Tony Wang

Stephen Casper, Xander Davies, Claudia Shi et al. · berkeley, eth-zurich

Reinforcement learning from human feedback (RLHF) is a technique for training AI systems to align with human goals. RLHF has emerged as the central method used to finetune state-of-the-art large language models (LLMs). Despite this popularity, there has been relatively little public work systematizing its flaws. In this paper, we (1) survey open problems and fundamental limitations of RLHF and related methods; (2) overview techniques to understand, improve, and complement RLHF in practice; and (3) propose auditing and disclosure standards to improve societal oversight of RLHF systems. Our work emphasizes the limitations of RLHF and highlights the importance of a multi-faceted approach to the development of safer AI systems.

Aarti Basant, Abhijit Khairnar, Abhijit Paithankar et al. · nvidia

We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

Xiangyu Wang, Yawen He, Shivendra Pratap Singh et al.

Large-scale recommendation systems operate across diverse domains, yet they face the challenges of data sparsity and noisy implicit feedback. Traditional approaches mitigate this via model-specific knowledge distillation from source domains to a target domain. Inspired by the transformative success of synthetic data generation in large language models (LLMs), we introduce Synthetic Cross-domain Augmentation and Learning for Recommendation (SCALR), a framework that generates synthetic user-item interaction events for a target recommendation domain by leveraging observed events from a source domain. SCALR decomposes cross-domain learning into two modular stages. First, it translates observed user events in source domains by framing event generation as estimating the likelihood that a user would interact with a target-domain item, conditioned on their observed interactions in a source domain. Second, downstream models train on these synthetic events as cross-domain learning objectives, where the synthetic events augment the target domain's training data in a model-agnostic manner. Our approach yields statistically significant improvements in online A/B tests on an industrial recommendation platform. To the best of our knowledge, this is among the first works to explicitly frame cross-domain event transfer as synthetic data generation for recommendation systems.

Bingjun Luo, Tony Wang, Chaoqi Chen et al.

Multimodal Large Language Models (MLLMs) face significant computational overhead when processing long videos due to the massive number of visual tokens required. To improve efficiency, existing methods primarily reduce redundancy by pruning or merging tokens based on importance or similarity. However, these approaches largely overlook a critical dimension of video content, i.e., changes and turning points, and they lack a collaborative model for spatio-temporal relationships. To address this, we propose a new perspective: similarity is for identifying redundancy, while difference is for capturing key events. Based on this, we designed a training-free framework named ST-SimDiff. We first construct a spatio-temporal graph from the visual tokens to uniformly model their complex associations. Subsequently, we employ a parallel dual-selection strategy: 1) similarity-based selection uses community detection to retain representative tokens, compressing static information; 2) temporal difference-based selection precisely locates content-changing points to preserve tokens that capture key dynamic shifts. This allows it to preserve both static and dynamic content with a minimal number of tokens. Extensive experiments show our method significantly outperforms state-of-the-art approaches while substantially reducing computational costs. Our code is available in https://github.com/bingjunluo/ST-SimDiff.

Bingjun Luo, Tony Wang, Hanqi Chen et al.

Recent advances in Multimodal Large Language Models (MLLMs) have significantly advanced video understanding tasks, yet challenges remain in efficiently compressing visual tokens while preserving spatiotemporal interactions. Existing methods, such as LLaVA family, utilize simplistic pooling or interpolation techniques that overlook the intricate dynamics of visual tokens. To bridge this gap, we propose ST-GridPool, a novel training-free visual token enhancement method designed specifically for Video LLMs. Our approach integrates Pyramid Temporal Gridding (PTG), which captures multi-grained spatiotemporal interactions through hierarchical temporal gridding, and Norm-based Spatial Pooling (NSP), which preserves high-information visual regions by leveraging the correlation between token norms and semantic richness. Extensive experiments on various benchmarks demonstrate that ST-GridPool consistently enhances performance of Video LLMs without requiring costly retraining. Our method offers an efficient and plug-and-play solution for improving visual token representations. Our code is available in https://github.com/bingjunluo/ST-GridPool.

Haijun Xia, Tony Wang, Aditya Gunturu et al.

Despite the advances and ubiquity of digital communication media such as videoconferencing and virtual reality, they remain oblivious to the rich intentions expressed by users. Beyond transmitting audio, videos, and messages, we envision digital communication media as proactive facilitators that can provide unobtrusive assistance to enhance communication and collaboration. Informed by the results of a formative study, we propose three key design concepts to explore the systematic integration of intelligence into communication and collaboration, including the panel substrate, language-based intent recognition, and lightweight interaction techniques. We developed CrossTalk, a videoconferencing system that instantiates these concepts, which was found to enable a more fluid and flexible communication and collaboration experience.

Ruiqi Deng, Geoffrey Martin, Tony Wang et al.

Clinical practice guidelines (CPGs) provide evidence-based recommendations for patient care; however, integrating them into Artificial Intelligence (AI) remains challenging. Previous approaches, such as rule-based systems, face significant limitations, including poor interpretability, inconsistent adherence to guidelines, and narrow domain applicability. To address this, we develop and validate CPGPrompt, an auto-prompting system that converts narrative clinical guidelines into large language models (LLMs). Our framework translates CPGs into structured decision trees and utilizes an LLM to dynamically navigate them for patient case evaluation. Synthetic vignettes were generated across three domains (headache, lower back pain, and prostate cancer) and distributed into four categories to test different decision scenarios. System performance was assessed on both binary specialty-referral decisions and fine-grained pathway-classification tasks. The binary specialty referral classification achieved consistently strong performance across all domains (F1: 0.85-1.00), with high recall (1.00 $\pm$ 0.00). In contrast, multi-class pathway assignment showed reduced performance, with domain-specific variations: headache (F1: 0.47), lower back pain (F1: 0.72), and prostate cancer (F1: 0.77). Domain-specific performance differences reflected the structure of each guideline. The headache guideline highlighted challenges with negation handling. The lower back pain guideline required temporal reasoning. In contrast, prostate cancer pathways benefited from quantifiable laboratory tests, resulting in more reliable decision-making.

Miriam Senne, Benjamin D. Killeen, Tony Wang et al.

Body composition assessment (BCA) provides detailed information about the distribution of different tissue types in the body, enabling more precise characterization of individuals than BMI or weight alone. Consistent and frequent BCA would be valuable for personalized medicine, but the gold standard methods for BCA, such as CT and MRI, are only practical for opportunistic monitoring of patients with clinical indications for imaging and are not suitable for routine use in the general population. Here, we consider an imaging modality which is not currently used in medical applications: millimeter wave (mmWave) radar. Commonly used in security settings, mmWave scans enable fast, non-intrusive, and privacy-preserving reconstruction of full body shape without the need to remove clothing. To demonstrate the feasibility of fast and convenient BCA from mmWave scans, we present a method for BCA value regression using a multi-task learning strategy that leverages synthetic mmWave-like point clouds derived from clinical imaging and parametric human models. We evaluate the model on a pilot cohort of real mmWave scans with bioimpedance-derived body fat measurements, supporting the feasibility of estimating VAT and body fat percentage (BFP) from mmWave data acquired through clothing in a standing posture. We find that the model can predict VAT and BFP with a mean absolute error of 1.0 L and 3.2\%, respectively, demonstrating the potential of mmWave scanning for routine BCA in a wide range of settings.

Xinjie Zhao, Moritz Blum, Rui Yang et al.

Large Language Models~(LLMs) have demonstrated capabilities across various applications but face challenges such as hallucination, limited reasoning abilities, and factual inconsistencies, especially when tackling complex, domain-specific tasks like question answering~(QA). While Knowledge Graphs~(KGs) have been shown to help mitigate these issues, research on the integration of LLMs with background KGs remains limited. In particular, user accessibility and the flexibility of the underlying KG have not been thoroughly explored. We introduce AGENTiGraph (Adaptive Generative ENgine for Task-based Interaction and Graphical Representation), a platform for knowledge management through natural language interaction. It integrates knowledge extraction, integration, and real-time visualization. AGENTiGraph employs a multi-agent architecture to dynamically interpret user intents, manage tasks, and integrate new knowledge, ensuring adaptability to evolving user requirements and data contexts. Our approach demonstrates superior performance in knowledge graph interactions, particularly for complex domain-specific tasks. Experimental results on a dataset of 3,500 test cases show AGENTiGraph significantly outperforms state-of-the-art zero-shot baselines, achieving 95.12\% accuracy in task classification and 90.45\% success rate in task execution. User studies corroborate its effectiveness in real-world scenarios. To showcase versatility, we extended AGENTiGraph to legislation and healthcare domains, constructing specialized KGs capable of answering complex queries in legal and medical contexts.

Tao Long, Sitong Wang, Émilie Fabre et al.

UIST researchers develop tools to address user challenges. However, user interactions with AI evolve over time through learning, adaptation, and repurposing, making one time evaluations insufficient. Capturing these dynamics requires longer-term studies, but challenges in deployment, evaluation design, and data collection have made such longitudinal research difficult to implement. Our workshop aims to tackle these challenges and prepare researchers with practical strategies for longitudinal studies. The workshop includes a keynote, panel discussions, and interactive breakout groups for discussion and hands-on protocol design and tool prototyping sessions. We seek to foster a community around longitudinal system research and promote it as a more embraced method for designing, building, and evaluating UIST tools.

Blerim Abdullai, Tony Wang, Xinyuan Qiao et al.

GNSS is unreliable, inaccurate, and insufficient in many real-time autonomous field applications. In this work, we present a GNSS-free global localization solution that contains a method of registering imaging radar on the ground with overhead RGB imagery, with joint optimization of relative poses from odometry and global poses from our overhead registration. Previous works have used various combinations of ground sensors and overhead imagery, and different feature extraction and matching methods. These include various handcrafted and deep-learning-based methods for extracting features from overhead imagery. Our work presents insights on extracting essential features from RGB overhead images for effective global localization against overhead imagery using only ground radar and a single georeferenced initial guess. We motivate our method by evaluating it on datasets in diverse geographic conditions and robotic platforms, including on an Unmanned Surface Vessel (USV) as well as urban and suburban driving datasets.

Simon Alford, Anshula Gandhi, Akshay Rangamani et al.

One of the challenges facing artificial intelligence research today is designing systems capable of utilizing systematic reasoning to generalize to new tasks. The Abstraction and Reasoning Corpus (ARC) measures such a capability through a set of visual reasoning tasks. In this paper we report incremental progress on ARC and lay the foundations for two approaches to abstraction and reasoning not based in brute-force search. We first apply an existing program synthesis system called DreamCoder to create symbolic abstractions out of tasks solved so far, and show how it enables solving of progressively more challenging ARC tasks. Second, we design a reasoning algorithm motivated by the way humans approach ARC. Our algorithm constructs a search graph and reasons over this graph structure to discover task solutions. More specifically, we extend existing execution-guided program synthesis approaches with deductive reasoning based on function inverse semantics to enable a neural-guided bidirectional search algorithm. We demonstrate the effectiveness of the algorithm on three domains: ARC, 24-Game tasks, and a 'double-and-add' arithmetic puzzle.

Tony Wang, DeaGyu Kim, Yifan Shi et al.

The recent advancements in nanoscale 3D printing and microfabrication techniques have reinvigorated research on microrobots. However, precise motion control of the microrobots on biological environments using compact actuation setups remains challenging to date. This work presents a novel control mechanism and contact design that enables bidirectional steering via biasing the neutral position of the microrobot. Equipped with rockers to contact the substrate, the microrobot, hence microrocker bot, is capable of well-controlled forward and backward movement on flat and non-flat biological surfaces. The 100um by 113um by 36um robots were 3D printed via two-photon lithography and subsequently deposited with nickel thin films. Under a relatively small static magnetic field, the microrocker bot tilts either forward or backward to align the thin film magnetization direction with the magnetic field lines. When combined with an oscillating magnetic field, the robot undergoes stick-slip motion in the predisposed direction, dictated by the neutral position tilt. The microrocker bots are further equipped with sharp mechanical tips that can be selectively engaged. When the frequency and offset of the actuation sawtooth waveform are optimized, the robot travels up to 100um/s (1 body length per second) forward and backward showing very linear trajectories. Finally, to prove the functionality of the microrocker bots in direct contact with biological surfaces, we demonstrate the robot's ability to traverse forward and backward on the surface of a Dracaena Fragrans leaf, and upend/engage on its mechanical tip.

Tom Velez, Tony Wang, Ioannis Koutroulis et al.

Background: While machine learning (ML) models are rapidly emerging as promising screening tools in critical care medicine, the identification of homogeneous subphenotypes within populations with heterogeneous conditions such as pediatric sepsis may facilitate attainment of high-predictive performance of these prognostic algorithms. This study is aimed to identify subphenotypes of pediatric sepsis and demonstrate the potential value of partitioned data/subtyping-based training. Methods: This was a retrospective study of clinical data extracted from medical records of 6,446 pediatric patients that were admitted at a major hospital system in the DC area. Vitals and labs associated with patients meeting the diagnostic criteria for sepsis were used to perform latent profile analysis. Modern ML algorithms were used to explore the predictive performance benefits of reduced training data heterogeneity via label profiling. Results: In total 134 (2.1%) patients met the diagnostic criteria for sepsis in this cohort and latent profile analysis identified four profiles/subphenotypes of pediatric sepsis. Profiles 1 and 3 had the lowest mortality and included pediatric patients from different age groups. Profile 2 were characterized by respiratory dysfunction; profile 4 by neurological dysfunction and highest mortality rate (22.2%). Machine learning experiments comparing the predictive performance of models derived without training data profiling against profile targeted models suggest statistically significant improved performance of prediction can be obtained. For example, area under ROC curve (AUC) obtained to predict profile 4 with 24-hour data (AUC = .998, p < .0001) compared favorably with the AUC obtained from the model considering all profiles as a single homogeneous group (AUC = .918) with 24-hour data.

Tony Wang, Tim Tschampel, Emilia Apostolova et al.

In this work, we utilize Machine Learning for early recognition of patients at high risk of acute respiratory distress syndrome (ARDS), which is critical for successful prevention strategies for this devastating syndrome. The difficulty in early ARDS recognition stems from its complex and heterogenous nature. In this study, we integrate knowledge of the heterogeneity of ARDS patients into predictive model building. Using MIMIC-III data, we first apply latent class analysis (LCA) to identify homogeneous sub-groups in the ARDS population, and then build predictive models on the partitioned data. The results indicate that significantly improved performances of prediction can be obtained for two of the three identified sub-phenotypes of ARDS. Experiments suggests that identifying sub-phenotypes is beneficial for building predictive model for ARDS.

Tony Wang, Tom Velez, Emilia Apostolova et al.

Although timely sepsis diagnosis and prompt interventions in Intensive Care Unit (ICU) patients are associated with reduced mortality, early clinical recognition is frequently impeded by non-specific signs of infection and failure to detect signs of sepsis-induced organ dysfunction in a constellation of dynamically changing physiological data. The goal of this work is to identify patient at risk of life-threatening sepsis utilizing a data-centered and machine learning-driven approach. We derive a mortality risk predictive dynamic Bayesian network (DBN) guided by a customized sepsis knowledgebase and compare the predictive accuracy of the derived DBN with the Sepsis-related Organ Failure Assessment (SOFA) score, the Quick SOFA (qSOFA) score, the Simplified Acute Physiological Score (SAPS-II) and the Modified Early Warning Score (MEWS) tools. A customized sepsis ontology was used to derive the DBN node structure and semantically characterize temporal features derived from both structured physiological data and unstructured clinical notes. We assessed the performance in predicting mortality risk of the DBN predictive model and compared performance to other models using Receiver Operating Characteristic (ROC) curves, area under curve (AUROC), calibration curves, and risk distributions. The derived dataset consists of 24,506 ICU stays from 19,623 patients with evidence of suspected infection, with 2,829 patients deceased at discharge. The DBN AUROC was found to be 0.91, which outperformed the SOFA (0.843), qSOFA (0.66), MEWS (0.73), and SAPS-II (0.77) scoring tools. Continuous Net Reclassification Index and Integrated Discrimination Improvement analysis supported the superiority DBN. Compared with conventional rule-based risk scoring tools, the sepsis knowledgebase-driven DBN algorithm offers improved performance for predicting mortality of infected patients in ICUs.