Lionel Levine

Shayan Fazeli, Lionel Levine, Mehrab Beikzadeh et al.

Recent advances in remote health monitoring systems have significantly benefited patients and played a crucial role in improving their quality of life. However, while physiological health-focused solutions have demonstrated increasing success and maturity, mental health-focused applications have seen comparatively limited success in spite of the fact that stress and anxiety disorders are among the most common issues people deal with in their daily lives. In the hopes of furthering progress in this domain through the development of a more robust analytic framework for the measurement of indicators of mental health, we propose a multi-modal semi-supervised framework for tracking physiological precursors of the stress response. Our methodology enables utilizing multi-modal data of differing domains and resolutions from wearable devices and leveraging them to map short-term episodes to semantically efficient embeddings for a given task. Additionally, we leverage an inter-modality contrastive objective, with the advantages of rendering our framework both modular and scalable. The focus on optimizing both local and global aspects of our embeddings via a hierarchical structure renders transferring knowledge and compatibility with other devices easier to achieve. In our pipeline, a task-specific pooling based on an attention mechanism, which estimates the contribution of each modality on an instance level, computes the final embeddings for observations. This additionally provides a thorough diagnostic insight into the data characteristics and highlights the importance of signals in the broader view of predicting episodes annotated per mental health status. We perform training experiments using a corpus of real-world data on perceived stress, and our results demonstrate the efficacy of the proposed approach in performance improvements.

Haniyeh Ehsani Oskouie, Sajjad Ghiasvand, Lionel Levine et al.

As Artificial Intelligence (AI) models are increasingly integrated into critical systems, the need for a robust framework to establish the trustworthiness of AI is increasingly paramount. While collaborative efforts have established conceptual foundations for such a framework, there remains a significant gap in developing concrete, technically robust methods for assessing AI model quality and performance. This paper introduces a novel approach for assessing a newly trained model's performance based on another known model by calculating correlation between neural networks. The proposed method evaluates correlations by determining if, for each neuron in one network, there exists a neuron in the other network that produces similar output. This approach has implications for memory efficiency, allowing for the use of smaller networks when high correlation exists between networks of different sizes. Experiments on five fully connected networks and a two layer CNN trained on MNIST family datasets show that higher alignment with the CNN tracks stronger performance and smaller degradation under black box transfer based attacks. On ImageNet pretrained ResNets and DenseNets, partial layer comparisons recover intuitive architectural affinities, indicating that the procedure scales with reasonable approximations. These results support representational alignment as a lightweight compatibility check that complements standard accuracy, calibration, and robustness evaluations and enables early external validation of new models. Code is available at https://github.com/aheldis/Cross-model-Correlation.git.

Suvadip Sana, Sami Wolf, Neer Mehta et al.

Masked language models (MLMs) define local conditional distributions over tokens but do not, in general, correspond to any consistent joint distribution over sequences. This raises a fundamental question: what global distributional behavior is induced when such conditionals are used iteratively for generation? We address this question by modeling iterative masked-token resampling as a Glauber dynamics Markov chain on the discrete space of token sequences. We first show that MLM conditionals are intrinsically incompatible: we introduce a rectangle test that certifies this incompatibility and empirically verify its prevalence across modern MLMs. We then provide a theoretical analysis of the induced Markov chain. Under bounded cross-token influence, we establish a high-temperature contraction result implying $O(n\log n)$ mixing time where $n$ is the sequence length. In contrast, we prove that under a uniform local margin condition, the chain exhibits metastability, with exponentially slow escape from semantic basins at low temperatures. Empirically, we demonstrate a phase transition in mixing behavior as a function of temperature and sequence length, consistent with the theoretical predictions. We further characterize the induced stationary behavior through semantic trajectories, identifying persistent structures such as long-lived traps and recurrent semantic basins, with political content serving as a measurable case study.

Elliot Glazer, Ege Erdil, Tamay Besiroglu et al.

We introduce FrontierMath, a benchmark of hundreds of original, exceptionally challenging mathematics problems crafted and vetted by expert mathematicians. The questions cover most major branches of modern mathematics -- from computationally intensive problems in number theory and real analysis to abstract questions in algebraic geometry and category theory. Solving a typical problem requires multiple hours of effort from a researcher in the relevant branch of mathematics, and for the upper end questions, multiple days. FrontierMath uses new, unpublished problems and automated verification to reliably evaluate models while minimizing risk of data contamination. Current state-of-the-art AI models solve under 2% of problems, revealing a vast gap between AI capabilities and the prowess of the mathematical community. As AI systems advance toward expert-level mathematical abilities, FrontierMath offers a rigorous testbed that quantifies their progress.

Lionel Levine, Haniyeh Ehsani Oskouie, Sajjad Ghiasvand et al.

We introduce MI-to-Mid Distilled Compression (M2M-DC), a two-scale, shape-safe compression framework that interleaves information-guided block pruning with progressive inner slicing and staged knowledge distillation (KD). First, M2M-DC ranks residual (or inverted-residual) blocks by a label-aware mutual information (MI) signal and removes the least informative units (structured prune-after-training). It then alternates short KD phases with stage-coherent, residual-safe channel slicing: (i) stage "planes" (co-slicing conv2 out-channels with the downsample path and next-stage inputs), and (ii) an optional mid-channel trim (conv1 out / bn1 / conv2 in). This targets complementary redundancy, whole computational motifs and within-stage width while preserving residual shape invariants. On CIFAR-100, M2M-DC yields a clean accuracy-compute frontier. For ResNet-18, we obtain 85.46% Top-1 with 3.09M parameters and 0.0139 GMacs (72% params, 63% GMacs vs. teacher; mean final 85.29% over three seeds). For ResNet-34, we reach 85.02% Top-1 with 5.46M params and 0.0195 GMacs (74% / 74% vs. teacher; mean final 84.62%). Extending to inverted-residuals, MobileNetV2 achieves a mean final 68.54% Top-1 at 1.71M params (27%) and 0.0186 conv GMacs (24%), improving over the teacher's 66.03% by +2.5 points across three seeds. Because M2M-DC exposes only a thin, architecture-aware interface (blocks, stages, and down sample/skip wiring), it generalizes across residual CNNs and extends to inverted-residual families with minor legalization rules. The result is a compact, practical recipe for deployment-ready models that match or surpass teacher accuracy at a fraction of the compute.

Wilson Wu, John X. Morris, Lionel Levine

Do transformers "think ahead" during inference at a given position? It is known transformers prepare information in the hidden states of the forward pass at time step $t$ that is then used in future forward passes $t+τ$. We posit two explanations for this phenomenon: pre-caching, in which off-diagonal gradient terms present during training result in the model computing features at $t$ irrelevant to the present inference task but useful for the future, and breadcrumbs, in which features most relevant to time step $t$ are already the same as those that would most benefit inference at time $t+τ$. We test these hypotheses by training language models without propagating gradients to past timesteps, a scheme we formalize as myopic training. In a constructed synthetic data setting, we find clear evidence for pre-caching. In the autoregressive language modeling setting, our experiments are more suggestive of the breadcrumbs hypothesis, though pre-caching increases with model scale.

Dan Hendrycks, Dawn Song, Christian Szegedy et al.

The lack of a concrete definition for Artificial General Intelligence (AGI) obscures the gap between today's specialized AI and human-level cognition. This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult. To operationalize this, we ground our methodology in Cattell-Horn-Carroll theory, the most empirically validated model of human cognition. The framework dissects general intelligence into ten core cognitive domains-including reasoning, memory, and perception-and adapts established human psychometric batteries to evaluate AI systems. Application of this framework reveals a highly "jagged" cognitive profile in contemporary models. While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage. The resulting AGI scores (e.g., GPT-4 at 27%, GPT-5 at 57%) concretely quantify both rapid progress and the substantial gap remaining before AGI.

Lionel Levine, John Santerre, Alex S. Young et al.

We introduce PRISM (Predictive Reasoning in Sequential Medicine), a transformer-based architecture designed to model the sequential progression of clinical decision-making processes. Unlike traditional approaches that rely on isolated diagnostic classification, PRISM frames clinical trajectories as tokenized sequences of events - including diagnostic tests, laboratory results, and diagnoses - and learns to predict the most probable next steps in the patient diagnostic journey. Leveraging a large custom clinical vocabulary and an autoregressive training objective, PRISM demonstrates the ability to capture complex dependencies across longitudinal patient timelines. Experimental results show substantial improvements over random baselines in next-token prediction tasks, with generated sequences reflecting realistic diagnostic pathways, laboratory result progressions, and clinician ordering behaviors. These findings highlight the feasibility of applying generative language modeling techniques to structured medical event data, enabling applications in clinical decision support, simulation, and education. PRISM establishes a foundation for future advancements in sequence-based healthcare modeling, bridging the gap between machine learning architectures and real-world diagnostic reasoning.

Lionel Levine, John Santerre, Alexander S. Young et al.

We present PRISM-Consult, a clinician-aligned panel-of-experts architecture that extends the compact PRISM sequence model into a routed family of domain specialists. Episodes are tokenized as structured clinical events; a light-weight router reads the first few tokens and dispatches to specialist models (Cardiac-Vascular, Pulmonary, Gastro-Oesophageal, Musculoskeletal, Psychogenic). Each specialist inherits PRISM's small transformer backbone and token template, enabling parameter efficiency and interpretability. On real-world Emergency Department cohorts, specialists exhibit smooth convergence with low development perplexities across domains, while the router achieves high routing quality and large compute savings versus consult-all under a safety-first policy. We detail the data methodology (initial vs. conclusive ICD-9 families), routing thresholds and calibration, and report per-domain results to avoid dominance by common events. The framework provides a practical path to safe, auditable, and low-latency consult at scale, and we outline validation steps-external/temporal replication, asymmetric life-threat thresholds, and multi-label arbitration-to meet prospective clinical deployment standards.

Jonathn Chang, Leonhard Piff, Suvadip Sana et al.

Aligning AI with human values is a pressing unsolved problem. To address the lack of quantitative metrics for value alignment, we propose EigenBench: a black-box method for comparatively benchmarking language models' values. Given an ensemble of models, a constitution describing a value system, and a dataset of scenarios, our method returns a vector of scores quantifying each model's alignment to the given constitution. To produce these scores, each model judges the outputs of other models across many scenarios, and these judgments are aggregated with EigenTrust (Kamvar et al., 2003), yielding scores that reflect a weighted consensus judgment of the whole ensemble. EigenBench uses no ground truth labels, as it is designed to quantify subjective traits for which reasonable judges may disagree on the correct label. Hence, to validate our method, we collect human judgments on the same ensemble of models and show that EigenBench's judgments align closely with those of human evaluators. We further demonstrate that EigenBench can recover model rankings on the GPQA benchmark without access to objective labels, supporting its viability as a framework for evaluating subjective values for which no ground truths exist.

Arya Hatamian, Lionel Levine, Haniyeh Ehsani Oskouie et al.

Having a sufficient quantity of quality data is a critical enabler of training effective machine learning models. Being able to effectively determine the adequacy of a dataset prior to training and evaluating a model's performance would be an essential tool for anyone engaged in experimental design or data collection. However, despite the need for it, the ability to prospectively assess data sufficiency remains an elusive capability. We report here on two experiments undertaken in an attempt to better ascertain whether or not basic descriptive statistical measures can be indicative of how effective a dataset will be at training a resulting model. Leveraging the effect size of our features, this work first explores whether or not a correlation exists between effect size, and resulting model performance (theorizing that the magnitude of the distinction between classes could correlate to a classifier's resulting success). We then explore whether or not the magnitude of the effect size will impact the rate of convergence of our learning rate, (theorizing again that a greater effect size may indicate that the model will converge more rapidly, and with a smaller sample size needed). Our results appear to indicate that this is not an effective heuristic for determining adequate sample size or projecting model performance, and therefore that additional work is still needed to better prospectively assess adequacy of data.