Wenqi Shi

Ran Xu, Hejie Cui, Yue Yu et al. · gatech

Clinical natural language processing requires methods that can address domain-specific challenges, such as complex medical terminology and clinical contexts. Recently, large language models (LLMs) have shown promise in this domain. Yet, their direct deployment can lead to privacy issues and are constrained by resources. To address this challenge, we delve into synthetic clinical text generation using LLMs for clinical NLP tasks. We propose an innovative, resource-efficient approach, ClinGen, which infuses knowledge into the process. Our model involves clinical knowledge extraction and context-informed LLM prompting. Both clinical topics and writing styles are drawn from external domain-specific knowledge graphs and LLMs to guide data generation. Our extensive empirical study across 7 clinical NLP tasks and 16 datasets reveals that ClinGen consistently enhances performance across various tasks, effectively aligning the distribution of real datasets and significantly enriching the diversity of generated training instances. Our code is available at \url{https://github.com/ritaranx/ClinGen}.

Yiyou Sun, Xinyang Han, Weichen Zhang et al.

Recent AI systems have achieved strong results on a wide range of benchmarks, yet these gains have not translated into economically meaningful deployment across many professional domains. We argue that this gap is largely an evaluation problem: widely used benchmarks lack sustained performance measurement on real and economically valuable workflows. This paper introduces Agents' Last Exam (ALE), a benchmark designed to evaluate AI agents on long-horizon, economically valuable, real-world tasks with verifiable outcomes. Developed in collaboration with 250+ industry experts, ALE covers non-physical industries defined with reference to O*NET / SOC 2018 (the U.S. federal occupational taxonomy). It is organized around a task taxonomy with 55 subfields grouped into 13 industry clusters covering 1K+ tasks. Current results show that the hardest tier remains far from saturated: across mainstream harness and backbone configurations, the average full pass rate is 2.6%. ALE is designed as a living benchmark: its task pool grows continuously as new workflows and industries are onboarded. More broadly, ALE is intended not merely as another leaderboard, but as an instrument for closing the gap between benchmark success and GDP-relevant impact.

Shuang Zeng, Lei Zhu, Xinliang Zhang et al. · pku

Medical image segmentation is a fundamental yet challenging task due to the arduous process of acquiring large volumes of high-quality labeled data from experts. Contrastive learning offers a promising but still problematic solution to this dilemma. Firstly existing medical contrastive learning strategies focus on extracting image-level representation, which ignores abundant multi-level representations. Furthermore they underutilize the decoder either by random initialization or separate pre-training from the encoder, thereby neglecting the potential collaboration between the encoder and decoder. To address these issues, we propose a novel multi-level asymmetric contrastive learning framework named MACL for volumetric medical image segmentation pre-training. Specifically, we design an asymmetric contrastive learning structure to pre-train encoder and decoder simultaneously to provide better initialization for segmentation models. Moreover, we develop a multi-level contrastive learning strategy that integrates correspondences across feature-level, image-level, and pixel-level representations to ensure the encoder and decoder capture comprehensive details from representations of varying scales and granularities during the pre-training phase. Finally, experiments on 8 medical image datasets indicate our MACL framework outperforms existing 11 contrastive learning strategies. i.e. Our MACL achieves a superior performance with more precise predictions from visualization figures and 1.72%, 7.87%, 2.49% and 1.48% Dice higher than previous best results on ACDC, MMWHS, HVSMR and CHAOS with 10% labeled data, respectively. And our MACL also has a strong generalization ability among 5 variant U-Net backbones. Our code will be released at https://github.com/stevezs315/MACL.

Yuxing Lu, Yushuhong Lin, Wenqi Shi et al.

Clinical practice is not the selection of an answer from enumerated options: a physician gathers heterogeneous information incrementally and commits to sequential, irreversible decisions under uncertainty. Static benchmarks cannot probe and existing interactive medical benchmarks each compromise on at least one of them. We present ClinEnv, an interactive benchmark that evaluates LLMs as attending physicians over real inpatient admissions under a paradigm we term Longitudinal Inpatient Simulation. Each case is automatically constructed into an ordered sequence of decision stages; at every stage the model must actively query four specialized agents before committing to medications, procedures, and diagnoses. ClinEnv scores both what the model decides, through deterministic ontology-grounded matching, and how it gathers information. Across seven models, the strongest reaches only 0.31 decision F1, and outcome quality is sharply decoupled from process quality. Difficulty concentrates in management decisions and later stages, where models recover discharge diagnoses far more reliably than management actions (0.51 vs. 0.17 F1) and continue to issue redundant queries as cases progress. ClinEnv makes this information-acquisition gap, invisible to outcome-only evaluation, directly measurable.

Wenqi Shi, Sheng Zhou, Zhisheng Niu et al.

Graphics processing units (GPUs) can improve deep neural network inference throughput via batch processing, where multiple tasks are concurrently processed. We focus on novel scenarios that the energy-constrained mobile devices offload inference tasks to an edge server with GPU. The inference task is partitioned into sub-tasks for a finer granularity of offloading and scheduling, and the user energy consumption minimization problem under inference latency constraints is investigated. To deal with the coupled offloading and scheduling introduced by concurrent batch processing, we first consider an offline problem with a constant edge inference latency and the same latency constraint. It is proven that optimizing the offloading policy of each user independently and aggregating all the same sub-tasks in one batch is optimal, and thus the independent partitioning and same sub-task aggregating (IP-SSA) algorithm is inspired. Further, the optimal grouping (OG) algorithm is proposed to optimally group tasks when the latency constraints are different. Finally, when future task arrivals cannot be precisely predicted, a deep deterministic policy gradient (DDPG) agent is trained to call OG. Experiments show that IP-SSA reduces up to 94.9\% user energy consumption in the offline setting, while DDPG-OG outperforms DDPG-IP-SSA by up to 8.92\% in the online setting.

Meng Lu, Yuxing Lu, Yuchen Zhuang et al.

Vision language models (VLMs) achieve strong performance on general image understanding but struggle to think with medical images, especially when performing multi-step reasoning through iterative visual interaction. Medical VLMs often rely on static visual embeddings and single-pass inference, preventing models from re-examining, verifying, or refining visual evidence during reasoning. While tool-integrated reasoning offers a promising path forward, open-source VLMs lack the training infrastructure to learn effective tool selection, invocation, and coordination in multi-modal medical reasoning. We introduce MedVistaGym, a scalable and interactive training environment that incentivizes tool-integrated visual reasoning for medical image analysis. MedVistaGym equips VLMs to determine when and which tools to invoke, localize task-relevant image regions, and integrate single or multiple sub-image evidence into interleaved multimodal reasoning within a unified, executable interface for agentic training. Using MedVistaGym, we train MedVistaGym-R1 to interleave tool use with agentic reasoning through trajectory sampling and end-to-end reinforcement learning. Across six medical VQA benchmarks, MedVistaGym-R1-8B exceeds comparably sized tool-augmented baselines by 19.10% to 24.21%, demonstrating that structured agentic training--not tool access alone--unlocks effective tool-integrated reasoning for medical image analysis.

Jingwei Huang, Kuroush Nezafati, Zhikai Chi et al. · gatech

Extracting structured information from clinical notes requires navigating a dense web of interdependent variables where the value of one attribute logically constrains others. Existing Large Language Model (LLM)-based extraction pipelines often struggle to capture these dependencies, leading to clinically inconsistent outputs. We propose deep reflective reasoning, a large language model agent framework that iteratively self-critiques and revises structured outputs by checking consistency among variables, the input text, and retrieved domain knowledge, stopping when outputs converge. We extensively evaluate the proposed method in three diverse oncology applications: (1) On colorectal cancer synoptic reporting from gross descriptions (n=217), reflective reasoning improved average F1 across eight categorical synoptic variables from 0.828 to 0.911 and increased mean correct rate across four numeric variables from 0.806 to 0.895; (2) On Ewing sarcoma CD99 immunostaining pattern identification (n=200), the accuracy improved from 0.870 to 0.927; (3) On lung cancer tumor staging (n=100), tumor stage accuracy improved from 0.680 to 0.833 (pT: 0.842 -> 0.884; pN: 0.885 -> 0.948). The results demonstrate that deep reflective reasoning can systematically improve the reliability of LLM-based structured data extraction under interdependence constraints, enabling more consistent machine-operable clinical datasets and facilitating knowledge discovery with machine learning and data science towards digital health.

Ran Xu, Wenqi Shi, Yue Yu et al. · gatech

Developing effective biomedical retrieval models is important for excelling at knowledge-intensive biomedical tasks but still challenging due to the deficiency of sufficient publicly annotated biomedical data and computational resources. We present BMRetriever, a series of dense retrievers for enhancing biomedical retrieval via unsupervised pre-training on large biomedical corpora, followed by instruction fine-tuning on a combination of labeled datasets and synthetic pairs. Experiments on 5 biomedical tasks across 11 datasets verify BMRetriever's efficacy on various biomedical applications. BMRetriever also exhibits strong parameter efficiency, with the 410M variant outperforming baselines up to 11.7 times larger, and the 2B variant matching the performance of models with over 5B parameters. The training data and model checkpoints are released at \url{https://huggingface.co/BMRetriever} to ensure transparency, reproducibility, and application to new domains.

Xiangru Tang, Daniel Shao, Jiwoong Sohn et al.

Large Language Models (LLMs) have shown impressive performance on existing medical question-answering benchmarks. This high performance makes it increasingly difficult to meaningfully evaluate and differentiate advanced methods. We present MedAgentsBench, a benchmark that focuses on challenging medical questions requiring multi-step clinical reasoning, diagnosis formulation, and treatment planning-scenarios where current models still struggle despite their strong performance on standard tests. Drawing from seven established medical datasets, our benchmark addresses three key limitations in existing evaluations: (1) the prevalence of straightforward questions where even base models achieve high performance, (2) inconsistent sampling and evaluation protocols across studies, and (3) lack of systematic analysis of the interplay between performance, cost, and inference time. Through experiments with various base models and reasoning methods, we demonstrate that the latest thinking models, DeepSeek R1 and OpenAI o3, exhibit exceptional performance in complex medical reasoning tasks. Additionally, advanced search-based agent methods offer promising performance-to-cost ratios compared to traditional approaches. Our analysis reveals substantial performance gaps between model families on complex questions and identifies optimal model selections for different computational constraints. Our benchmark and evaluation framework are publicly available at https://github.com/gersteinlab/medagents-benchmark.

Ran Xu, Wenqi Shi, Yue Yu et al. · gatech

We present RAM-EHR, a Retrieval AugMentation pipeline to improve clinical predictions on Electronic Health Records (EHRs). RAM-EHR first collects multiple knowledge sources, converts them into text format, and uses dense retrieval to obtain information related to medical concepts. This strategy addresses the difficulties associated with complex names for the concepts. RAM-EHR then augments the local EHR predictive model co-trained with consistency regularization to capture complementary information from patient visits and summarized knowledge. Experiments on two EHR datasets show the efficacy of RAM-EHR over previous knowledge-enhanced baselines (3.4% gain in AUROC and 7.2% gain in AUPR), emphasizing the effectiveness of the summarized knowledge from RAM-EHR for clinical prediction tasks. The code will be published at \url{https://github.com/ritaranx/RAM-EHR}.

Junchao Cui, Wenqi Shi, Shaoyong Du et al.

Worldwide image geo-localization aims to infer the geographic location of an image captured anywhere on Earth, spanning street, city, regional, national, and continental scales. Existing methods rely on visual features that are sensitive to environmental variations (e.g., lighting, season, and weather) and lack effective post-processing to filter outlier candidates, limiting localization accuracy. To address these limitations, we propose DualGeo, a two-stage framework for worldwide image geo-localization. First, it establishes a geo-representational foundation by fusing image and semantic segmentation features via bidirectional cross-attention. The fused features are then aligned with GPS coordinates through dual-view contrastive learning to build a global retrieval database. Second, it performs geo-cognitive refinement by re-ranking retrieved candidates using geographic clustering. It then feeds them into large multimodal models (LMMs) for final coordinate prediction. Experiments on IM2GPS, IM2GPS3k, and YFCC4k show that DualGeo outperforms state-of-the-art methods, improving street-level (<1 km) and city-level (<25 km) localization accuracy by 3.6%-16.58% and 1.29%-8.77%, respectively. Our code and datasets are available : https://github.com/CJ310177/DualGeo.

Ran Xu, Wenqi Shi, Yuchen Zhuang et al. · gatech

Retrieval-Augmented Generation (RAG) systems often struggle to handle multi-hop question-answering tasks accurately due to irrelevant context retrieval and limited complex reasoning capabilities. We introduce Collab-RAG, a collaborative training framework that leverages mutual enhancement between a white-box small language model (SLM) and a blackbox large language model (LLM) for RAG. Specifically, the SLM decomposes complex queries into simpler sub-questions, thus enhancing the accuracy of the retrieval and facilitating more effective reasoning by the black-box LLM. Concurrently, the black-box LLM provides feedback signals to improve the SLM's decomposition capability. We observe that Collab-RAG relies solely on supervision from an affordable black-box LLM without additional distillation from frontier LLMs, yet demonstrates strong generalization across multiple black-box LLMs. Experimental evaluations across five multi-hop QA datasets demonstrate that Collab-RAG substantially outperforms existing black-box-only and SLM fine-tuning baselines by 1.8%-14.2% on average. In particular, our fine-tuned 3B SLM surpasses a frozen 32B LLM in question decomposition, highlighting the efficiency of Collab-RAG in improving reasoning and retrieval for complex questions. The code of Collab-RAG is available on https://github.com/ritaranx/Collab-RAG/.

Xiangru Tang, Zhuoyun Yu, Jiapeng Chen et al.

Virtual cell modeling represents an emerging frontier at the intersection of artificial intelligence and biology, aiming to predict quantities such as responses to diverse perturbations quantitatively. However, autonomously building computational models for virtual cells is challenging due to the complexity of biological systems, the heterogeneity of data modalities, and the need for domain-specific expertise across multiple disciplines. Here, we introduce CellForge, an agentic system that leverages a multi-agent framework that transforms presented biological datasets and research objectives directly into optimized computational models for virtual cells. More specifically, given only raw single-cell multi-omics data and task descriptions as input, CellForge outputs both an optimized model architecture and executable code for training virtual cell models and inference. The framework integrates three core modules: Task Analysis for presented dataset characterization and relevant literature retrieval, Method Design, where specialized agents collaboratively develop optimized modeling strategies, and Experiment Execution for automated generation of code. The agents in the Design module are separated into experts with differing perspectives and a central moderator, and have to collaboratively exchange solutions until they achieve a reasonable consensus. We demonstrate CellForge's capabilities in single-cell perturbation prediction, using six diverse datasets that encompass gene knockouts, drug treatments, and cytokine stimulations across multiple modalities. CellForge consistently outperforms task-specific state-of-the-art methods. Overall, CellForge demonstrates how iterative interaction between LLM agents with differing perspectives provides better solutions than directly addressing a modeling challenge. Our code is publicly available at https://github.com/gersteinlab/CellForge.

Ran Xu, Yuchen Zhuang, Yue Yu et al. · gatech

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by integrating external knowledge retrieved at inference time. While RAG demonstrates strong performance on benchmarks largely derived from general-domain corpora like Wikipedia, its effectiveness under realistic, diverse retrieval scenarios remains underexplored. We evaluated RAG systems using MassiveDS, a large-scale datastore with mixture of knowledge, and identified critical limitations: retrieval mainly benefits smaller models, rerankers add minimal value, and no single retrieval source consistently excels. Moreover, current LLMs struggle to route queries across heterogeneous knowledge sources. These findings highlight the need for adaptive retrieval strategies before deploying RAG in real-world settings. Our code and data can be found at https://github.com/ritaranx/RAG_in_the_Wild.

Ran Xu, Yuchen Zhuang, Zihan Dong et al. · gatech

Search-augmented LLMs often struggle with complex reasoning tasks due to ineffective multi-hop retrieval and limited reasoning ability. We propose AceSearcher, a cooperative self-play framework that trains a single large language model (LLM) to alternate between two roles: a decomposer that breaks down complex queries and a solver that integrates retrieved contexts for answer generation. AceSearcher couples supervised fine-tuning on a diverse mixture of search, reasoning, and decomposition tasks with reinforcement fine-tuning optimized for final answer accuracy, eliminating the need for intermediate annotations. Extensive experiments on three reasoning-intensive tasks across 10 datasets show that AceSearcher outperforms state-of-the-art baselines, achieving an average exact match improvement of 7.6%. Remarkably, on document-level finance reasoning tasks, AceSearcher-32B matches the performance of the DeepSeek-V3 model using less than 5% of its parameters. Even at smaller scales (1.5B and 8B), AceSearcher often surpasses existing search-augmented LLMs with up to 9x more parameters, highlighting its exceptional efficiency and effectiveness in tackling complex reasoning tasks. Our code will be published at https://github.com/ritaranx/AceSearcher and https://huggingface.co/AceSearcher.

Wenqi Shi, Wenkai Xu

Learning causal relationships is a fundamental problem in science. Anchor regression has been developed to address this problem for a large class of causal graphical models, though the relationships between the variables are assumed to be linear. In this work, we tackle the nonlinear setting by proposing kernel anchor regression (KAR). Beyond the natural formulation using a classic two-stage least square estimator, we also study an improved variant that involves nonparametric regression in three separate stages. We provide convergence results for the proposed KAR estimators and the identifiability conditions for KAR to learn the nonlinear structural equation models (SEM). Experimental results demonstrate the superior performances of the proposed KAR estimators over existing baselines.

Meng Lu, Ran Xu, Yi Fang et al.

While recent vision-language models (VLMs) demonstrate strong image understanding, their ability to "think with images", i.e., to reason through multi-step visual interactions, remains limited. We introduce VISTA-Gym, a scalable training environment for incentivizing tool-integrated visual reasoning capabilities in VLMs. VISTA-Gym unifies diverse real-world multimodal reasoning tasks (7 tasks from 13 datasets in total) with a standardized interface for visual tools (e.g., grounding, parsing), executable interaction loops, verifiable feedback signals, and efficient trajectory logging, enabling visual agentic reinforcement learning at scale. While recent VLMs exhibit strong text-only reasoning, both proprietary and open-source models still struggle with tool selection, invocation, and coordination. With VISTA-Gym, we train VISTA-R1 to interleave tool-use with agentic reasoning via multi-turn trajectory sampling and end-to-end reinforcement learning. Extensive experiments across 11 public reasoning-intensive VQA benchmarks show that VISTA-R1-8B outperforms state-of-the-art baselines with similar sizes by 9.51%-18.72%, demonstrating VISTA-Gym as an effective training ground to unlock the tool-integrated reasoning capabilities for VLMs.

J. Ben Tamo, Nishant S. Chouhan, Micky C. Nnamdi et al.

Surgical decision-making is complex and requires understanding causal relationships between patient characteristics, interventions, and outcomes. In high-stakes settings like spinal fusion or scoliosis correction, accurate estimation of individualized treatment effects (ITEs) remains limited due to the reliance on traditional statistical methods that struggle with complex, heterogeneous data. In this study, we develop a multi-task meta-learning framework, X-MultiTask, for ITE estimation that models each surgical decision (e.g., anterior vs. posterior approach, surgery vs. no surgery) as a distinct task while learning shared representations across tasks. To strengthen causal validity, we incorporate the inverse probability weighting (IPW) into the training objective. We evaluate our approach on two datasets: (1) a public spinal fusion dataset (1,017 patients) to assess the effect of anterior vs. posterior approaches on complication severity; and (2) a private AIS dataset (368 patients) to analyze the impact of posterior spinal fusion (PSF) vs. non-surgical management on patient-reported outcomes (PROs). Our model achieves the highest average AUC (0.84) in the anterior group and maintains competitive performance in the posterior group (0.77). It outperforms baselines in treatment effect estimation with the lowest overall $ε_{\text{NN-PEHE}}$ (0.2778) and $ε_{\text{ATE}}$ (0.0763). Similarly, when predicting PROs in AIS, X-MultiTask consistently shows superior performance across all domains, with $ε_{\text{NN-PEHE}}$ = 0.2551 and $ε_{\text{ATE}}$ = 0.0902. By providing robust, patient-specific causal estimates, X-MultiTask offers a powerful tool to advance personalized surgical care and improve patient outcomes. The code is available at https://github.com/Wizaaard/X-MultiTask.

Ran Xu, Yuchen Zhuang, Yishan Zhong et al. · gatech

We introduce MedAgentGym, a scalable and interactive training environment designed to enhance coding-based biomedical reasoning capabilities in large language model (LLM) agents. MedAgentGym comprises 72,413 task instances across 129 categories derived from 12 authentic real-world biomedical scenarios. Tasks are encapsulated within executable sandbox environments, each featuring detailed task specifications, interactive feedback mechanisms, verifiable ground truth annotations, and scalable training trajectory generation. Extensive benchmarking of 29 LLMs reveals substantial performance disparities in biomedical data science between commercial and open-source LLMs. Leveraging efficient multi-threaded and multi-turn trajectory sampling in MedAgentGym, Med-Copilot achieves performance gains of +43.02% and +45.28% from offline and online reinforcement learning, respectively, demonstrating MedAgentGym as an effective training ground while establishing itself as a cost-effective, privacy-preserving alternative competitive with proprietary LLMs (gpt-4o). By offering a unified execution environment with a comprehensive benchmark and accessible, extensible training resources, MedAgentGym delivers an integrated platform to develop LLM-based coding assistants for advanced biomedical data science.

Shuang Zeng, Chee Hong Lee, Micky C Nnamdi et al. · pku

Retinal vessel segmentation is a vital early detection method for several severe ocular diseases. Despite significant progress in retinal vessel segmentation with the advancement of Neural Networks, there are still challenges to overcome. Specifically, retinal vessel segmentation aims to predict the class label for every pixel within a fundus image, with a primary focus on intra-image discrimination, making it vital for models to extract more discriminative features. Nevertheless, existing methods primarily focus on minimizing the difference between the output from the decoder and the label, but ignore fully using feature-level fine-grained representations from the encoder. To address these issues, we propose a novel Attention U-shaped Kolmogorov-Arnold Network named AttUKAN along with a novel Label-guided Pixel-wise Contrastive Loss for retinal vessel segmentation. Specifically, we implement Attention Gates into Kolmogorov-Arnold Networks to enhance model sensitivity by suppressing irrelevant feature activations and model interpretability by non-linear modeling of KAN blocks. Additionally, we also design a novel Label-guided Pixel-wise Contrastive Loss to supervise our proposed AttUKAN to extract more discriminative features by distinguishing between foreground vessel-pixel pairs and background pairs. Experiments are conducted across four public datasets including DRIVE, STARE, CHASE_DB1, HRF and our private dataset. AttUKAN achieves F1 scores of 82.50%, 81.14%, 81.34%, 80.21% and 80.09%, along with MIoU scores of 70.24%, 68.64%, 68.59%, 67.21% and 66.94% in the above datasets, which are the highest compared to 11 networks for retinal vessel segmentation. Quantitative and qualitative results show that our AttUKAN achieves state-of-the-art performance and outperforms existing retinal vessel segmentation methods. Our code will be available at https://github.com/stevezs315/AttUKAN.

Jingwei Huang, Kuroush Nezafati, Ismael Villanueva-Miranda et al.

This study introduces a LLMs powered multiagent ensemble method to address challenges in hallucination and data labeling, particularly in large-scale EHR datasets. Manual labeling of such datasets requires domain expertise and is labor-intensive, time-consuming, expensive, and error-prone. To overcome this bottleneck, we developed an ensemble LLMs method and demonstrated its effectiveness in two real-world tasks: (1) labeling a large-scale unlabeled ECG dataset in MIMIC-IV; (2) identifying social determinants of health (SDOH) from the clinical notes of EHR. Trading off benefits and cost, we selected a pool of diverse open source LLMs with satisfactory performance. We treat each LLM's prediction as a vote and apply a mechanism of majority voting with minimal winning threshold for ensemble. We implemented an ensemble LLMs application for EHR data labeling tasks. By using the ensemble LLMs and natural language processing, we labeled MIMIC-IV ECG dataset of 623,566 ECG reports with an estimated accuracy of 98.2%. We applied the ensemble LLMs method to identify SDOH from social history sections of 1,405 EHR clinical notes, also achieving competitive performance. Our experiments show that the ensemble LLMs can outperform individual LLM even the best commercial one, and the method reduces hallucination errors. From the research, we found that (1) the ensemble LLMs method significantly reduces the time and effort required for labeling large-scale EHR data, automating the process with high accuracy and quality; (2) the method generalizes well to other text data labeling tasks, as shown by its application to SDOH identification; (3) the ensemble of a group of diverse LLMs can outperform or match the performance of the best individual LLM; and (4) the ensemble method substantially reduces hallucination errors. This approach provides a scalable and efficient solution to data-labeling challenges.

Amritpal Singh, Wenqi Shi, May D Wang

In the context of surgery, robots can provide substantial assistance by performing small, repetitive tasks such as suturing, needle exchange, and tissue retraction, thereby enabling surgeons to concentrate on more complex aspects of the procedure. However, existing surgical task learning mainly pertains to rigid body interactions, whereas the advancement towards more sophisticated surgical robots necessitates the manipulation of soft bodies. Previous work focused on tissue phantoms for soft tissue task learning, which can be expensive and can be an entry barrier to research. Simulation environments present a safe and efficient way to learn surgical tasks before their application to actual tissue. In this study, we create a Robot Operating System (ROS)-compatible physics simulation environment with support for both rigid and soft body interactions within surgical tasks. Furthermore, we investigate the soft tissue interactions facilitated by the patient-side manipulator of the DaVinci surgical robot. Leveraging the pybullet physics engine, we simulate kinematics and establish anchor points to guide the robotic arm when manipulating soft tissue. Using demonstration-guided reinforcement learning (RL) algorithms, we investigate their performance in comparison to traditional reinforcement learning algorithms. Our in silico trials demonstrate a proof-of-concept for autonomous surgical soft tissue retraction. The results corroborate the feasibility of learning soft body manipulation through the application of reinforcement learning agents. This work lays the foundation for future research into the development and refinement of surgical robots capable of managing both rigid and soft tissue interactions. Code is available at https://github.com/amritpal-001/tissue_retract.

Hang Zhang, Ruheng Wang, Yuelyu Ji et al.

Large language models have achieved strong performance on medical reasoning benchmarks, yet their deployment in clinical settings demands rigorous verification to ensure factual accuracy. While reward models offer a scalable approach for reasoning trace verification, existing methods face two limitations: they produce only scalar reward values without explicit justification, and they rely on single-pass retrieval that precludes adaptive knowledge access as verification unfolds. We introduce $\method$, an agentic framework that addresses these limitations by training medical reasoning verifiers to iteratively query external medical corpora during evaluation. Our approach combines tool-augmented verification with an iterative reinforcement learning paradigm that requires only trace-level supervision, alongside an adaptive curriculum mechanism that dynamically adjusts training data distribution. Across four medical reasoning benchmarks, $\method$ achieves substantial gains over existing methods, improving MedQA accuracy by 23.5% and MedXpertQA by 32.0% relative to the base generator in particular. Crucially, $\method$ demonstrates an $\mathbf{8\times}$ reduction in sampling budget requirement compared to prior reward model baselines. These findings establish that grounding verification in dynamically retrieved evidence offers a principled path toward more reliable medical reasoning systems.

Wenqi Shi, Ran Xu, Yuchen Zhuang et al. · gatech

Large language models (LLMs) have demonstrated exceptional capabilities in planning and tool utilization as autonomous agents, but few have been developed for medical problem-solving. We propose EHRAgent, an LLM agent empowered with a code interface, to autonomously generate and execute code for multi-tabular reasoning within electronic health records (EHRs). First, we formulate an EHR question-answering task into a tool-use planning process, efficiently decomposing a complicated task into a sequence of manageable actions. By integrating interactive coding and execution feedback, EHRAgent learns from error messages and improves the originally generated code through iterations. Furthermore, we enhance the LLM agent by incorporating long-term memory, which allows EHRAgent to effectively select and build upon the most relevant successful cases from past experiences. Experiments on three real-world multi-tabular EHR datasets show that EHRAgent outperforms the strongest baseline by up to 29.6% in success rate. EHRAgent leverages the emerging few-shot learning capabilities of LLMs, enabling autonomous code generation and execution to tackle complex clinical tasks with minimal demonstrations.

Wenqi Shi, Ran Xu, Yuchen Zhuang et al. · gatech

Despite their improved capabilities in generation and reasoning, adapting large language models (LLMs) to the biomedical domain remains challenging due to their immense size and corporate privacy. In this work, we propose MedAdapter, a unified post-hoc adapter for test-time adaptation of LLMs towards biomedical applications. Instead of fine-tuning the entire LLM, MedAdapter effectively adapts the original model by fine-tuning only a small BERT-sized adapter to rank candidate solutions generated by LLMs. Experiments demonstrate that MedAdapter effectively adapts both white-box and black-box LLMs in biomedical reasoning, achieving average performance improvements of 25.48% and 11.31%, respectively, without requiring extensive computational resources or sharing data with third parties. MedAdapter also yields superior performance when combined with train-time adaptation, highlighting a flexible and complementary solution to existing adaptation methods. Faced with the challenges of balancing model performance, computational resources, and data privacy, MedAdapter provides an efficient, privacy-preserving, cost-effective, and transparent solution for adapting LLMs to the biomedical domain.

Zhenyu Han, Ansheng You, Haibo Wang et al.

Reinforcement learning (RL) has become a pivotal technology in the post-training phase of large language models (LLMs). Traditional task-colocated RL frameworks suffer from significant scalability bottlenecks, while task-separated RL frameworks face challenges in complex dataflows and the corresponding resource idling and workload imbalance. Moreover, most existing frameworks are tightly coupled with LLM training or inference engines, making it difficult to support custom-designed engines. To address these challenges, we propose AsyncFlow, an asynchronous streaming RL framework for efficient post-training. Specifically, we introduce a distributed data storage and transfer module that provides a unified data management and fine-grained scheduling capability in a fully streamed manner. This architecture inherently facilitates automated pipeline overlapping among RL tasks and dynamic load balancing. Moreover, we propose a producer-consumer-based asynchronous workflow engineered to minimize computational idleness by strategically deferring parameter update process within staleness thresholds. Finally, the core capability of AsynFlow is architecturally decoupled from underlying training and inference engines and encapsulated by service-oriented user interfaces, offering a modular and customizable user experience. Extensive experiments demonstrate an average of 1.59 throughput improvement compared with state-of-the-art baseline. The presented architecture in this work provides actionable insights for next-generation RL training system designs.

Yuchen Zhuang, Di Jin, Jiaao Chen et al. · gatech

Large language models (LLMs)-empowered web agents enables automating complex, real-time web navigation tasks in enterprise environments. However, existing web agents relying on supervised fine-tuning (SFT) often struggle with generalization and robustness due to insufficient reasoning capabilities when handling the inherently dynamic nature of web interactions. In this study, we introduce WorkForceAgent-R1, an LLM-based web agent trained using a rule-based R1-style reinforcement learning framework designed explicitly to enhance single-step reasoning and planning for business-oriented web navigation tasks. We employ a structured reward function that evaluates both adherence to output formats and correctness of actions, enabling WorkForceAgent-R1 to implicitly learn robust intermediate reasoning without explicit annotations or extensive expert demonstrations. Extensive experiments on the WorkArena benchmark demonstrate that WorkForceAgent-R1 substantially outperforms SFT baselines by 10.26-16.59%, achieving competitive performance relative to proprietary LLM-based agents (gpt-4o) in workplace-oriented web navigation tasks.

Jinrui Fang, Runhan Chen, Xu Yang et al.

Large language models (LLMs) achieve high accuracy in medical diagnosis when all clinical information is provided in a single turn, yet how they behave under multi-turn evidence accumulation closer to real clinical reasoning remains unexplored. We introduce MINT (Medical Incremental N-Turn Benchmark), a high-fidelity, multi-turn medical diagnosis benchmark comprising 1,035 cases with clinically labeled evidence shards, controlled turn granularity, and information-preserving decomposition. Through systematic evaluation of 11 LLMs on MINT, we uncover three persistent behavioral patterns that significantly impact diagnostic decisions: (1) intent to answer, models rush to answer before sufficient evidence has been observed, with over 55% of answers committed within the first two turns; (2) self-correction, incorrect-to-correct answer revisions occur at up to 10.6 times the rate of correct-to-incorrect flips, revealing a latent capacity for self-correction that premature commitment forecloses; and (3) strong lures, clinically salient information such as laboratory results trigger premature answering even when models are explicitly instructed to wait. We translate these findings into clinically actionable guidance: deferring the diagnostic question to later turns reduces premature answering and improves accuracy at the first point of commitment by up to 62.6%, while reserving salient clinical evidence for later turns prevents a catastrophic accuracy drop of up to 23.3% caused by premature commitment. Our work provides both a controlled evaluation framework and concrete recommendations for improving the reliability of LLMs in multi-turn medical diagnosis.

Aaron Tu, Weihao Xuan, Heli Qi et al. · gatech

Reinforcement learning with verifiable rewards (RLVR) is a practical and scalable approach to enhancing large language models in areas such as math, code, and other structured tasks. Two questions motivate this paper: how much of the reported gains survive under strictly parity-controlled evaluation, and whether RLVR is cost-free or exacts a measurable tax. We argue that progress is real, but gains are often overstated due to three forces - an RLVR tax, evaluation pitfalls, and data contamination. Using a partial-prompt contamination audit and matched-budget reproductions across base and RL models, we show that several headline gaps shrink or vanish under clean, parity-controlled evaluation. We then propose a tax-aware training and evaluation protocol that co-optimizes accuracy, grounding, and calibrated abstention and standardizes budgeting and provenance checks. Applied to recent RLVR setups, this protocol yields more reliable estimates of reasoning gains and, in several cases, revises prior conclusions. Our position is constructive: RLVR is valuable and industry-ready; we advocate keeping its practical benefits while prioritizing reliability, safety, and measurement.

Micky C. Nnamdi, J. Ben Tamo, Wenqi Shi et al.

Problem-Based Learning (PBL) has significantly impacted biomedical engineering (BME) education since its introduction in the early 2000s, effectively enhancing critical thinking and real-world knowledge application among students. With biomedical engineering rapidly converging with artificial intelligence (AI), integrating effective AI education into established curricula has become challenging yet increasingly necessary. Recent advancements, including AI's recognition by the 2024 Nobel Prize, have highlighted the importance of training students comprehensively in biomedical AI. However, effective biomedical AI education faces substantial obstacles, such as diverse student backgrounds, limited personalized mentoring, constrained computational resources, and difficulties in safely scaling hands-on practical experiments due to privacy and ethical concerns associated with biomedical data. To overcome these issues, we conducted a three-year (2021-2023) case study implementing an advanced PBL framework tailored specifically for biomedical AI education, involving 92 undergraduate and 156 graduate students from the joint Biomedical Engineering program of Georgia Institute of Technology and Emory University. Our approach emphasizes collaborative, interdisciplinary problem-solving through authentic biomedical AI challenges. The implementation led to measurable improvements in learning outcomes, evidenced by high research productivity (16 student-authored publications), consistently positive peer evaluations, and successful development of innovative computational methods addressing real biomedical challenges. Additionally, we examined the role of generative AI both as a teaching subject and an educational support tool within the PBL framework. Our study presents a practical and scalable roadmap for biomedical engineering departments aiming to integrate robust AI education into their curricula.

Yuxing Lu, J. Ben Tamo, Weichen Zhao et al.

Large language models are strong sequence predictors, yet standard inference relies on immutable context histories. After making an error at generation step t, the model lacks an updatable memory mechanism that improves predictions for step t+1. We propose LLM-as-RNN, an inference-only framework that turns a frozen LLM into a recurrent predictor by representing its hidden state as natural-language memory. This state, implemented as a structured system-prompt summary, is updated at each timestep via feedback-driven text rewrites, enabling learning without parameter updates. Under a fixed token budget, LLM-as-RNN corrects errors and retains task-relevant patterns, effectively performing online learning through language. We evaluate the method on three sequential benchmarks in healthcare, meteorology, and finance across Llama, Gemma, and GPT model families. LLM-as-RNN significantly outperforms zero-shot, full-history, and MemPrompt baselines, improving predictive accuracy by 6.5% on average, while producing interpretable, human-readable learning traces absent in standard context accumulation.

Micky C Nnamdi, Wenqi Shi, Cheng Wan et al.

Obstructive sleep apnea (OSA) is a sleep disorder that affects nearly one billion people globally and significantly elevates cardiovascular risk. Traditional diagnosis through polysomnography is resource-intensive and limits widespread access, creating a critical need for accurate and efficient alternatives. In this paper, we introduce KindSleep, a deep learning framework that integrates clinical knowledge with single-channel patient-specific oximetry signals and clinical data for precise OSA diagnosis. KindSleep first learns to identify clinically interpretable concepts, such as desaturation indices and respiratory disturbance events, directly from raw oximetry signals. It then fuses these AI-derived concepts with multimodal clinical data to estimate the Apnea-Hypopnea Index (AHI). We evaluate KindSleep on three large, independent datasets from the National Sleep Research Resource (SHHS, CFS, MrOS; total n = 9,815). KindSleep demonstrates excellent performance in estimating AHI scores (R2 = 0.917, ICC = 0.957) and consistently outperforms existing approaches in classifying OSA severity, achieving weighted F1-scores from 0.827 to 0.941 across diverse populations. By grounding its predictions in a layer of clinically meaningful concepts, KindSleep provides a more transparent and trustworthy diagnostic tool for sleep medicine practices.

Huawei Bai, Yifan Huang, Wenqi Shi et al.

The training efficiency and scalability of language models on massive clusters currently remain a critical bottleneck. Mainstream approaches like ND parallelism are often cumbersome and complex, while flexible alternatives such as the Zero Redundancy Optimizer (ZeRO) are frequently hampered by communication overhead. In this paper, we propose Asynchronous Hierarchical Zero Parallelism (AsyncHZP), a novel asynchronous variant of ZeRO designed to achieve superior performance while maintaining simplicity and memory efficiency. Unlike traditional ZeRO, which employs over-fine-grained sharding that can lead to inefficient communication, AsyncHZP adaptively reshards parameters, gradients, and optimizer states across different replica groups. This strategy optimizes device memory utilization and significantly reduces communication overhead. In addition, we also design a multi-stream asynchronous scheduling method that executes parameter all-gather and gradient reduce-scatter operations in dedicated background threads, effectively overlapping communication with computation while incurring negligible memory fragmentation. Empirical evaluations on both Dense and Mixture-of-Experts (MoE) models confirm that AsyncHZP maintains robust stability at scale. It consistently outperforms classic ND parallelism, achieving state-of-the-art performance without complex strategic tuning, thereby simplifying the path to efficient large-scale training.

Matthew Chen, Micky Nnamdi, Justin Shao et al. · gatech

Foundation models for electroencephalography (EEG) signals have recently demonstrated success in learning generalized representations of EEGs, outperforming specialized models in various downstream tasks. However, many of these models lack transparency in their pretraining dynamics and offer limited insight into how well EEG information is preserved within their embeddings. For successful clinical integration, EEG foundation models must ensure transparency in pretraining, downstream fine-tuning, and the interpretability of learned representations. Current approaches primarily operate in the temporal domain, overlooking advancements in digital signal processing that enable the extraction of deterministic and traceable features, such as wavelet-based representations. We propose MENDR (Manifold Explainable Neural Data Representations), a filter bank-based EEG foundation model built on a novel Riemannian Manifold Transformer architecture to resolve these issues. MENDR learns symmetric positive definite matrix embeddings of EEG signals and is pretrained on a large corpus comprising over 4,000 hours of EEG data, decomposed via discrete wavelet packet transforms into multi-resolution coefficients. MENDR significantly enhances interpretability by visualizing symmetric positive definite embeddings as geometric ellipsoids and supports accurate reconstruction of EEG signals from learned embeddings. Evaluations across multiple clinical EEG tasks demonstrate that MENDR achieves near state-of-the-art performance with substantially fewer parameters, underscoring its potential for efficient, interpretable, and clinically applicable EEG analysis.

Felipe Giuste, Wenqi Shi, Yuanda Zhu et al.

Despite the myriad peer-reviewed papers demonstrating novel Artificial Intelligence (AI)-based solutions to COVID-19 challenges during the pandemic, few have made significant clinical impact. The impact of artificial intelligence during the COVID-19 pandemic was greatly limited by lack of model transparency. This systematic review examines the use of Explainable Artificial Intelligence (XAI) during the pandemic and how its use could overcome barriers to real-world success. We find that successful use of XAI can improve model performance, instill trust in the end-user, and provide the value needed to affect user decision-making. We introduce the reader to common XAI techniques, their utility, and specific examples of their application. Evaluation of XAI results is also discussed as an important step to maximize the value of AI-based clinical decision support systems. We illustrate the classical, modern, and potential future trends of XAI to elucidate the evolution of novel XAI techniques. Finally, we provide a checklist of suggestions during the experimental design process supported by recent publications. Common challenges during the implementation of AI solutions are also addressed with specific examples of potential solutions. We hope this review may serve as a guide to improve the clinical impact of future AI-based solutions.

Wenqi Shi, Sheng Zhou, Zhisheng Niu et al.

In federated learning (FL), devices contribute to the global training by uploading their local model updates via wireless channels. Due to limited computation and communication resources, device scheduling is crucial to the convergence rate of FL. In this paper, we propose a joint device scheduling and resource allocation policy to maximize the model accuracy within a given total training time budget for latency constrained wireless FL. A lower bound on the reciprocal of the training performance loss, in terms of the number of training rounds and the number of scheduled devices per round, is derived. Based on the bound, the accuracy maximization problem is solved by decoupling it into two sub-problems. First, given the scheduled devices, the optimal bandwidth allocation suggests allocating more bandwidth to the devices with worse channel conditions or weaker computation capabilities. Then, a greedy device scheduling algorithm is introduced, which in each step selects the device consuming the least updating time obtained by the optimal bandwidth allocation, until the lower bound begins to increase, meaning that scheduling more devices will degrade the model accuracy. Experiments show that the proposed policy outperforms state-of-the-art scheduling policies under extensive settings of data distributions and cell radius.

Wenqi Shi, Sheng Zhou, Zhisheng Niu

Owing to the increasing need for massive data analysis and model training at the network edge, as well as the rising concerns about the data privacy, a new distributed training framework called federated learning (FL) has emerged. In each iteration of FL (called round), the edge devices update local models based on their own data and contribute to the global training by uploading the model updates via wireless channels. Due to the limited spectrum resources, only a portion of the devices can be scheduled in each round. While most of the existing work on scheduling focuses on the convergence of FL w.r.t. rounds, the convergence performance under a total training time budget is not yet explored. In this paper, a joint bandwidth allocation and scheduling problem is formulated to capture the long-term convergence performance of FL, and is solved by being decoupled into two sub-problems. For the bandwidth allocation sub-problem, the derived optimal solution suggests to allocate more bandwidth to the devices with worse channel conditions or weaker computation capabilities. For the device scheduling sub-problem, by revealing the trade-off between the number of rounds required to attain a certain model accuracy and the latency per round, a greedy policy is inspired, that continuously selects the device that consumes the least time in model updating until achieving a good trade-off between the learning efficiency and latency per round. The experiments show that the proposed policy outperforms other state-of-the-art scheduling policies, with the best achievable model accuracy under training time budgets.

Wenqi Shi, Yunzhong Hou, Sheng Zhou et al.

Deep neural networks (DNNs) are state-of-the-art solutions for many machine learning applications, and have been widely used on mobile devices. Running DNNs on resource-constrained mobile devices often requires the help from edge servers via computation offloading. However, offloading through a bandwidth-limited wireless link is non-trivial due to the tight interplay between the computation resources on mobile devices and wireless resources. Existing studies have focused on cooperative inference where DNN models are partitioned at different neural network layers, and the two parts are executed at the mobile device and the edge server, respectively. Since the output data size of a DNN layer can be larger than that of the raw data, offloading intermediate data between layers can suffer from high transmission latency under limited wireless bandwidth. In this paper, we propose an efficient and flexible 2-step pruning framework for DNN partition between mobile devices and edge servers. In our framework, the DNN model only needs to be pruned once in the training phase where unimportant convolutional filters are removed iteratively. By limiting the pruning region, our framework can greatly reduce either the wireless transmission workload of the device or the total computation workload. A series of pruned models are generated in the training phase, from which the framework can automatically select to satisfy varying latency and accuracy requirements. Furthermore, coding for the intermediate data is added to provide extra transmission workload reduction. Our experiments show that the proposed framework can achieve up to 25.6$\times$ reduction on transmission workload, 6.01$\times$ acceleration on total computation and 4.81$\times$ reduction on end-to-end latency as compared to partitioning the original DNN model without pruning.