David Wu

Nicholas Heller, Fabian Isensee, Dasha Trofimova et al.

This paper presents the challenge report for the 2021 Kidney and Kidney Tumor Segmentation Challenge (KiTS21) held in conjunction with the 2021 international conference on Medical Image Computing and Computer Assisted Interventions (MICCAI). KiTS21 is a sequel to its first edition in 2019, and it features a variety of innovations in how the challenge was designed, in addition to a larger dataset. A novel annotation method was used to collect three separate annotations for each region of interest, and these annotations were performed in a fully transparent setting using a web-based annotation tool. Further, the KiTS21 test set was collected from an outside institution, challenging participants to develop methods that generalize well to new populations. Nonetheless, the top-performing teams achieved a significant improvement over the state of the art set in 2019, and this performance is shown to inch ever closer to human-level performance. An in-depth meta-analysis is presented describing which methods were used and how they faired on the leaderboard, as well as the characteristics of which cases generally saw good performance, and which did not. Overall KiTS21 facilitated a significant advancement in the state of the art in kidney tumor segmentation, and provides useful insights that are applicable to the field of semantic segmentation as a whole.

Hengyuan Hu, Samuel Sokota, David Wu et al. · meta-ai, oxford

Fully cooperative, partially observable multi-agent problems are ubiquitous in the real world. In this paper, we focus on a specific subclass of coordination problems in which humans are able to discover self-explaining deviations (SEDs). SEDs are actions that deviate from the common understanding of what reasonable behavior would be in normal circumstances. They are taken with the intention of causing another agent or other agents to realize, using theory of mind, that the circumstance must be abnormal. We first motivate SED with a real world example and formalize its definition. Next, we introduce a novel algorithm, improvement maximizing self-explaining deviations (IMPROVISED), to perform SEDs. Lastly, we evaluate IMPROVISED both in an illustrative toy setting and the popular benchmark setting Hanabi, where it is the first method to produce so called finesse plays, which are regarded as one of the more iconic examples of human theory of mind.

Andrew Lee, David Wu, Emily Dinan et al.

Despite many recent advancements in language modeling, state-of-the-art language models lack grounding in the real world and struggle with tasks involving complex reasoning. Meanwhile, advances in the symbolic reasoning capabilities of AI have led to systems that outperform humans in games like chess and Go (Silver et al., 2018). Chess commentary provides an interesting domain for bridging these two fields of research, as it requires reasoning over a complex board state and providing analyses in natural language. In this work we demonstrate how to combine symbolic reasoning engines with controllable language models to generate chess commentaries. We conduct experiments to demonstrate that our approach generates commentaries that are preferred by human judges over previous baselines.

David Wu, Sebastian Jaimungal

The objectives of option hedging/trading extend beyond mere protection against downside risks, with a desire to seek gains also driving agent's strategies. In this study, we showcase the potential of robust risk-aware reinforcement learning (RL) in mitigating the risks associated with path-dependent financial derivatives. We accomplish this by leveraging a policy gradient approach that optimises robust risk-aware performance criteria. We specifically apply this methodology to the hedging of barrier options, and highlight how the optimal hedging strategy undergoes distortions as the agent moves from being risk-averse to risk-seeking. As well as how the agent robustifies their strategy. We further investigate the performance of the hedge when the data generating process (DGP) varies from the training DGP, and demonstrate that the robust strategies outperform the non-robust ones.

Payal Chandak, Victoria Alkin, David Wu et al.

Medicine is inherently pluralistic. Principles such as autonomy, beneficence, nonmaleficence, and justice routinely conflict, and such ethical dilemmas often sharply divide reasonable physicians. Good clinical practice navigates these tensions in concert with each patient's values rather than imposing a single ethical stance. The ethical values that large language models bring to medical advice, however, have not been systematically examined. We present a framework for auditing value pluralism in medical AI, comprising a benchmark of clinician-verified dilemmas and an attribution method that recovers value priorities directly from decisions. The ecosystem of frontier models spans physician-level value heterogeneity, and models discuss competing values in their reasoning (Overton pluralism) before committing to a decision. However, individual model decisions are near-deterministic across repeated sampling and semantic variations, failing to reproduce the distributional pluralism of the physician panel. Across benchmark cases, these consistent decisions reflect committed, systematic value preferences. While most model priorities fall within the natural range of inter-physician variation, some significantly underweight patient autonomy. A single LLM deployed without regard for its value priorities could amplify those priorities at scale to every patient it serves. Without explicit efforts to balance ethical perspectives with one or multiple models, these tools risk replacing clinical pluralism with a deployment monoculture.

Zijie Yan, Hongxiao Bai, Xin Yao et al.

Scaling Mixture-of-Experts (MoE) training introduces systems challenges absent in dense models. Because each token activates only a subset of experts, this sparsity allows total parameters to grow much faster than per-token computation, creating coupled constraints across memory, communication, and computation. Optimizing one dimension often shifts pressure to another, demanding co-design across the full system stack. We address these challenges for MoE training through integrated optimizations spanning memory (fine-grained recomputation, offloading, etc.), communication (optimized dispatchers, overlapping, etc.), and computation (Grouped GEMM, fusions, CUDA Graphs, etc.). The framework also provides Parallel Folding for flexible multi-dimensional parallelism, low-precision training support for FP8 and NVFP4, and efficient long-context training. On NVIDIA GB300 and GB200, it achieves 1,233/1,048 TFLOPS/GPU for DeepSeek-V3-685B and 974/919 TFLOPS/GPU for Qwen3-235B. As a performant, scalable, and production-ready open-source solution, it has been used across academia and industry for training MoE models ranging from billions to trillions of parameters on clusters scaling up to thousands of GPUs. This report explains how these techniques work, their trade-offs, and their interactions at the systems level, providing practical guidance for scaling MoE models with Megatron Core.

Dennis Liu, Zijie Yan, Xin Yao et al. · nvidia

Mixture of Experts (MoE) models enhance neural network scalability by dynamically selecting relevant experts per input token, enabling larger model sizes while maintaining manageable computation costs. However, efficient training of large-scale MoE models across thousands of GPUs presents significant challenges due to limitations in existing parallelism strategies. We introduce an end-to-end training framework for large-scale MoE models that utilizes five-dimensional hybrid parallelism: Tensor Parallelism, Expert Parallelism, Context Parallelism, Data Parallelism, and Pipeline Parallelism. Central to our approach is MoE Parallel Folding, a novel strategy that decouples the parallelization of attention and MoE layers in Transformer models, allowing each layer type to adopt optimal parallel configurations. Additionally, we develop a flexible token-level dispatcher that supports both token-dropping and token-dropless MoE training across all five dimensions of parallelism. This dispatcher accommodates dynamic tensor shapes and coordinates different parallelism schemes for Attention and MoE layers, facilitating complex parallelism implementations. Our experiments demonstrate significant improvements in training efficiency and scalability. We achieve up to 49.3% Model Flops Utilization (MFU) for the Mixtral 8x22B model and 39.0% MFU for the Qwen2-57B-A14B model on H100 GPUs, outperforming existing methods. The framework scales efficiently up to 1,024 GPUs and maintains high performance with sequence lengths up to 128K tokens, validating its effectiveness for large-scale MoE model training. The code is available in Megatron-Core.

David Wu, Sanjiban Choudhury

Aligning large language models (LLMs) to human preferences is challenging in domains where preference data is unavailable. We address the problem of learning reward models for such target domains by leveraging feedback collected from simpler source domains, where human preferences are easier to obtain. Our key insight is that, while domains may differ significantly, human preferences convey \emph{domain-agnostic} concepts that can be effectively captured by a reward model. We propose \method, a framework that trains domain-invariant reward models by optimizing a dual loss: a domain loss that minimizes the divergence between source and target distribution, and a source loss that optimizes preferences on the source domain. We show \method is a general approach that we evaluate and analyze across 4 distinct settings: (1) Cross-lingual transfer (accuracy: $0.621 \rightarrow 0.661$), (2) Clean-to-noisy (accuracy: $0.671 \rightarrow 0.703$), (3) Few-shot-to-full transfer (accuracy: $0.845 \rightarrow 0.920$), and (4) Simple-to-complex tasks transfer (correlation: $0.508 \rightarrow 0.556$). Our code, models and data are available at \url{https://github.com/portal-cornell/dial}.

John Wu, David Wu, Jimeng Sun

Predicting high-dimensional or extreme multilabels, such as in medical coding, requires both accuracy and interpretability. Existing works often rely on local interpretability methods, failing to provide comprehensive explanations of the overall mechanism behind each label prediction within a multilabel set. We propose a mechanistic interpretability module called DIctionary Label Attention (\method) that disentangles uninterpretable dense embeddings into a sparse embedding space, where each nonzero element (a dictionary feature) represents a globally learned medical concept. Through human evaluations, we show that our sparse embeddings are more human understandable than its dense counterparts by at least 50 percent. Our automated dictionary feature identification pipeline, leveraging large language models (LLMs), uncovers thousands of learned medical concepts by examining and summarizing the highest activating tokens for each dictionary feature. We represent the relationships between dictionary features and medical codes through a sparse interpretable matrix, enhancing the mechanistic and global understanding of the model's predictions while maintaining competitive performance and scalability without extensive human annotation.

John Wu, David Wu, Jimeng Sun

Medical coding, the translation of unstructured clinical text into standardized medical codes, is a crucial but time-consuming healthcare practice. Though large language models (LLM) could automate the coding process and improve the efficiency of such tasks, interpretability remains paramount for maintaining patient trust. Current efforts in interpretability of medical coding applications rely heavily on label attention mechanisms, which often leads to the highlighting of extraneous tokens irrelevant to the ICD code. To facilitate accurate interpretability in medical language models, this paper leverages dictionary learning that can efficiently extract sparsely activated representations from dense language model embeddings in superposition. Compared with common label attention mechanisms, our model goes beyond token-level representations by building an interpretable dictionary which enhances the mechanistic-based explanations for each ICD code prediction, even when the highlighted tokens are medically irrelevant. We show that dictionary features can steer model behavior, elucidate the hidden meanings of upwards of 90% of medically irrelevant tokens, and are human interpretable.

David Wu, Fateme Nateghi Haredasht, Saloni Kumar Maharaj et al.

Large language models (LLMs) are routinely used by physicians and patients for medical advice, yet their clinical safety profiles remain poorly characterized. We present NOHARM (Numerous Options Harm Assessment for Risk in Medicine), a benchmark using 100 real primary-care-to-specialist consultation cases to measure harm frequency and severity from LLM-generated medical recommendations. NOHARM covers 10 specialties, with 12,747 expert annotations for 4,249 clinical management options. Across 31 LLMs, severe harm occurs in up to 22.2% (95% CI 21.6-22.8%) of cases, with harms of omission accounting for 76.6% (95% CI 76.4-76.8%) of errors. Safety performance is only moderately correlated (r = 0.61-0.64) with existing AI and medical knowledge benchmarks. The best models outperform generalist physicians on safety (mean difference 9.7%, 95% CI 7.0-12.5%), and a diverse multi-agent approach reduces harm compared to solo models (mean difference 8.0%, 95% CI 4.0-12.1%). Therefore, despite strong performance on existing evaluations, widely used AI models can produce severely harmful medical advice at nontrivial rates, underscoring clinical safety as a distinct performance dimension necessitating explicit measurement.

Liam G. McCoy, Fateme Nateghi Haredasht, Kanav Chopra et al.

This study evaluates the capacity of large language models (LLMs) to generate structured clinical consultation templates for electronic consultation. Using 145 expert-crafted templates developed and routinely used by Stanford's eConsult team, we assess frontier models -- including o3, GPT-4o, Kimi K2, Claude 4 Sonnet, Llama 3 70B, and Gemini 2.5 Pro -- for their ability to produce clinically coherent, concise, and prioritized clinical question schemas. Through a multi-agent pipeline combining prompt optimization, semantic autograding, and prioritization analysis, we show that while models like o3 achieve high comprehensiveness (up to 92.2\%), they consistently generate excessively long templates and fail to correctly prioritize the most clinically important questions under length constraints. Performance varies across specialties, with significant degradation in narrative-driven fields such as psychiatry and pain medicine. Our findings demonstrate that LLMs can enhance structured clinical information exchange between physicians, while highlighting the need for more robust evaluation methods that capture a model's ability to prioritize clinically salient information within the time constraints of real-world physician communication.

David Wu, Gokul Swamy, J. Andrew Bagnell et al.

Inverse Reinforcement Learning (IRL) is a powerful framework for learning complex behaviors from expert demonstrations. However, it traditionally requires repeatedly solving a computationally expensive reinforcement learning (RL) problem in its inner loop. It is desirable to reduce the exploration burden by leveraging expert demonstrations in the inner-loop RL. As an example, recent work resets the learner to expert states in order to inform the learner of high-reward expert states. However, such an approach is infeasible in the real world. In this work, we consider an alternative approach to speeding up the RL subroutine in IRL: \emph{pessimism}, i.e., staying close to the expert's data distribution, instantiated via the use of offline RL algorithms. We formalize a connection between offline RL and IRL, enabling us to use an arbitrary offline RL algorithm to improve the sample efficiency of IRL. We validate our theory experimentally by demonstrating a strong correlation between the efficacy of an offline RL algorithm and how well it works as part of an IRL procedure. By using a strong offline RL algorithm as part of an IRL procedure, we are able to find policies that match expert performance significantly more efficiently than the prior art.

David Wu, Sanjiban Choudhury

Existing inverse reinforcement learning methods (e.g. MaxEntIRL, $f$-IRL) search over candidate reward functions and solve a reinforcement learning problem in the inner loop. This creates a rather strange inversion where a harder problem, reinforcement learning, is in the inner loop of a presumably easier problem, imitation learning. In this work, we show that better utilization of expert demonstrations can reduce the need for hard exploration in the inner RL loop, hence accelerating learning. Specifically, we propose two simple recipes: (1) placing expert transitions into the replay buffer of the inner RL algorithm (e.g. Soft-Actor Critic) which directly informs the learner about high reward states instead of forcing the learner to discover them through extensive exploration, and (2) using expert actions in Q value bootstrapping in order to improve the target Q value estimates and more accurately describe high value expert states. Our methods show significant gains over a MaxEntIRL baseline on the benchmark MuJoCo suite of tasks, speeding up recovery to 70\% of deterministic expert performance by 2.13x on HalfCheetah-v2, 2.6x on Ant-v2, 18x on Hopper-v2, and 3.36x on Walker2d-v2.

David Wu, Yunnan Wu

The 2021 Waymo Interaction Prediction Challenge introduced a problem of predicting the future trajectories and confidences of two interacting agents jointly. We developed a solution that takes an anchored marginal motion prediction model with rasterization and augments it to model agent interaction. We do this by predicting the joint confidences using a rasterized image that highlights the ego agent and the interacting agent. Our solution operates on the cartesian product space of the anchors; hence the $"^2"$ in $AIR^2$. Our model achieved the highest mAP (the primary metric) on the leaderboard.

David Wu, Yunnan Wu

Self-supervised representation learning is a fundamental problem in computer vision with many useful applications (e.g., image search, instance level recognition, copy detection). In this paper we present a new contrastive self-supervised representation learning algorithm in the context of Copy Detection in the 2021 Image Similarity Challenge hosted by Facebook AI Research. Previous work in contrastive self-supervised learning has identified the importance of being able to optimize representations while ``pushing'' against a large number of negative examples. Representative previous solutions either use large batches enabled by modern distributed training systems or maintain queues or memory banks holding recently evaluated representations while relaxing some consistency properties. We approach this problem from a new angle: We directly learn a query model and a key model jointly and push representations against a very large number (e.g., 1 million) of negative representations in each SGD step. We achieve this by freezing the backbone on one side and by alternating between a Q-optimization step and a K-optimization step. During the competition timeframe, our algorithms achieved a micro-AP score of 0.3401 on the Phase 1 leaderboard, significantly improving over the baseline $μ$AP of 0.1556. On the final Phase 2 leaderboard, our model scored 0.1919, while the baseline scored 0.0526. Continued training yielded further improvement. We conducted an empirical study to compare the proposed approach with a SimCLR style strategy where the negative examples are taken from the batch only. We found that our method ($μ$AP of 0.3403) significantly outperforms this SimCLR-style baseline ($μ$AP of 0.2001).

Anton Bakhtin, David Wu, Adam Lerer et al.

Prior AI successes in complex games have largely focused on settings with at most hundreds of actions at each decision point. In contrast, Diplomacy is a game with more than 10^20 possible actions per turn. Previous attempts to address games with large branching factors, such as Diplomacy, StarCraft, and Dota, used human data to bootstrap the policy or used handcrafted reward shaping. In this paper, we describe an algorithm for action exploration and equilibrium approximation in games with combinatorial action spaces. This algorithm simultaneously performs value iteration while learning a policy proposal network. A double oracle step is used to explore additional actions to add to the policy proposals. At each state, the target state value and policy for the model training are computed via an equilibrium search procedure. Using this algorithm, we train an agent, DORA, completely from scratch for a popular two-player variant of Diplomacy and show that it achieves superhuman performance. Additionally, we extend our methods to full-scale no-press Diplomacy and for the first time train an agent from scratch with no human data. We present evidence that this agent plays a strategy that is incompatible with human-data bootstrapped agents. This presents the first strong evidence of multiple equilibria in Diplomacy and suggests that self play alone may be insufficient for achieving superhuman performance in Diplomacy.

Hengyuan Hu, Adam Lerer, Brandon Cui et al.

The standard problem setting in Dec-POMDPs is self-play, where the goal is to find a set of policies that play optimally together. Policies learned through self-play may adopt arbitrary conventions and implicitly rely on multi-step reasoning based on fragile assumptions about other agents' actions and thus fail when paired with humans or independently trained agents at test time. To address this, we present off-belief learning (OBL). At each timestep OBL agents follow a policy $π_1$ that is optimized assuming past actions were taken by a given, fixed policy ($π_0$), but assuming that future actions will be taken by $π_1$. When $π_0$ is uniform random, OBL converges to an optimal policy that does not rely on inferences based on other agents' behavior (an optimal grounded policy). OBL can be iterated in a hierarchy, where the optimal policy from one level becomes the input to the next, thereby introducing multi-level cognitive reasoning in a controlled manner. Unlike existing approaches, which may converge to any equilibrium policy, OBL converges to a unique policy, making it suitable for zero-shot coordination (ZSC). OBL can be scaled to high-dimensional settings with a fictitious transition mechanism and shows strong performance in both a toy-setting and the benchmark human-AI & ZSC problem Hanabi.

David Wu, David R. Palmer, Daryl R. Deford

We present a convex cone program to infer the latent probability matrix of a random dot product graph (RDPG). The optimization problem maximizes the Bernoulli maximum likelihood function with an added nuclear norm regularization term. The dual problem has a particularly nice form, related to the well-known semidefinite program relaxation of the MaxCut problem. Using the primal-dual optimality conditions, we bound the entries and rank of the primal and dual solutions. Furthermore, we bound the optimal objective value and prove asymptotic consistency of the probability estimates of a slightly modified model under mild technical assumptions. Our experiments on synthetic RDPGs not only recover natural clusters, but also reveal the underlying low-dimensional geometry of the original data. We also demonstrate that the method recovers latent structure in the Karate Club Graph and synthetic U.S. Senate vote graphs and is scalable to graphs with up to a few hundred nodes.