Publications

EZH2 Inhibition Induces a Metabolic Stress Response Sensitizing Triple-Negative Breast Cancer to Glutaminase Targeting
Lucas Porras
Marina Fukano
Ann-Sophie Gironne
Hugo Philippeau
Gabriel Alzial
Elise Quadri
Jeanne Cormery
Yousef Alaessa
Anie Monast
Faustine Gorse
Myriame St-Arnaud
Mariana De Sa Tavares Russo
Sylvie Mader
Daina Zofija Avizonis
Morag Park
Geneviève Deblois
EZH2, the catalytic subunit of the histone methyltransferase complex PRC2, is overexpressed and associated with poor prognosis in triple-neg… (see more)ative breast cancer (TNBC). Although EZH2 inhibition significantly alters chromatin landscapes and gene expression, it has limited impact on the growth of TNBC models, suggesting adaptive compensatory mechanisms. Here, we demonstrated that EZH2 inhibition causes the accumulation of misfolded proteins and double-stranded RNA (dsRNA), triggering an essential integrated stress response (ISR) through PKR and PERK activation. By inducing ISR-mediated ATF4, EZH2 inhibition enhanced amino acid flux and promoted glutaminolysis to support TNBC cell survival. Pharmacological targeting of this metabolic axis with a glutaminase inhibitor in combination with EZH2 inhibition significantly impaired TNBC cell proliferation and tumor growth. These findings reveal a stress-driven metabolic adaptation that sustains TNBC survival upon EZH2 blockade and highlight inhibition of this pathway as a strategy to enhance the efficacy of EZH2 inhibitors in TNBC.
HyperBrowseComp: A Multilingual and Multimodal Stress Test for Web-Browsing Agents
Alham Fikri Aji
Faiz Rizki Ramadhan
Zayd M. K. Zuhri
Seung Hun Eddie Han
Ryandito Diandaru
Qinrong Cui
Jan Christian Blaise Cruz
Badrinath Chandana
Peerawat Chomphooyod
Ahmed Attia
Jonibek Mansurov
Emilio Villa-Cueva
Canh Duong Nguyen
Imran Turganov
Minghao Wu
Peerat Limkonchotiwat
Irina Nikishina
We introduce HyperBrowseComp, a multilingual and multimodal browsing benchmark comprising 423 manually authored and human-validated question… (see more)s across 13 languages, written by native or highly proficient speakers. Questions are designed to be extremely challenging. Each question targets a concise, publicly verifiable answer whose discovery requires locating obscure evidence, following multi-step clue chains, or inspecting heterogeneous sources such as videos, scanned documents, images, or maps. Easier questions are filtered out by evaluating them with models without internet access to reduce the likelihood that they can be answered with parametric knowledge alone. We evaluate several models using provider-native search and a shared external retrieval harness under a common agent protocol. To contextualize model performance and effort, we also conduct a human evaluation on a sample of the questions. HyperBrowseComp provides a challenging testbed for persistent information seeking across languages and evidence modalities, with difficulty arising from discovering and connecting evidence on the open web.
Same Proteins, Different Verdicts: Toxicity-Classifier Clearance Is Not a Portable Safety Measure
Ananya Krishna
Bruce James Wittmann
DNA synthesis screening creates one of the few control points between a designed protein and its physical production. Deployed screens mainl… (see more)y search for similarity to catalogued sequences of concern. Generative methods can reduce that similarity while preserving predicted structural properties, so recent studies have also used academic toxicity classifiers to assess whether a redesign remains harmful. This practice assumes that classifier clearance is stable enough to support safety claims and comparisons among redesign methods. We test that assumption after an apparent success in our own protein redesign study failed to reproduce. We rescored the same historical sequences with a specified ToxinPred3 version and found that it labelled 94 of 100 sequences non-toxic for each of two redesign methods, removing the original separation. We then compared three evaluators on 176 unmodified proteins that CommEC flags. ToxinPred2 calls 156 toxic, whereas the published ToxinPred3 rule calls only 6 toxic. A fixed negative motif term changes 50 ToxinPred3 decisions. The disagreement also changes conclusions about redesign: OAE passes CommEC for 36 proteins, but random substitutions at the same positions pass 37 (McNemar p=1.0). Protein-language-model probes and sparse-autoencoder features do not provide an independent functional label. Toxicity-classifier clearance therefore depends on the evaluator, version, threshold, input type, and positive class. It does not constitute a general safety property of a redesign.
Selective Backpropagation for Efficient Few-Shot Class-Incremental Learning
Eeham Khan
Abdulmoumen Al-Atrash
Few-Shot Class-Incremental Learning (FSCIL) requires models to continuously learn new classes from limited samples while retaining prior kno… (see more)wledge, under strict constraints on compute and memory. Existing approaches lie along a difficult trade-off: simple fine-tuning is computationally efficient but suffers from catastrophic forgetting, replay-based methods mitigate forgetting at the cost of substantial compute and memory, and exemplar-free methods often reduce forgetting by freezing most of the backbone, improving efficiency at the expense of adaptability. We propose Selective Backpropagation (SBP), a deterministic parameter budgeting framework that bridges this gap. SBP restricts gradient updates to a pre-allocated subset of network parameters, freezing past knowledge and preserving unbiased capacity for future learning, enabling rapid adaptation without costly mask optimization. We show that SBP achieves strong performance across standard FSCIL benchmarks while requiring training time close to that of naive fine-tuning and substantially lower training time than prior SOTA methods. Crucially, our experiments expose a limitation of standard FSCIL evaluation: performance on short, distribution-consistent benchmarks does not necessarily predict behavior under distribution shift or over substantially longer learning horizons. We therefore evaluate FSCIL methods in cross-domain settings and over an 80-session ImageNet-1K stream. SBP remains strong across these regimes while maintaining low training cost, providing a favorable stability-plasticity-efficiency trade-off. Our code is available at https://github.com/PaInt-Lab/sbp-main-public.
Training LLMs to Verbalize Evaluation Awareness
Usman Anwar
Sahar Abdelnabi
David Krueger
Evaluation awareness (EA) can cause large language models (LLMs) to behave differently during audits than in deployment, yet measuring and a… (see more)ccounting for EA remains challenging. We introduce verbalization training (VT), a method for making LLMs less reticent about verbalizing evaluation awareness while avoiding to supervise the latent belief itself. VT uses a model's spontaneous verbalizations as evidence that awareness is present and truncates each rollout immediately before the verbalization, producing training prefixes at which the model is presumed to be aware. The model is then trained with an RL objective designed to increase verbalization in a calibrated way. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increases verbalized EA by 2.4-2.9 times and transfers to held-out agentic settings, while measured latent EA and behavior remain largely stable. In a causal experiment, we independently implant meta-knowledge about evaluations through synthetic-document fine-tuning and show that VT-induced verbalizations reflect the richer knowledge acquired by the model.
Accelerating Wildfire Hazard Mapping Across Canada with an AI Surrogate Modelling Framework
Hager Radi Abdelwahed
Tom Swystun
Sandy Erni
Francisco Emiliano Lopez Saavedra
Yan Boulanger
Canada's increasing wildfire activity requires frequent national hazard mapping to guide proactive mitigation, including fuel management, la… (see more)nd-use planning, infrastructure protection, and evacuation routing. However, process-based wildfire simulation frameworks (e.g., BurnP3 and BurnP3+) rely on computationally intensive Monte Carlo simulations to model burn probability, fire intensity, and rate of spread. These workflows can take days to weeks to process a single regional modelling unit. To overcome these computational bottlenecks, we present three contributions developed through collaboration between AI researchers and wildfire modelling experts: (a) a standardized framework for training, fine-tuning, and evaluating AI surrogate models against process-based wildfire simulators; (b) a multi-task deep learning surrogate trained on national-scale data to emulate burn probability, mean fire intensity, and mean rate of spread at a fraction of the computational cost (seconds); and (c) domain-specific metrics and counterfactual evaluation protocols to probe out-of-distribution generalization. Integrating geospatial data (fuel, terrain, ignition likelihood) and regional weather, our model approximates BurnP3+ simulation outputs, achieving strong performance across agreement, rank-correlation, error, and spatial-overlap metrics. It preserves downstream hazard classifications and produces physically plausible responses to expert-informed counterfactual interventions. Furthermore, we demonstrate zero-shot transfer to a separate regional BurnP3+ study and show that lightweight fine-tuning can substantially improve regional calibration and predictive agreement. On a representative region, the surrogate achieves 806-fold end-to-end CPU acceleration and uses 13% of BurnP3+'s host memory, facilitating rapid scenario analysis.
AgentInspect: Diagnosing Behavioral Failures in Artificial Intelligence Agents
Ruchira Manke
Mohammad Wardat
Hridesh Rajan
Effectively testing Artificial Intelligence (AI) agents remains a fundamental challenge due to their stochastic reasoning, vast and diverse … (see more)input space, reliance on external tools, and operation in dynamic execution environments; factors that demand new testing methodologies explicitly tailored to the complex and interactive nature of agent-based systems. This work presents a novel methodology for testing AI agents, with a particular focus on assessing their behavioral robustness under varied operational conditions. Our approach relies on following key technical innovations: (1) a coverage-guided test input generation strategy based on agent- specific coverage objectives, (2) a capture-and-simulate mechanism that systematically emulates abnormal tool behaviors to mimic real-world execution failures, and (3) a deterministic behavioral failure detection approach that enables consistent identification of failures across different test inputs. We developed AgentInspect, a framework that automatically detects six types of behavioral failures in LangChain-based AI agents by analyzing their execution trajectories across three evaluation settings: a baseline setting using real tool responses, a simulated setting incorporating synthetic tool responses, and a hybrid setting that combines the real and simulated tool responses. To evaluate our approach, we curated a benchmark of 35 AI agents obtained from GitHub. Our results show that AgentInspect consistently identifies different behavioral failures with high precision and recall across all three execution settings. In particular, the simulated and hybrid settings expose failure modes that do not emerge during baseline execution with real tool responses, thereby enabling a more comprehensive assessment of agent robustness. Our findings highlight AgentInspect’s effectiveness in revealing critical failures and its practical utility for systematic robustness evaluation of AI agents.
Climate vulnerability assessment framework choice influences species prioritization and prediction of population decline
Janaína de Andrade Serrano
Andrea Brown
Climate change vulnerability assessments (CCVAs) are regarded as an essential strategy to conservation strategies but are rarely used to inf… (see more)orm conservation planning and prioritize species for management and adaptation. The main factor inhibiting its use is the challenge of combining exposure, sensitivity, and adaptive-capacity variables and, especially, of evaluating the ability of frameworks to predict species decline. This raises a fundamental challenge for conservation science: can climate vulnerability reliably identify species of conservation concern before widespread declines occur? Using extensive long-term monitoring data for 236 breeding bird species in North America, we compared a Trait-based Vulnerability Index (TVI), which combines variables into a fixed-weight composite index, with a Climate Vulnerability Model (CVM) that estimates the relative contribution of the same variables statistically. The two approaches showed limited agreement in the species identified as highly vulnerable (35.6% overlap), indicating that vulnerability estimates differed substantially among assessment frameworks. This divergence varied by guild and biome, with the greatest disagreement between frameworks found for seabirds, and CVM outperforming TVI in most biomes except for grassland-associated species. Exposure variables, particularly projected habitat loss and cumulative threat burden, outperformed intrinsic traits in predicting both climate-change threat listings and observed population decline. Species assigned higher model-based climate vulnerability were significantly more likely to exhibit long-term population declines, whereas sensitivity and adaptive-capacity traits provided comparatively little additional predictive information. Our study provides a rare empirical evaluation of climate vulnerability assessments and their ability to identify species experiencing biological decline. Our results suggest that different vulnerability assessments can lead to different conservation priorities, indicating that additional work is needed to improve and validate climate-risk frameworks. The positive associations between vulnerability estimates and demographic decline indicate that large-scale assessments can provide useful information for evaluating risk and prioritizing conservation action where direct demographic data are limited.
A foundation for systematic analysis of transformers and RNNs for tractography
Emmanuelle Renauld
Philippe Poulin
Antoine Théberge
Maxime Descoteaux
Machine learning (ML) has emerged as a promising approach for improving diffusion MRI (dMRI) tractography, a task that remains limited by th… (see more)e intrinsic tension between local diffusion information and global anatomical plausibility. In this work, we systematically evaluate recurrent neural networks (RNNs) and Transformer models for iterative tractography, with particular attention to training strategies, input representations (including convolutional neural network (CNN)-based embeddings and end-of-sequence (EOS) tokens), and hyperparameter selection. We introduce a generation-validation phase enabling supervision at the streamline level during training, allowing supervision despite the mismatch between local loss functions and global streamline quality. Using the ISMRM2015 tractography challenge dataset, our models achieve the highest reported performance to date. Through controlled experiments, we quantify the impact of missing bundles, noisy or imperfect training streamlines, and invalid fibers in the training set. Finally, we demonstrate the applicability of our best-performing models for in vivo data from the Tractoinferno database. Overall, our results highlight both the potential and the limits of sequence-based deep learning models such as Transformers and RNNs for tractography, and emphasize the need for improved phantoms and evaluation methods for in vivo validation. We provide takeaways and recommendations for future researchers training and validating sequence-based supervised methods for tractography.
Improved mapping of Arctic fractional land cover and land cover change from multi-resolution optical remote sensing
Xiaoran Zhu
Jonathan A. Wang
Oliver Sonnentag
Isla H. Myers-Smith
Daryl Yang
Kathleen M. Orndahl
Leon Nill
Mark A. Friedl
Changes in Arctic tundra vegetation, driven by climate change, may be inducing major shifts in ecosystem services and the Arctic carbon budg… (see more)et, and altering high latitude feedbacks to the climate system. Field-based studies have documented warming-induced shrub expansion, and remote sensing has revealed heterogeneous, but primarily positive, trends in peak summer greenness across the Arctic. However, efforts to move beyond remotely sensed measures of spectral greening to quantify the spatial extent and rate of shrub expansion have been constrained by spectral similarities among tundra vegetation types, limited ground truth data, low revisit frequency of satellite observations, and sub-pixel heterogeneity of land cover at medium spatial resolution (30 m). To address these challenges, we developed a methodology that integrates high spatial resolution (2 m) commercial satellite imagery with Harmonized Landsat and Sentinel-2 observations in a machine learning framework, and used it to produce annual maps for 2016 to 2023 of sub-pixel land cover fractions at 30-m spatial resolution across three Arctic tundra ecoregions spanning 3.35 × 105 km2 between the Seward and Tuktoyaktuk Peninsulas. Uncertainty was quantified at each pixel via Monte Carlo resampling. Independent accuracy assessments yielded good accuracies (mean squared errors of 15.98% and 11.89% for low-stature vegetation and erect shrub cover, respectively), that were comparable to or exceeded previous mapping efforts. Further, repeat commercial satellite image pairs enabled the first assessment of mapped fractional cover change in Arctic tundra (R2 of 0.46 and 0.55, change direction accuracies of 77% and 78% for low-stature vegetation and erect shrub cover, respectively). This novel, scalable, multi-sensor approach to fractional land cover mapping produced the first annual maps of land cover fractions in the Arctic tundra, which support more accurate representation of vegetation dynamics and their linkages to climate change and disturbance processes.
An Optimal Design of Time-of-Use Tariff to Mitigate Overloads From Residential Electric Vehicle Charging on Distribution Networks
High penetration of electric vehicle (EV) charging creates equipment overload risks in power distribution networks, potentially reducing ass… (see more)et lifetime. This article develops an optimal time-of-use (TOU) tariff strategy to mitigate overloads by shifting residential EV charging demand, without requiring communication infrastructure or direct load control policies. We formulate a bi-level optimization problem to determine optimal rate levels that restore equipment lifetime while minimizing utility costs. The problem is reformulated as a convex optimization with proven convergence properties, enabling an efficient solution for large-scale networks. The solution guarantees that no new overload occurs after shifting the EV loads. A novel time-series analysis method is used to estimate loading impacts and to evaluate the effectiveness of candidate strategies. The proposed strategy has a simple TOU structure consistent with utilities' rate structure in place, making it suitable for adoption. We validate the approach on the IEEE 8500-node test feeder. Results demonstrate that at 80% EV penetration rate, the strategy reduces transformer overload peaks by 20%, while maintaining equipment's nominal lifetime and achieving 42.2% cost savings for utilities compared to early equipment replacement.
RETRACE: From Entangled Repair Histories to Reusable Experience for CI Repair
Rabeya Khatun Muna
Muhammad Ahasanuzzaman
Jinqiu Yang
Tse-Hsun Chen
Large language model (LLM) agents increasingly reuse prior experience, but most approaches assume that problems and solutions are already al… (see more)igned. Software histories rarely provide this alignment: a pull request (PR) may contain multiple continuous integration (CI) problems, failed attempts, reverted edits, and unrelated changes, obscuring which changes resolve each problem. We present RETRACE, a framework for reconstructing problem-level repair experience from such histories. RETRACE combines an endpoint view that reasons backward from changes retained in the passing revision with a development view that traces repair evolution forward through commit history. CI execution evidence reconciles the two views, and the recovered experience is represented at three abstraction levels, from concrete fixes to transferable repair patterns. For new failures, RETRACE retrieves relevant problem-level experience to guide repair. On CI-REPAIR-BENCH, comprising 565 PR-level repairs from 101 repositories across 12 failure categories, RETRACE improves mini-SWE-agent Pass@1 from 19.6% to 31.9% with MiniMax-M2.5 and from 23.3% to 32.8% with DeepSeek-V4-Flash. On a matched subset, Codex improves from 15.5% to 27.5%. Combining both views consistently outperforms either alone, showing that recovering problem-change alignment enables historical CI repairs to serve as reusable repair experience.