Sean Wu

Sean Wu, Michael Koo, Lesley Blum et al.

In recent years, there have been significant breakthroughs in the field of natural language processing, particularly with the development of large language models (LLMs). These LLMs have showcased remarkable capabilities on various benchmarks. In the healthcare field, the exact role LLMs and other future AI models will play remains unclear. There is a potential for these models in the future to be used as part of adaptive physician training, medical co-pilot applications, and digital patient interaction scenarios. The ability of AI models to participate in medical training and patient care will depend in part on their mastery of the knowledge content of specific medical fields. This study investigated the medical knowledge capability of LLMs, specifically in the context of internal medicine subspecialty multiple-choice test-taking ability. We compared the performance of several open-source LLMs (Koala 7B, Falcon 7B, Stable-Vicuna 13B, and Orca Mini 13B), to GPT-4 and Claude 2 on multiple-choice questions in the field of Nephrology. Nephrology was chosen as an example of a particularly conceptually complex subspecialty field within internal medicine. The study was conducted to evaluate the ability of LLM models to provide correct answers to nephSAP (Nephrology Self-Assessment Program) multiple-choice questions. The overall success of open-sourced LLMs in answering the 858 nephSAP multiple-choice questions correctly was 17.1% - 25.5%. In contrast, Claude 2 answered 54.4% of the questions correctly, whereas GPT-4 achieved a score of 73.3%. We show that current widely used open-sourced LLMs do poorly in their ability for zero-shot reasoning when compared to GPT-4 and Claude 2. The findings of this study potentially have significant implications for the future of subspecialty medical training and patient care.

Sean Wu, Michael Koo, Li Yo Kao et al.

Open-source LLMs have shown great potential as fine-tuned chatbots, and demonstrate robust abilities in reasoning and surpass many existing benchmarks. Retrieval-Augmented Generation (RAG) is a technique for improving the performance of LLMs on tasks that the models weren't explicitly trained on, by leveraging external knowledge databases. Numerous studies have demonstrated the effectiveness of RAG to more successfully accomplish downstream tasks when using vector datasets that consist of relevant background information. It has been implicitly assumed by those in the field that if adversarial background information is utilized in this context, that the success of using a RAG-based approach would be nonexistent or even negatively impact the results. To address this assumption, we tested several open-source LLMs on the ability of RAG to improve their success in answering multiple-choice questions (MCQ) in the medical subspecialty field of Nephrology. Unlike previous studies, we examined the effect of RAG in utilizing both relevant and adversarial background databases. We set up several open-source LLMs, including Llama 3, Phi-3, Mixtral 8x7b, Zephyr$β$, and Gemma 7B Instruct, in a zero-shot RAG pipeline. As adversarial sources of information, text from the Bible and a Random Words generated database were used for comparison. Our data show that most of the open-source LLMs improve their multiple-choice test-taking success as expected when incorporating relevant information vector databases. Surprisingly however, adversarial Bible text significantly improved the success of many LLMs and even random word text improved test taking ability of some of the models. In summary, our results demonstrate for the first time the countertintuitive ability of adversarial information datasets to improve the RAG-based LLM success.

Jinge Wu, Hongjian Zhou, Mingde Zeng et al.

Building a deep research agent today is an exercise in glue code: the same backbone evaluated on the same benchmark can report different accuracies in different papers because harness and tool registry all differ, and integrating a new foundation model into a comparable evaluation surface costs weeks of model-specific engineering. We call this the per-paper engineering tax and release BioMedArena, an open-source toolkit that not only alleviates it but also provides an arena for fair comparison of different foundation models when evaluating them as deep-research agents. BioMedArena decouples six layers of biomedical agent evaluation -- benchmark loading, tool exposure, tool selection, execution mode, context management, and scoring -- and exposes 147 biomedical benchmarks and 75 biomedical tools across 9 functional families. Adding a new model, benchmark, or tool reduces to registering a few-line provider adapter. We further provide 6 agent harnesses with 6 context-management strategies, which provide 12 backbones with competitive research capabilities and significantly improved performance, achieving state-of-the-art (SOTA) results on 8 representative biomedical benchmarks, with an average lift of +15.03 percentage points over prior SOTA. The toolkit, configurations, and per-task traces are available at https://github.com/AI-in-Health/BioMedArena

Sean Wu, Pan Lu, Yupeng Chen et al.

Artificial intelligence (AI) is increasingly embedded in scientific discovery, yet whether it can anticipate scientific progress remains unclear. To study this question, we introduce a temporally grounded evaluation framework for forecasting scientific progress under controlled knowledge constraints. We present CUSP (Cutoff-conditioned Unseen Scientific Progress), a multi-disciplinary and event-level benchmark that evaluates scientific forecasting in AI systems through feasibility assessment, mechanistic reasoning, generative solution design, and temporal prediction. Across 4,760 scientific events, we observe systematic and domain-dependent limitations in current frontier models. While models can identify plausible research directions from competing candidates, they fail to reliably predict whether scientific advances will be realized and systematically misestimate when they will occur. Performance is highly heterogeneous across domains, with the timing of AI progress more predictable than advances in biology, chemistry, and physics. Performance is largely insensitive to whether events occur before or after the training cutoff, suggesting these limitations cannot be explained solely by knowledge exposure in training data. Under controlled information access, additional pre-cutoff knowledge improves performance but does not close the gap to full-information settings, which becomes more pronounced for high-citation advances. Models also exhibit systematic overconfidence and strong response biases, indicating unreliable uncertainty estimation. Taken together, current AI systems fall short as predictive tools for scientific progress. Access to prior knowledge does not translate into reliable forecasting, and performance benefits more from post-event information than from forward-looking prediction.

William Peng, Josheev Rai, Kevin Tseng et al.

Multi-stream transformer architectures have recently been proposed as a promising direction for managing representation collapse and the vanishing gradient problem for residual connections, yet their internal mechanisms remain unexplored. In particular, the recently introduced Manifold-Constrained Hyper-Connections (mHC) architecture posits multiple residual streams with constrained interaction, but lacks in-depth mechanistic analysis. We present the first open-source mHC language model (https://huggingface.co/wgpeng/mhc-780m) and analyze the multiple-stream architecture with a suite of representation-level metrics and causal interventions to probe how parallel streams encode and utilize information. Specifically, we introduce a systematic stream ablation-and-rescue framework that enables direct causal comparison of residual streams during inference. Through targeted pairwise interventions and controlled recovery experiments, we distinguish functional redundancy from asymmetric utilization and reveal how information is distributed across streams beyond what is observable from representational similarity alone.

Sean Wu, Michael Koo, Fabien Scalzo et al.

Large language models (LLMs) have demonstrated increasingly sophisticated performance in medical and other fields of knowledge. Traditional methods of creating specialist LLMs require extensive fine-tuning and training of models on large datasets. Recently, prompt engineering, instead of fine-tuning, has shown potential to boost the performance of general foundation models. However, prompting methods such as chain-of-thought (CoT) may not be suitable for all subspecialty, and k-shot approaches may introduce irrelevant tokens into the context space. We present AutoMedPrompt, which explores the use of textual gradients to elicit medically relevant reasoning through system prompt optimization. AutoMedPrompt leverages TextGrad's automatic differentiation via text to improve the ability of general foundation LLMs. We evaluated AutoMedPrompt on Llama 3, an open-source LLM, using several QA benchmarks, including MedQA, PubMedQA, and the nephrology subspecialty-specific NephSAP. Our results show that prompting with textual gradients outperforms previous methods on open-source LLMs and surpasses proprietary models such as GPT-4, Claude 3 Opus, and Med-PaLM 2. AutoMedPrompt sets a new state-of-the-art (SOTA) performance on PubMedQA with an accuracy of 82.6$\%$, while also outperforming previous prompting strategies on open-sourced models for MedQA (77.7$\%$) and NephSAP (63.8$\%$).

Edward Phillips, Fredrik K. Gustafsson, Sean Wu et al.

Selective prediction systems can mitigate harms resulting from language model hallucinations by abstaining from answering in high-risk cases. Uncertainty quantification techniques are often employed to identify such cases, but are rarely evaluated in the context of the wider selective prediction policy and its ability to operate at low target error rates. We identify a model-dependent failure mode of entropy-based uncertainty methods that leads to unreliable abstention behaviour, and address it by combining entropy scores with a correctness probe signal. We find that across three QA benchmarks (TriviaQA, BioASQ, MedicalQA) and four model families, the combined score generally improves both the risk--coverage trade-off and calibration performance relative to entropy-only baselines. Our results highlight the importance of deployment-facing evaluation of uncertainty methods, using metrics that directly reflect whether a system can be trusted to operate at a stated risk level.

Sean Wu, Shamik Basu, Tim Broedermann et al.

We tackle the ill-posed inverse rendering problem in 3D reconstruction with a Neural Radiance Field (NeRF) approach informed by Physics-Based Rendering (PBR) theory, named PBR-NeRF. Our method addresses a key limitation in most NeRF and 3D Gaussian Splatting approaches: they estimate view-dependent appearance without modeling scene materials and illumination. To address this limitation, we present an inverse rendering (IR) model capable of jointly estimating scene geometry, materials, and illumination. Our model builds upon recent NeRF-based IR approaches, but crucially introduces two novel physics-based priors that better constrain the IR estimation. Our priors are rigorously formulated as intuitive loss terms and achieve state-of-the-art material estimation without compromising novel view synthesis quality. Our method is easily adaptable to other inverse rendering and 3D reconstruction frameworks that require material estimation. We demonstrate the importance of extending current neural rendering approaches to fully model scene properties beyond geometry and view-dependent appearance. Code is publicly available at https://github.com/s3anwu/pbrnerf

Kittimate Chulajata, Sean Wu, Fabien Scalzo et al.

Forecasting winners in E-sports with real-time analytics has the potential to further engage audiences watching major tournament events. However, making such real-time predictions is challenging due to unpredictable variables within the game involving diverse player strategies and decision-making. Our work attempts to enhance audience engagement within video game tournaments by introducing a real-time method of predicting wins. Our Long Short Term Memory Network (LSTMs) based approach enables efficient predictions of win-lose outcomes by only using the health indicator of each player as a time series. As a proof of concept, we evaluate our model's performance within a classic, two-player arcade game, Super Street Fighter II Turbo. We also benchmark our method against state of the art methods for time series forecasting; i.e. Transformer models found in large language models (LLMs). Finally, we open-source our data set and code in hopes of furthering work in predictive analysis for arcade games.

Edward Phillips, Sean Wu, Boyan Gao et al.

Large language models present challenges for principled uncertainty quantification, in part due to their complexity and the diversity of their outputs. Semantic dispersion, or the variance in the meaning of sampled answers, has been proposed as a useful proxy for model uncertainty, but the associated computational cost prohibits its use in latency-critical applications. We show that sampled semantic distributions can be distilled into lightweight student models which estimate a prompt-conditioned uncertainty before the language model generates an answer token. The student model predicts a semantic distribution over possible answers; the entropy of this distribution provides an effective uncertainty signal for hallucination prediction, and the probability density allows candidate answers to be evaluated for reliability. On TriviaQA, our student models match or outperform finite-sample semantic dispersion for hallucination prediction and provide a strong signal for out-of-domain answer detection. We term this technique Semantic Self-Distillation (SSD), which we suggest provides a general framework for distilling predictive uncertainty in complex output spaces beyond language.

Evangelos Ntavelis, Sean Wu, Mohamad Shahbazi et al.

We propose HeadsUp, a scalable feed-forward method for reconstructing high-quality 3D Gaussian heads from large-scale multi-camera setups. Our method employs an efficient encoder-decoder architecture that compresses input views into a compact latent representation. This latent representation is then decoded into a set of UV-parameterized 3D Gaussians anchored to a neutral head template. This UV representation decouples the number of 3D Gaussians from the number and resolution of input images, enabling training with many high-resolution input views. We train and evaluate our model on an internal dataset with more than 10,000 subjects, which is an order of magnitude larger than existing multi-view human head datasets. HeadsUp achieves state-of-the-art reconstruction quality and generalizes to novel identities without test-time optimization. We extensively analyze the scaling behavior of our model across identities, views, and model capacity, revealing practical insights for quality-compute trade-offs. Finally, we highlight the strength of our latent space by showcasing two downstream applications: generating novel 3D identities and animating the 3D heads with expression blendshapes.

Sean Wu, Fredrik K. Gustafsson, Edward Phillips et al.

Large language models (LLMs) often produce confident but incorrect answers in settings where abstention would be safer. Standard evaluation protocols, however, require a response and do not account for how confidence should guide decisions under different risk preferences. To address this gap, we introduce the Behavioral Alignment Score (BAS), a decision-theoretic metric for evaluating how well LLM confidence supports abstention-aware decision making. BAS is derived from an explicit answer-or-abstain utility model and aggregates realized utility across a continuum of risk thresholds, yielding a measure of decision-level reliability that depends on both the magnitude and ordering of confidence. We show theoretically that truthful confidence estimates uniquely maximize expected BAS utility, linking calibration to decision-optimal behavior. BAS is related to proper scoring rules such as log loss, but differs structurally: log loss penalizes underconfidence and overconfidence symmetrically, whereas BAS imposes an asymmetric penalty that strongly prioritizes avoiding overconfident errors. Using BAS alongside widely used metrics such as ECE and AURC, we then construct a benchmark of self-reported confidence reliability across multiple LLMs and tasks. Our results reveal substantial variation in decision-useful confidence, and while larger and more accurate models tend to achieve higher BAS, even frontier models remain prone to severe overconfidence. Importantly, models with similar ECE or AURC can exhibit very different BAS due to highly overconfident errors, highlighting limitations of standard metrics. We further show that simple interventions, such as top-$k$ confidence elicitation and post-hoc calibration, can meaningfully improve confidence reliability. Overall, our work provides both a principled metric and a comprehensive benchmark for evaluating LLM confidence reliability.

Edward Phillips, Sean Wu, Soheila Molaei et al.

Large language models demonstrate impressive results across diverse tasks but are still known to hallucinate, generating linguistically plausible but incorrect answers to questions. Uncertainty quantification has been proposed as a strategy for hallucination detection, but no existing black-box approach provides estimates for both global and local uncertainty. The former attributes uncertainty to a batch of responses, while the latter attributes uncertainty to individual responses. Current local methods typically rely on white-box access to internal model states, whilst black-box methods only provide global uncertainty estimates. We introduce a geometric framework to address this, based on archetypal analysis of batches of responses sampled with only black-box model access. At the global level, we propose Geometric Volume, which measures the convex hull volume of archetypes derived from response embeddings. At the local level, we propose Geometric Suspicion, which ranks responses by reliability and enables hallucination reduction through preferential response selection. Unlike prior dispersion methods which yield only a single global score, our approach provides semantic boundary points which have utility for attributing reliability to individual responses. Experiments show that our framework performs comparably to or better than prior methods on short form question-answering datasets, and achieves superior results on medical datasets where hallucinations carry particularly critical risks. We also provide theoretical justification by proving a link between convex hull volume and entropy.

Sean Wu, Jun Yu Chen, Vahid Mohammadzadeh et al.

Perimetric measurements provide insight into a patient's peripheral vision and day-to-day functioning and are the main outcome measure for identifying progression of visual damage from glaucoma. However, visual field data can be noisy, exhibiting high variance, especially with increasing damage. In this study, we demonstrate the utility of self-supervised deep learning in denoising visual field data from over 4000 patients to enhance its signal-to-noise ratio and its ability to detect true glaucoma progression. We deployed both a variational autoencoder (VAE) and a masked autoencoder to determine which self-supervised model best smooths the visual field data while reconstructing salient features that are less noisy and more predictive of worsening disease. Our results indicate that including a categorical p-value at every visual field location improves the smoothing of visual field data. Masked autoencoders led to cleaner denoised data than previous methods, such as variational autoencoders. A 4.7% increase in detection of progressing eyes with pointwise linear regression (PLR) was observed. The masked and variational autoencoders' smoothed data predicted glaucoma progression 2.3 months earlier when p-values were included compared to when they were not. The faster prediction of time to progression (TTP) and the higher percentage progression detected support our hypothesis that masking out visual field elements during training while including p-values at each location would improve the task of detection of visual field progression. Our study has clinically relevant implications regarding masking when training neural networks to denoise visual field data, resulting in earlier and more accurate detection of glaucoma progression. This denoising model can be integrated into future models for visual field analysis to enhance detection of glaucoma progression.

Sean Wu, Jordan Hong, Keyu Bai et al. · eth-zurich

Due to their weak inductive bias, Multi-Layer Perceptrons (MLPs) have subpar performance at low-compute levels compared to standard architectures such as convolution-based networks (CNN). Recent work, however, has shown that the performance gap drastically reduces as the amount of compute is increased without changing the amount of inductive bias. In this work, we study the converse: in the low-compute regime, how does the incremental increase of inductive bias affect performance? To quantify inductive bias, we propose a "soft MLP" approach, which we coin Interpolated MLP (I-MLP). We control the amount of inductive bias in the standard MLP by introducing a novel algorithm based on interpolation between fixed weights from a prior model with high inductive bias. We showcase our method using various prior models, including CNNs and the MLP-Mixer architecture. This interpolation scheme allows fractional control of inductive bias, which may be attractive when full inductive bias is not desired (e.g. in the mid-compute regime). We find experimentally that for Vision Tasks in the low-compute regime, there is a continuous and two-sided logarithmic relationship between inductive bias and performance when using CNN and MLP-Mixer prior models.

Sean Wu, Nicole Amenta, Jiachen Zhou et al.

Following four successful years in the SAE AutoDrive Challenge Series I, the University of Toronto is participating in the Series II competition to develop a Level 4 autonomous passenger vehicle capable of handling various urban driving scenarios by 2025. Accurate detection of traffic lights and correct identification of their states is essential for safe autonomous operation in cities. Herein, we describe our recently-redesigned traffic light perception system for autonomous vehicles like the University of Toronto's self-driving car, Artemis. Similar to most traffic light perception systems, we rely primarily on camera-based object detectors. We deploy the YOLOv5 detector for bounding box regression and traffic light classification across multiple cameras and fuse the observations. To improve robustness, we incorporate priors from high-definition semantic maps and perform state filtering using hidden Markov models. We demonstrate a multi-camera, real time-capable traffic light perception pipeline that handles complex situations including multiple visible intersections, traffic light variations, temporary occlusion, and flashing light states. To validate our system, we collected and annotated a varied dataset incorporating flashing states and a range of occlusion types. Our results show superior performance in challenging real-world scenarios compared to single-frame, single-camera object detection.