to multiple long documents, motivating the use of extract-then-abstract summarization cascades. We introduce two new measures, faithfulness and hallucination rate for evaluation in this task, which complement existing measures for fluency and informativeness. Results across seven medical sections and five models show that a summarization architecture that supports traceability yields promising results, and that a sentence-rewriting approach performs consistently on the measure used for faithfulness (faithfulness-adjusted $F_3$) over a diverse range of generated sections.

Kristjan Arumae, Qing Sun, Parminder Bhatia

Pre-training large language models has become a standard in the natural language processing community. Such models are pre-trained on generic data (e.g. BookCorpus and English Wikipedia) and often fine-tuned on tasks in the same domain. However, in order to achieve state-of-the-art performance on out of domain tasks such as clinical named entity recognition and relation extraction, additional in domain pre-training is required. In practice, staged multi-domain pre-training presents performance deterioration in the form of catastrophic forgetting (CF) when evaluated on a generic benchmark such as GLUE. In this paper we conduct an empirical investigation into known methods to mitigate CF. We find that elastic weight consolidation provides best overall scores yielding only a 0.33% drop in performance across seven generic tasks while remaining competitive in bio-medical tasks. Furthermore, we explore gradient and latent clustering based data selection techniques to improve coverage when using elastic weight consolidation and experience replay methods.

Luyang Kong, Lifan Chen, Ming Chen et al.

Anomaly detectors are often designed to catch statistical anomalies. End-users typically do not have interest in all of the detected outliers, but only those relevant to their application. Given an existing black-box sequential anomaly detector, this paper proposes a method to improve its user relevancy using a small number of human feedback. As our first contribution, the method is agnostic to the detector: it only assumes access to its anomaly scores, without requirement on any additional information inside it. Inspired by a fact that anomalies are of different types, our approach identifies these types and utilizes user feedback to assign relevancy to types. This relevancy score, as our second contribution, is used to adjust the subsequent anomaly selection process. Empirical results on synthetic and real-world datasets show that our approach yields significant improvements on precision and recall over a range of anomaly detectors.

Colby Wise, Vassilis N. Ioannidis, Miguel Romero Calvo et al.

The coronavirus disease (COVID-19) has claimed the lives of over 350,000 people and infected more than 6 million people worldwide. Several search engines have surfaced to provide researchers with additional tools to find and retrieve information from the rapidly growing corpora on COVID-19. These engines lack extraction and visualization tools necessary to retrieve and interpret complex relations inherent to scientific literature. Moreover, because these engines mainly rely upon semantic information, their ability to capture complex global relationships across documents is limited, which reduces the quality of similarity-based article recommendations for users. In this work, we present the COVID-19 Knowledge Graph (CKG), a heterogeneous graph for extracting and visualizing complex relationships between COVID-19 scientific articles. The CKG combines semantic information with document topological information for the application of similar document retrieval. The CKG is constructed using the latent schema of the data, and then enriched with biomedical entity information extracted from the unstructured text of articles using scalable AWS technologies to form relations in the graph. Finally, we propose a document similarity engine that leverages low-dimensional graph embeddings from the CKG with semantic embeddings for similar article retrieval. Analysis demonstrates the quality of relationships in the CKG and shows that it can be used to uncover meaningful information in COVID-19 scientific articles. The CKG helps power www.cord19.aws and is publicly available.

Parminder Bhatia, Lan Liu, Kristjan Arumae et al.

Coronavirus disease (COVID-19) has been declared as a pandemic by WHO with thousands of cases being reported each day. Numerous scientific articles are being published on the disease raising the need for a service which can organize, and query them in a reliable fashion. To support this cause we present AWS CORD-19 Search (ACS), a public, COVID-19 specific, neural search engine that is powered by several machine learning systems to support natural language based searches. ACS with capabilities such as document ranking, passage ranking, question answering and topic classification provides a scalable solution to COVID-19 researchers and policy makers in their search and discovery for answers to high priority scientific questions. We present a quantitative evaluation and qualitative analysis of the system against other leading COVID-19 search platforms. ACS is top performing across these systems yielding quality results which we detail with relevant examples in this work.

Shaoqing Yuan, Parminder Bhatia, Busra Celikkaya et al.

Recent advancements in medical entity linking have been applied in the area of scientific literature and social media data. However, with the adoption of telemedicine and conversational agents such as Alexa in healthcare settings, medical name inference has become an important task. Medication name inference is the task of mapping user friendly medication names from a free-form text to a concept in a normalized medication list. This is challenging due to the differences in the use of medical terminology from health care professionals and user conversations coming from the lay public. We begin with mapping descriptive medication phrases (DMP) to standard medication names (SMN). Given the prescriptions of each patient, we want to provide them with the flexibility of referring to the medication in their preferred ways. We approach this as a ranking problem which maps SMN to DMP by ordering the list of medications in the patient's prescription list obtained from pharmacies. Furthermore, we leveraged the output of intermediate layers and performed medication clustering. We present the Medication Inference Model (MIM) achieving state-of-the-art results. By incorporating medical entities based attention, we have obtained further improvement for ranking models.

Miguel Ballesteros, Rishita Anubhai, Shuai Wang et al.

In this paper, we propose a neural architecture and a set of training methods for ordering events by predicting temporal relations. Our proposed models receive a pair of events within a span of text as input and they identify temporal relations (Before, After, Equal, Vague) between them. Given that a key challenge with this task is the scarcity of annotated data, our models rely on either pretrained representations (i.e. RoBERTa, BERT or ELMo), transfer and multi-task learning (by leveraging complementary datasets), and self-training techniques. Experiments on the MATRES dataset of English documents establish a new state-of-the-art on this task.

Kristjan Arumae, Parminder Bhatia

Training large language representation models has become a standard in the natural language processing community. This allows for fine tuning on any number of specific tasks, however, these large high capacity models can continue to train on domain specific unlabeled data to make initialization even more robust for supervised tasks. We demonstrate that in practice these pre-trained models present performance deterioration in the form of catastrophic forgetting when evaluated on tasks from a general domain such as GLUE. In this work we propose CALM, Continuous Adaptive Learning for Language Modeling: techniques to render models which retain knowledge across multiple domains. With these methods, we are able to reduce the performance gap across supervised tasks introduced by task specific models which we demonstrate using a continual learning setting in biomedical and clinical domains.

Ming Zhu, Busra Celikkaya, Parminder Bhatia et al.

Entity linking is the task of linking mentions of named entities in natural language text, to entities in a curated knowledge-base. This is of significant importance in the biomedical domain, where it could be used to semantically annotate a large volume of clinical records and biomedical literature, to standardized concepts described in an ontology such as Unified Medical Language System (UMLS). We observe that with precise type information, entity disambiguation becomes a straightforward task. However, fine-grained type information is usually not available in biomedical domain. Thus, we propose LATTE, a LATent Type Entity Linking model, that improves entity linking by modeling the latent fine-grained type information about mentions and entities. Unlike previous methods that perform entity linking directly between the mentions and the entities, LATTE jointly does entity disambiguation, and latent fine-grained type learning, without direct supervision. We evaluate our model on two biomedical datasets: MedMentions, a large scale public dataset annotated with UMLS concepts, and a de-identified corpus of dictated doctor's notes that has been annotated with ICD concepts. Extensive experimental evaluation shows our model achieves significant performance improvements over several state-of-the-art techniques.

Kristjan Arumae, Parminder Bhatia, Fei Liu

Highlighting is a powerful tool to pick out important content and emphasize. Creating summary highlights at the sub-sentence level is particularly desirable, because sub-sentences are more concise than whole sentences. They are also better suited than individual words and phrases that can potentially lead to disfluent, fragmented summaries. In this paper we seek to generate summary highlights by annotating summary-worthy sub-sentences and teaching classifiers to do the same. We frame the task as jointly selecting important sentences and identifying a single most informative textual unit from each sentence. This formulation dramatically reduces the task complexity involved in sentence compression. Our study provides new benchmarks and baselines for generating highlights at the sub-sentence level.

Gaurav Singh, Parminder Bhatia

Relation Extraction (RE) aims to label relations between groups of marked entities in raw text. Most current RE models learn context-aware representations of the target entities that are then used to establish relation between them. This works well for intra-sentence RE and we call them first-order relations. However, this methodology can sometimes fail to capture complex and long dependencies. To address this, we hypothesize that at times two target entities can be explicitly connected via a context token. We refer to such indirect relations as second-order relations and describe an efficient implementation for computing them. These second-order relation scores are then combined with first-order relation scores. Our empirical results show that the proposed method leads to state-of-the-art performance over two biomedical datasets.

Parminder Bhatia, Kristjan Arumae, Busra Celikkaya

State-of-the-art named entity recognition (NER) systems have been improving continuously using neural architectures over the past several years. However, many tasks including NER require large sets of annotated data to achieve such performance. In particular, we focus on NER from clinical notes, which is one of the most fundamental and critical problems for medical text analysis. Our work centers on effectively adapting these neural architectures towards low-resource settings using parameter transfer methods. We complement a standard hierarchical NER model with a general transfer learning framework consisting of parameter sharing between the source and target tasks, and showcase scores significantly above the baseline architecture. These sharing schemes require an exponential search over tied parameter sets to generate an optimal configuration. To mitigate the problem of exhaustively searching for model optimization, we propose the Dynamic Transfer Networks (DTN), a gated architecture which learns the appropriate parameter sharing scheme between source and target datasets. DTN achieves the improvements of the optimized transfer learning framework with just a single training setting, effectively removing the need for exponential search.

Parminder Bhatia, Busra Celikkaya, Mohammed Khalilia

Negative medical findings are prevalent in clinical reports, yet discriminating them from positive findings remains a challenging task for information extraction. Most of the existing systems treat this task as a pipeline of two separate tasks, i.e., named entity recognition (NER) and rule-based negation detection. We consider this as a multi-task problem and present a novel end-to-end neural model to jointly extract entities and negations. We extend a standard hierarchical encoder-decoder NER model and first adopt a shared encoder followed by separate decoders for the two tasks. This architecture performs considerably better than the previous rule-based and machine learning-based systems. To overcome the problem of increased parameter size especially for low-resource settings, we propose the Conditional Softmax Shared Decoder architecture which achieves state-of-art results for NER and negation detection on the 2010 i2b2/VA challenge dataset and a proprietary de-identified clinical dataset.

Mengqi Jin, Mohammad Taha Bahadori, Aaron Colak et al.

Clinical text provides essential information to estimate the acuity of a patient during hospital stays in addition to structured clinical data. In this study, we explore how clinical text can complement a clinical predictive learning task. We leverage an internal medical natural language processing service to perform named entity extraction and negation detection on clinical notes and compose selected entities into a new text corpus to train document representations. We then propose a multimodal neural network to jointly train time series signals and unstructured clinical text representations to predict the in-hospital mortality risk for ICU patients. Our model outperforms the benchmark by 2% AUC.

Parminder Bhatia, Marsal Gavalda, Arash Einolghozati

While liking or upvoting a post on a mobile app is easy to do, replying with a written note is much more difficult, due to both the cognitive load of coming up with a meaningful response as well as the mechanics of entering the text. Here we present a novel textual reply generation model that goes beyond the current auto-reply and predictive text entry models by taking into account the content preferences of the user, the idiosyncrasies of their conversational style, and even the structure of their social graph. Specifically, we have developed two types of models for personalized user interactions: a content-based conversation model, which makes use of location together with user information, and a social-graph-based conversation model, which combines content-based conversation models with social graphs.

Parminder Bhatia, Robert Guthrie, Jacob Eisenstein

Word embeddings allow natural language processing systems to share statistical information across related words. These embeddings are typically based on distributional statistics, making it difficult for them to generalize to rare or unseen words. We propose to improve word embeddings by incorporating morphological information, capturing shared sub-word features. Unlike previous work that constructs word embeddings directly from morphemes, we combine morphological and distributional information in a unified probabilistic framework, in which the word embedding is a latent variable. The morphological information provides a prior distribution on the latent word embeddings, which in turn condition a likelihood function over an observed corpus. This approach yields improvements on intrinsic word similarity evaluations, and also in the downstream task of part-of-speech tagging.

Parminder Bhatia, Yangfeng Ji, Jacob Eisenstein

Discourse structure is the hidden link between surface features and document-level properties, such as sentiment polarity. We show that the discourse analyses produced by Rhetorical Structure Theory (RST) parsers can improve document-level sentiment analysis, via composition of local information up the discourse tree. First, we show that reweighting discourse units according to their position in a dependency representation of the rhetorical structure can yield substantial improvements on lexicon-based sentiment analysis. Next, we present a recursive neural network over the RST structure, which offers significant improvements over classification-based methods.