Michael Chen

Peter S. Park, Simon Goldstein, Aidan O'Gara et al. · berkeley

This paper argues that a range of current AI systems have learned how to deceive humans. We define deception as the systematic inducement of false beliefs in the pursuit of some outcome other than the truth. We first survey empirical examples of AI deception, discussing both special-use AI systems (including Meta's CICERO) built for specific competitive situations, and general-purpose AI systems (such as large language models). Next, we detail several risks from AI deception, such as fraud, election tampering, and losing control of AI systems. Finally, we outline several potential solutions to the problems posed by AI deception: first, regulatory frameworks should subject AI systems that are capable of deception to robust risk-assessment requirements; second, policymakers should implement bot-or-not laws; and finally, policymakers should prioritize the funding of relevant research, including tools to detect AI deception and to make AI systems less deceptive. Policymakers, researchers, and the broader public should work proactively to prevent AI deception from destabilizing the shared foundations of our society.

Hjalmar Wijk, Tao Lin, Joel Becker et al.

Frontier AI safety policies highlight automation of AI research and development (R&D) by AI agents as an important capability to anticipate. However, there exist few evaluations for AI R&D capabilities, and none that are highly realistic and have a direct comparison to human performance. We introduce RE-Bench (Research Engineering Benchmark, v1), which consists of 7 challenging, open-ended ML research engineering environments and data from 71 8-hour attempts by 61 distinct human experts. We confirm that our experts make progress in the environments given 8 hours, with 82% of expert attempts achieving a non-zero score and 24% matching or exceeding our strong reference solutions. We compare humans to several public frontier models through best-of-k with varying time budgets and agent designs, and find that the best AI agents achieve a score 4x higher than human experts when both are given a total time budget of 2 hours per environment. However, humans currently display better returns to increasing time budgets, narrowly exceeding the top AI agent scores given an 8-hour budget, and achieving 2x the score of the top AI agent when both are given 32 total hours (across different attempts). Qualitatively, we find that modern AI agents possess significant expertise in many ML topics -- e.g. an agent wrote a faster custom Triton kernel than any of our human experts' -- and can generate and test solutions over ten times faster than humans, at much lower cost. We open-source the evaluation environments, human expert data, analysis code and agent trajectories to facilitate future research.

Siqiao Huang, Partha Kaushik, Michael Chen et al.

World models have become a central paradigm for learning predictive simulators that support generation, planning, and decision-making. Yet, despite rapid progress in industry-scale interactive video generation, the broader research community still lacks compact, reproducible, and easily extensible implementations for studying the design choices underlying modern world models. We introduce Nano World Models, a minimalist codebase for future video prediction centered around diffusion forcing. Nano World Models provides a unified interface for generative objectives, model scales, action-conditioning mechanisms, latent observation spaces, datasets, evaluation protocols, and long-horizon rollout procedures. This design enables controlled studies of world-modeling components that are often entangled across separate implementations. Through experiments across simple control environments, game simulation, and real-robot data, we examine how prediction parameterization, architecture scale, action injection, sampling budget, and domain complexity affect video prediction quality and autoregressive rollout behavior. By releasing code, configurations, evaluation scripts, and pretrained checkpoints, Nano World Models aims to provide a compact yet extensible experimental substrate for open, reproducible, and scientific world-model research.

Emmy Liu, Amanda Bertsch, Lintang Sutawika et al. · cmu

Improvements in language model capabilities are often attributed to increasing model size or training data, but in some cases smaller models trained on curated data or with different architectural decisions can outperform larger ones trained on more tokens. What accounts for this? To quantify the impact of these design choices, we meta-analyze 92 open-source pretrained models across a wide array of scales, including state-of-the-art open-weights models as well as less performant models and those with less conventional design decisions. We find that by incorporating features besides model size and number of training tokens, we can achieve a relative 3-28% increase in ability to predict downstream performance compared with using scale alone. Analysis of model design decisions reveal insights into data composition, such as the trade-off between language and code tasks at 15-25\% code, as well as the better performance of some architectural decisions such as choosing rotary over learned embeddings. Broadly, our framework lays a foundation for more systematic investigation of how model development choices shape final capabilities.

Jennah Gosciak, Eric Giannella, Zhaowen Guo et al.

Social service programs like the Supplemental Nutrition Assistance Program (SNAP, or food stamps) have eligibility rules that can be challenging to understand. For nonprofit caseworkers who often support clients in navigating a dozen or more complex programs, LLM-based chatbots may offer a means to provide better, faster help to clients whose situations may be less common. In this paper, we measure the potential effects of LLM-based chatbot suggestions on caseworkers' ability to provide accurate guidance. We first created a 770-question multiple-choice benchmark dataset of difficult, but realistic questions that a caseworker might receive. Next, using these benchmark questions and corresponding expert-verified answers, we conducted a randomized experiment with caseworkers recruited from nonprofit outreach organizations in Los Angeles. Caseworkers in the control condition did not see chatbot suggestions and had a mean accuracy of 49%. Caseworkers in the treatment condition saw chatbot suggestions that we artificially varied to range in aggregate accuracy from low (53%) to high (100%). Caseworker performance significantly improves as chatbot quality improves: high-quality chatbots (96-100% accurate) improved caseworker accuracy by 27 percentage points. At the question-level, incorrect chatbot suggestions substantially reduce caseworker accuracy, with a two-thirds reduction on easy questions where the control group performed best (without chatbot suggestions). Finally, improvements in caseworker accuracy level off as chatbot accuracy increases, a phenomenon that we call the "AI underreliance plateau," which is a concern for real-world deployment and highlights the importance of evaluating human-in-the-loop tools with their users.

Nathaniel Li, Alexander Pan, Anjali Gopal et al. · berkeley, cmu

The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks of malicious use, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private, preventing further research into mitigating risk. Furthermore, they focus on only a few, highly specific pathways for malicious use. To fill these gaps, we publicly release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. WMDP was developed by a consortium of academics and technical consultants, and was stringently filtered to eliminate sensitive information prior to public release. WMDP serves two roles: first, as an evaluation for hazardous knowledge in LLMs, and second, as a benchmark for unlearning methods to remove such hazardous knowledge. To guide progress on unlearning, we develop RMU, a state-of-the-art unlearning method based on controlling model representations. RMU reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Michael Chen, Raghav Kansal, Abhijith Gandrakota et al.

We present a systematic evaluation of representation learning objectives for particle physics within a unified framework. Our study employs a shared transformer-based particle-cloud encoder with standardized preprocessing, matched sampling, and a consistent evaluation protocol on a jet classification dataset. We compare contrastive (supervised and self-supervised), masked particle modeling, and generative reconstruction objectives under a common training regimen. In addition, we introduce targeted supervised architectural modifications that achieve state-of-the-art performance on benchmark evaluations. This controlled comparison isolates the contributions of the learning objective, highlights their respective strengths and limitations, and provides reproducible baselines. We position this work as a reference point for the future development of foundation models in particle physics, enabling more transparent and robust progress across the community.

Michael Chen

There is a growing need to evaluate Large Language Models (LLMs) on complex, high-impact, real-world tasks to assess their true readiness as reasoning agents. To address this gap, we introduce AgentCaster, a contamination-free framework employing multimodal LLMs end-to-end for the challenging, long-horizon task of tornado forecasting. Within AgentCaster, models interpret heterogeneous spatiotemporal data from a high-resolution convection-allowing forecast archive. We assess model performance over a 40-day period featuring diverse historical data, spanning several major tornado outbreaks and including over 500 tornado reports. Each day, models query interactively from a pool of 3,625 forecast maps and 40,125 forecast soundings for a forecast horizon of 12-36 hours. Probabilistic tornado-risk polygon predictions are verified against ground truths derived from geometric comparisons across disjoint risk bands in projected coordinate space. To quantify accuracy, we propose domain-specific TornadoBench and TornadoHallucination metrics, with TornadoBench highly challenging for both LLMs and domain expert human forecasters. Notably, human experts significantly outperform state-of-the-art models, which demonstrate a strong tendency to hallucinate and overpredict risk intensity, struggle with precise geographic placement, and exhibit poor spatiotemporal reasoning in complex, dynamically evolving systems. AgentCaster aims to advance research on improving LLM agents for challenging reasoning tasks in critical domains.

Zongyi Li, Samuel Lanthaler, Catherine Deng et al.

Machine learning (ML) models have emerged as a promising approach for solving partial differential equations (PDEs) in science and engineering. Previous ML models typically cannot generalize outside the training data; for example, a trained ML model for the Navier-Stokes equations only works for a fixed Reynolds number ($Re$) on a pre-defined domain. To overcome these limitations, we propose a data augmentation scheme based on scale-consistency properties of PDEs and design a scale-informed neural operator that can model a wide range of scales. Our formulation leverages the facts: (i) PDEs can be rescaled, or more concretely, a given domain can be re-scaled to unit size, and the parameters and the boundary conditions of the PDE can be appropriately adjusted to represent the original solution, and (ii) the solution operators on a given domain are consistent on the sub-domains. We leverage these facts to create a scale-consistency loss that encourages matching the solutions evaluated on a given domain and the solution obtained on its sub-domain from the rescaled PDE. Since neural operators can fit to multiple scales and resolutions, they are the natural choice for incorporating scale-consistency loss during training of neural PDE solvers. We experiment with scale-consistency loss and the scale-informed neural operator model on the Burgers' equation, Darcy Flow, Helmholtz equation, and Navier-Stokes equations. With scale-consistency, the model trained on $Re$ of 1000 can generalize to $Re$ ranging from 250 to 10000, and reduces the error by 34% on average of all datasets compared to baselines.

Lily Stelling, Mick Yang, Rokas Gipiškis et al.

This report provides a detailed comparison between the Safety and Security measures proposed in the EU AI Act's General-Purpose AI (GPAI) Code of Practice (Third Draft) and the current commitments and practices voluntarily adopted by leading AI companies. As the EU moves toward enforcing binding obligations for GPAI model providers, the Code of Practice will be key for bridging legal requirements with concrete technical commitments. Our analysis focuses on the draft's Safety and Security section (Commitments II.1-II.16), documenting excerpts from current public-facing documents that are relevant to each individual measure. We systematically reviewed different document types, such as companies' frontier safety frameworks and model cards, from over a dozen companies, including OpenAI, Anthropic, Google DeepMind, Microsoft, Meta, Amazon, and others. This report is not meant to be an indication of legal compliance, nor does it take any prescriptive viewpoint about the Code of Practice or companies' policies. Instead, it aims to inform the ongoing dialogue between regulators and General-Purpose AI model providers by surfacing evidence of industry precedent for various measures. Nonetheless, we were able to find relevant quotes from at least 5 companies' documents for the majority of the measures in Commitments II.1-II.16.

Emily Wang, Michael Chen, Chao Li

In this paper, we propose a novel \emph{uncertainty-aware graph self-training} approach for semi-supervised node classification. Our method introduces an Expectation-Maximization (EM) regularization scheme to incorporate an uncertainty mechanism during pseudo-label generation and model retraining. Unlike conventional graph self-training pipelines that rely on fixed pseudo-labels, our approach iteratively refines label confidences with an EM-inspired uncertainty measure. This ensures that the predictive model focuses on reliable graph regions while gradually incorporating ambiguous nodes. Inspired by prior work on uncertainty-aware self-training techniques~\cite{wang2024uncertainty}, our framework is designed to handle noisy graph structures and feature spaces more effectively. Through extensive experiments on several benchmark graph datasets, we demonstrate that our method outperforms strong baselines by a margin of up to 2.5\% in accuracy while maintaining lower variance in performance across multiple runs.

Michael Chen, Mike D'Arcy, Alisa Liu et al.

Commonsense reasoning is a critical AI capability, but it is difficult to construct challenging datasets that test common sense. Recent neural question answering systems, based on large pre-trained models of language, have already achieved near-human-level performance on commonsense knowledge benchmarks. These systems do not possess human-level common sense, but are able to exploit limitations of the datasets to achieve human-level scores. We introduce the CODAH dataset, an adversarially-constructed evaluation dataset for testing common sense. CODAH forms a challenging extension to the recently-proposed SWAG dataset, which tests commonsense knowledge using sentence-completion questions that describe situations observed in video. To produce a more difficult dataset, we introduce a novel procedure for question acquisition in which workers author questions designed to target weaknesses of state-of-the-art neural question answering systems. Workers are rewarded for submissions that models fail to answer correctly both before and after fine-tuning (in cross-validation). We create 2.8k questions via this procedure and evaluate the performance of multiple state-of-the-art question answering systems on our dataset. We observe a significant gap between human performance, which is 95.3%, and the performance of the best baseline accuracy of 67.5% by the BERT-Large model.

Michael Kellman, Emrah Bostan, Michael Chen et al.

Fourier Ptychographic Microscopy (FPM) is a computational imaging method that is able to super-resolve features beyond the diffraction-limit set by the objective lens of a traditional microscope. This is accomplished by using synthetic aperture and phase retrieval algorithms to combine many measurements captured by an LED array microscope with programmable source patterns. FPM provides simultaneous large field-of-view and high resolution imaging, but at the cost of reduced temporal resolution, thereby limiting live cell applications. In this work, we learn LED source pattern designs that compress the many required measurements into only a few, with negligible loss in reconstruction quality or resolution. This is accomplished by recasting the super-resolution reconstruction as a Physics-based Neural Network and learning the experimental design to optimize the network's overall performance. Specifically, we learn LED patterns for different applications (e.g. amplitude contrast and quantitative phase imaging) and show that the designs we learn through simulation generalize well in the experimental setting. Further, we discuss a context-specific loss function, practical memory limitations, and interpretability of our learned designs.

David Hallac, Suvrat Bhooshan, Michael Chen et al.

With automobiles becoming increasingly reliant on sensors to perform various driving tasks, it is important to encode the relevant CAN bus sensor data in a way that captures the general state of the vehicle in a compact form. In this paper, we develop a deep learning-based method, called Drive2Vec, for embedding such sensor data in a low-dimensional yet actionable form. Our method is based on stacked gated recurrent units (GRUs). It accepts a short interval of automobile sensor data as input and computes a low-dimensional representation of that data, which can then be used to accurately solve a range of tasks. With this representation, we (1) predict the exact values of the sensors in the short term (up to three seconds in the future), (2) forecast the long-term average values of these same sensors, (3) infer additional contextual information that is not encoded in the data, including the identity of the driver behind the wheel, and (4) build a knowledge base that can be used to auto-label data and identify risky states. We evaluate our approach on a dataset collected by Audi, which equipped a fleet of test vehicles with data loggers to store all sensor readings on 2,098 hours of driving on real roads. We show in several experiments that our method outperforms other baselines by up to 90%, and we further demonstrate how these embeddings of sensor data can be used to solve a variety of real-world automotive applications.

Aditya Grover, Todor Markov, Peter Attia et al.

We propose a generalization of the best arm identification problem in stochastic multi-armed bandits (MAB) to the setting where every pull of an arm is associated with delayed feedback. The delay in feedback increases the effective sample complexity of standard algorithms, but can be offset if we have access to partial feedback received before a pull is completed. We propose a general framework to model the relationship between partial and delayed feedback, and as a special case we introduce efficient algorithms for settings where the partial feedback are biased or unbiased estimators of the delayed feedback. Additionally, we propose a novel extension of the algorithms to the parallel MAB setting where an agent can control a batch of arms. Our experiments in real-world settings, involving policy search and hyperparameter optimization in computational sustainability domains for fast charging of batteries and wildlife corridor construction, demonstrate that exploiting the structure of partial feedback can lead to significant improvements over baselines in both sequential and parallel MAB.

Li-Hao Yeh, Jonathan Dong, Jingshan Zhong et al.

Fourier ptychography is a new computational microscopy technique that provides gigapixel-scale intensity and phase images with both wide field-of-view and high resolution. By capturing a stack of low-resolution images under different illumination angles, a nonlinear inverse algorithm can be used to computationally reconstruct the high-resolution complex field. Here, we compare and classify multiple proposed inverse algorithms in terms of experimental robustness. We find that the main sources of error are noise, aberrations and mis-calibration (i.e. model mis-match). Using simulations and experiments, we demonstrate that the choice of cost function plays a critical role, with amplitude-based cost functions performing better than intensity-based ones. The reason for this is that Fourier ptychography datasets consist of images from both brightfield and darkfield illumination, representing a large range of measured intensities. Both noise (e.g. Poisson noise) and model mis-match errors are shown to scale with intensity. Hence, algorithms that use an appropriate cost function will be more tolerant to both noise and model mis-match. Given these insights, we propose a global Newton's method algorithm which is robust and computationally efficient. Finally, we discuss the impact of procedures for algorithmic correction of aberrations and mis-calibration.