Publications
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Our teams aspire to make discoveries that impact everyone, and core to our approach is sharing our research and tools to fuel progress in the field.
Sort By
1 - 15 of 11189 publications
Expert evaluation of LLM world models: A high-Tc superconductivity case study
Haoyu Guo
Maria Tikhanovskaya
Paul Raccuglia
Alexey Vlaskin
Chris Co
Scott Ellsworth
Matthew Abraham
Lizzie Dorfman
Peter Armitage
Chunhan Feng
Antoine Georges
Olivier Gingras
Dominik Kiese
Steve Kivelson
Vadim Oganesyan
Brad Ramshaw
Subir Sachdev
Senthil Todadri
John Tranquada
Eun-Ah Kim
Proceedings of the National Academy of Sciences (2026)
Preview abstract
Large Language Models (LLMs) show great promise as a powerful tool for scientific literature exploration. However, their effectiveness in providing scientifically accurate and comprehensive answers to complex questions within specialized domains remains an active area of research. This work evaluates the performance of six different LLM-based systems for answering scientific literature questions, including commercially available closed models and a custom retrieval-augmented generation (RAG) system capable of retrieving images alongside text. We conduct a rigorous expert evaluation of the systems in the domain of high-temperature cuprate superconductors, a research area that involves material science, experimental physics, computation, and theoretical physics. We use an expert-curated database of 1726 scientific papers and a set of 67 expert-formulated questions. The evaluation employs a multi-faceted rubric assessing balanced perspectives, factual comprehensiveness, succinctness, evidentiary support, and image relevance. Our results demonstrate that RAG-based systems, powered by curated data and multimodal retrieval, outperform existing closed models across key metrics, particularly in providing comprehensive and well-supported answers, and in retrieving relevant visual information. This study provides valuable insights into designing and evaluating specialized scientific literature understanding systems, particularly with expert involvement, while also highlighting the importance of rich, domain-specific data in such systems.
View details
Preview abstract
In some multi-stage software build pipelines, downstream compiler errors may be reported against ephemeral, machine-generated intermediate artifacts rather than original, human-written source code, which can make remediation challenging. A system and method may address this by intercepting a downstream error, mapping its location back to the original source file, and programmatically injecting a dormant suppression tag into the original source code. During a subsequent build, an intermediate transpiler can propagate this tag into a newly generated intermediate artifact. In the intermediate file, the tag may become active and be recognized by the downstream compiler as a directive to suppress the specific error. This approach can facilitate an automated remediation process for certain build failures that avoids direct modification of ephemeral files and uses the original source code as a record for suppression.
View details
Preview abstract
Managing compiler build errors that can arise during infrastructure upgrades in large, polyglot codebases may be challenging, as manual remediation can be slow and some automated tools may not support modern language syntax. A system can provide automated error remediation by ingesting compiler diagnostics and analyzing source code using an Abstract Syntax Tree (AST). A recursive scope resolution algorithm, for example, can traverse the AST to identify a specific and narrowly-scoped code block at which to apply an error suppression. Conversely, this algorithmic complexity can be bypassed when lexical scope resolution is not required, and the system can identify the specific location of error suppressions directly from the error's exact coordinates. The system may then generate and apply language-specific patches, such as structured comments for JavaScript source files or line-scoped comments for TypeScript source files, for example, by using a transactional rewrite engine. This approach can provide a scalable method for managing automated code remediation, which may facilitate infrastructure upgrades by reducing the need for manual intervention.
View details
A probabilistic framework for learning non‐intrusive corrections to long‐time climate simulations from short‐time training data
Benedikt Barthel
Rob Carver
Fei Sha
Themistoklis Sapsis
Journal of Advances in Modeling Earth Systems (2026)
Preview abstract
Despite advances in high performance computing, accurate numerical simulations of global atmospheric dynamics remain a challenge. The resolution required to fully resolve the vast range scales as well as the strong coupling with—often not fully-understood—physics renders such simulations computationally infeasible over time horizons relevant for long-term climate risk assessment. While data-driven parameterizations have shown some promise of alleviating these obstacles, the scarcity of high-quality training data and their lack of long-term stability typically hinders their ability to capture the risk of rare extreme events. In this work we present a general strategy for training variational (probabilistic) neural network models to non-intrusively correct under-resolved long-time simulations of turbulent climate systems. The approach is based on the paradigm introduced by Barthel Sorensen et al. (2024, https://doi.org/10.1029/2023ms004122) which involves training a post-processing correction operator on under-resolved simulations nudged toward a high-fidelity reference. Our variational framework enables us to learn the dynamics of the underlying system from very little training data and thus drastically improve the extrapolation capabilities of the previous deterministic state-of-the art—even when the statistics of that training data are far from converged. We investigate and compare three recently introduced variational network architectures and illustrate the benefits of our approach on an anisotropic quasi-geostrophic flow. For this prototype model our approach is able to not only accurately capture global statistics, but also the anistropic regional variation and the statistics of multiple extreme event metrics—demonstrating significant improvement over previously introduced deterministic architectures.
View details
KVCIS: Activation-Based Token Importance Prediction for Intelligent KV-Cache Compression
Zenodo (2026)
Preview abstract
We introduce KVCIS (KV-Cache Importance Scoring), a novel approach to KV-cache compression that predicts token importance from intermediate-layer activations before attention is computed. Unlike existing methods (H2O, StreamingLLM, Scissorhands) that make compression decisions based on attention scores computed during generation, KVCIS enables proactive compression at cache insertion time—determining how to store each token before paying the computational cost of attention. We discover a two-level importance structure in decoder-only transformers: the beginning-of-sequence (BOS) token acts as an "attention sink" receiving ~76% of attention, while the remaining ~24% is distributed across content tokens with 10-11× importance spread. A simple linear probe achieves R² = 0.998 overall and R² = 0.68–0.79 for discriminating among content tokens. Extensive validation across 3 model families (Llama, Mistral, Gemma), 8 layer depths, context lengths from 256 to 2048 tokens, and multiple downstream tasks demonstrates: 50% memory reduction with zero degradation on NarrativeQA (F1 = 0.064 matching baseline exactly), while uniform quantization degrades by 7.8% at the same compression ratio. KVCIS consistently achieves 5–8× better quality preservation than uniform quantization across all tested context lengths. The memory savings enable increased batch sizes and longer context support; the probe itself adds minimal overhead (~16KB direction vector, 0.06ms per token). This work extends activation-based probing from safety classification to inference optimization, demonstrating that intermediate-layer activations encode predictive signals about token importance for generation.
View details
Phoenix: Rowhammer Attacks on DDR5 with Self-Correcting Synchronization
Preview
Michele Marazzi
Kaveh Razavi
Salman Qazi
Diego Meyer
Patrick Jattke
IEEE Security & Privacy (S&P) (2026)
Preview abstract
We study the problem of allocating access point bandwidth to users of a wireless network in the presence of adversarial jamming. Specifically, we consider a setting in which the network designer acts first and allocates access point bandwidth to the users of the network, before an adversary applies a jamming strategy to reduce the bandwidth of a subset (or all) of the access points. We consider a strong adversary who has complete information and can optimize the jamming strategy, subject to power budget constraints. In turn, the network designer must allocate the resources in anticipation of the adversary's actions.
We explain that our model gives rise to a special network interdiction model, which differs from the standard setting in two ways: The first is that the interdictor is given the benefit of responding, rather than leading the game. The second is that the interdiction is fractional and performed at the node level of the network. The interdiction then propagates to all edges incident to the access point.
In terms of technical results, we provide an allocation algorithm that is based on linear programming duality and show that the algorithm can solve the problem optimally, assuming knowledge of the adversary's budget constraints. We conduct experiments on synthetic data to show the extent to which the algorithm improves the total utilized bandwidth over the algorithm that optimizes bandwidth allocation while being oblivious to the adversary's existence.
View details
Preview abstract
We introduce ALPS (Activation-based Length Prediction for Scheduling), a method for predicting LLM generation length from prefill activations before any tokens are generated. Unlike existing approaches that require model fine-tuning or complex entropy-weighted pooling, ALPS uses a simple linear probe on the last-token activation at intermediate layers. We discover that generation length is encoded in prefill representations: a ridge regression probe achieves R-squared > 0.85 across three model families. Validation across Llama-3.1-8B, Gemma-2-9B, and Qwen-2.5-7B demonstrates: (1) intermediate layers generally perform well, with some architectural variation; (2) simple last-token extraction outperforms complex pooling strategies; (3) activations improve substantially over surface-feature baselines (24 percentage points over input length plus lexical features). The best models achieve R-squared = 0.943 (Gemma), R-squared = 0.880 (Llama), and R-squared = 0.857 (Qwen) with MAE of 38-80 tokens. All test prompts terminated naturally (100% EOS), eliminating truncation confounds. While our evaluation uses 200 curated prompts—sufficient for demonstrating the phenomenon but requiring broader validation—cross-validation confirms generalization beyond training data. ALPS enables practical applications including budget-constrained inference, request scheduling, and resource allocation. The probe adds negligible overhead (~16KB direction vector, single dot product), making ALPS practical for production deployment.
View details
From Correctness to Collaboration: A Human-Centered Taxonomy of AI Agent Behavior in Software Engineering
Sherry Shi
Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26), ACM, New York, NY, USA (2026)
Preview abstract
The ongoing transition of Large Language Models in software engineering from code generators into autonomous agents requires a shift in how we define and measure success. While models are becoming more capable, the industry lacks a clear understanding of the behavioral norms that make an agent effective in collaborative software development in the enterprise. This work addresses this gap by presenting a taxonomy of desirable agent behaviors, synthesized from 91 sets of developer-defined rules for coding agents. We identify four core expectations: Adhere to Standards and Processes, Ensure Code Quality and Reliability, Solve Problems Effectively, and Collaborate with the Developer. These findings offer a concrete vocabulary for agent behavior, enabling researchers to move beyond correctness-only benchmarks and start designing evaluations that reflect the socio-technical nature of professional software development in enterprises.
View details
Approximate vs Precise: An experiment in what impacts user choice when apps request location access
Jessica Johnson
Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26), April 13–17, 2026, Barcelona, Spain (2026)
Preview abstract
User location data is highly sensitive, yet commonly requested by mobile apps for both core functionality and monetization. To improve user privacy, the major mobile platforms, Android and iOS, made changes so that when apps request precise location access, users can choose to share only their approximate location. However, the platforms have diverging interfaces: Android offers a side-by-side choice and iOS offers a corner toggle. This study evaluates which factors impact users’ choices when apps request location access via a randomized controlled experiment with 2579 US Android users. We tested the impact of app type, whether a reason for the request was provided, and the quality and content of the reason, including monetization. We do not find the reasons have an effect. Instead, we find users’ choices are impacted by app type and user demographics. We find that when users are given a side-by-side choice to allow approximate versus precise location access, they make reasonable choices. Of users who allowed access, the vast majority (90.7%) chose precise for a rideshare app versus the majority (71.3%) chose approximate for a local news app. Concerningly, the majority also allowed location access to a wallpaper app, and older users were significantly more likely to allow apps precise location access. We conclude by discussing implications for app platforms and future work.
View details
Neural general circulation models for modeling precipitation
Stephan Hoyer
Dmitrii Kochkov
Janni Yuval
Ian Langmore
Science Advances (2026)
Preview abstract
Climate models struggle to accurately simulate precipitation, particularly extremes and the diurnal cycle. While hybrid models combining machine learning and physics have emerged with the premise of improving precipitation simulations, none have proven sufficiently skillful or stable enough to outperform existing models in simulating precipitation.
Here, we present the first hybrid model that is trained directly on precipitation observations. The model runs at 2.8 degrees resolution and is built on the differentiable NeuralGCM framework. This model is stable for decadal simulations and demonstrates significant improvements over existing GCMs, ERA5 reanalysis, and a Global Cloud-Resolving Model in simulating precipitation. Our approach yields reduced biases, a more realistic precipitation distribution, improved representation of extremes, and a more accurate diurnal cycle. Furthermore, it outperforms the ECMWF ensemble for mid-range weather forecasting.
This advance paves the way for more reliable simulations of current climate and for the ability to fully utilize the abundance of existing observations to further improve GCMs.
View details
Preview abstract
We introduce AASE (Activation-based AI Safety Enforcement), a framework for post-perception safety monitoring in large language models. Unlike pre-perception approaches that analyze input or output text, AASE monitors the model's internal activation patterns—what the model "understands" rather than what text it processes or generates—enabling detection of safety-relevant states before harmful outputs are produced. The framework comprises three techniques: Activation Fingerprinting (AF) for harmful content detection, Agent Action Gating (AAG) for prompt injection defense, and Activation Policy Compliance (APC) for enterprise policy enforcement. We introduce paired contrastive training to isolate safety-relevant signals from confounding factors such as topic and style, addressing signal entanglement in polysemantic activations. Validation across 7 models from 3 architecture families shows strong class separation: Gemma-2-9B achieves AUC 1.00 with 7.2σ separation across all probes; AAG achieves AUC ≥0.88 across all models on the InjecAgent benchmark; APC achieves 0.97-1.00 AUC across three enterprise policies. Model size correlates with probe quality—Gemma-2-9B (7.2σ separation) outperforms Gemma-2-2B (4.3σ). All techniques survive INT4 quantization with minimal separation degradation. AASE is 9× faster than Llama Guard 3 (33ms vs 306ms) with higher TPR (88% vs 50%) at a tunable threshold that trades FPR for detection sensitivity, adding only 0.002ms probe overhead to existing inference.
View details
"It didn’t feel right but I needed a job so desperately": Understanding People's Emotions & Help Needs During Financial Scams
G. Jake Chanenson
Tara Matthews
Jessica McClearn
Sarah Meiklejohn
Mia Hassoun
2026
Preview abstract
Online financial scams represent a long-standing and serious threat for which people seek help. We present a study to understand people’s in situ motivations for engaging with scams and the help needs they express before, during, and after encountering a scam. We identify the main emotions scammers exploited (e.g., fear, hope) and characterize how they did so. We examine factors—such as financial insecurity and legal precarity—which elevate people’s risk of engaging with specific scams and experiencing harm. We indicate when people sought help and describe their help-seeking needs and emotions at different stages of the scam. We discuss how these needs could be met through the design of contextually-specific prevention, diagnostic, mitigation, and recovery interventions.
View details
ConvApparel: A Benchmark Dataset and Validation Framework for User Simulators in Conversational Recommenders
Jihwan Jeong
The 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL-26), Rabat, Morocco (2026)
Preview abstract
LLM-based user simulators are a scalable solution for improving conversational AI, but a critical realism gap undermines their effectiveness. To close this gap, we introduce a framework for building and validating high-fidelity simulators. We present a novel dataset of human-AI shopping conversations designed to capture a wide spectrum of user experiences. To measure fidelity, we propose a hybrid evaluation protocol that combines statistical alignment with a learned, discriminator-based Human-Likeness Score. Our most sophisticated simulator, trained via reinforcement learning with iterative critique, achieves a significant leap in realism. Critically, we demonstrate through counterfactual validation that our simulator—trained exclusively on optimal interactions—realistically adapts its behavior to suboptimal system responses, mirroring real user reactions and marking a key advance in creating reliable simulators for robust AI development.
View details
