Publications

Multi-Modal Learning meets Genetic Programming: Analyzing Alignment in Latent Space Optimization
Benjamin Léger
Kazem Meidani
Christian Gagné
Symbolic regression (SR) aims to discover mathematical expressions from data, a task traditionally tackled using Genetic Programming (GP) th… (voir plus)rough combinatorial search over symbolic structures. Latent Space Optimization (LSO) methods use neural encoders to map symbolic expressions into continuous spaces, transforming the combinatorial search into continuous optimization. SNIP (Meidani et al., 2024), a contrastive pre-training model inspired by CLIP, advances LSO by introducing a multi-modal approach: aligning symbolic and numeric encoders in a shared latent space to learn the phenotype-genotype mapping, enabling optimization in the numeric space to implicitly guide symbolic search. However, this relies on fine-grained cross-modal alignment, whereas literature on similar models like CLIP reveals that such an alignment is typically coarse-grained. In this paper, we investigate whether SNIP delivers on its promise of effective bi-modal optimization for SR. Our experiments show that: (1) cross-modal alignment does not improve during optimization, even as fitness increases, and (2) the alignment learned by SNIP is too coarse to efficiently conduct principled search in the symbolic space. These findings reveal that while multi-modal LSO holds significant potential for SR, effective alignment-guided optimization remains unrealized in practice, highlighting fine-grained alignment as a critical direction for future work.
2026-04-08
arXiv (prépublication)
doi.org
arxiv.org
Sequence Generation and Phylogenetic Inference with Generative Flow Networks
Qichen Huang
Carlos Michel Mourra-Díaz
Xiaozhen Wen
D. Payette
A bstract Phylogenetic inference remains computationally challenging due to the exponentially growing tree topology search space, and curren… (voir plus)t methods rely heavily on multiple sequence alignments (MSAs) which are expensive and error-prone. We propose AncestorGFN, a proof-of-concept approach leveraging Generative Flow Networks (GFlowNets) for simultaneous sequence generation and phylogenetic exploration without requiring explicit MSAs. Our method learns to generate sequences matching a target distribution while the flow trajectories implicitly encode structural relationships among sequences. We demonstrate that greedy traceback on maximum-flow trajectories recovers shared intermediate states suggestive of common ancestry, and evaluate on the let-7 microRNA family where the learned flow structure qualitatively captures phylogenetic branching patterns. Furthermore, beam search at inference time discovers novel sequences clustering near known targets, suggesting applications in de novo sequence design. This work establishes an initial foundation for alignment-free phylogenetic exploration using generative models.
2026-04-08
bioRxiv (accepté)
doi.org
Toward Hardware-Agnostic Quadrupedal World Models via Morphology Conditioning
Mohamad H. Danesh
Chenhao Li
Amin Abyaneh
Anas Houssaini
Kirsty Ellis
Glen Berseth
Marco Hutter
Hsiu-Chin Lin
World models promise a paradigm shift in robotics, where an agent learns the underlying physics of its environment once to enable efficient … (voir plus)planning and behavior learning. However, current world models are often hardware-locked specialists: a model trained on a Boston Dynamics Spot robot fails catastrophically on a Unitree Go1 due to the mismatch in kinematic and dynamic properties, as the model overfits to specific embodiment constraints rather than capturing the universal locomotion dynamics. Consequently, a slight change in actuator dynamics or limb length necessitates training a new model from scratch. In this work, we take a step towards a framework for training a generalizable Quadrupedal World Model (QWM) that disentangles environmental dynamics from robot morphology. We address the limitations of implicit system identification, where treating static physical properties (like mass or limb length) as latent variables to be inferred from motion history creates an adaptation lag that can compromise zero-shot safety and efficiency. Instead, we explicitly condition the generative dynamics on the robot's engineering specifications. By integrating a physical morphology encoder and a reward normalizer, we enable the model to serve as a neural simulator capable of generalizing across morphologies. This capability unlocks zero-shot control across a range of embodiments. We introduce, for the first time, a world model that enables zero-shot generalization to new morphologies for locomotion. While we carefully study the limitations of our method, QWM operates as a distribution-bounded interpolator within the quadrupedal morphology family rather than a universal physics engine, this work represents a significant step toward morphology-conditioned world models for legged locomotion.
2026-04-08
arXiv (prépublication)
doi.org
arxiv.org
Epistemic Robust Offline Reinforcement Learning
Abhilash Reddy Chenreddy
Erick Delage
Offline reinforcement learning learns policies from fixed datasets without further environment interaction. A key challenge in this setting … (voir plus)is epistemic uncertainty, arising from limited or biased data coverage, particularly when the behavior policy systematically avoids certain actions. This can lead to inaccurate value estimates and unreliable generalization. Ensemble-based methods like SAC-N mitigate this by conservatively estimating Q-values using the ensemble minimum, but they require large ensembles and often conflate epistemic with aleatoric uncertainty. To address these limitations, we propose a unified and generalizable framework that replaces discrete ensembles with compact uncertainty sets over Q-values. %We further introduce an Epinet based model that directly shapes the uncertainty sets to optimize the cumulative reward under the robust Bellman objective without relying on ensembles. We also introduce a benchmark for evaluating offline RL algorithms under risk-sensitive behavior policies, and demonstrate that our method achieves improved robustness and generalization over ensemble-based baselines across both tabular and continuous state domains.
2026-04-07
arXiv (prépublication)
doi.org
arxiv.org
LSST Strong Lensing Systems Dark Matter Sensitivity Analysis with Neural Ratio Estimators
Andreas Filipp
Yashar Hezaveh
Laurence Perreault-Levasseur
Daniel Gilman
LSST Dark Energy Science Collaboration
Strong gravitational lensing offers a unique probe of dark matter (DM) on sub-galactic scales, where the abundance and distribution of low-m… (voir plus)ass halos are highly sensitive to the underlying properties of DM particles. In this work, we forecast LSST's sensitivity to DM substructure in galaxy-galaxy strong lenses using simulated samples and neural ratio estimators (NREs). Our simulations include both subhalos within the main deflector and line-of-sight (LOS) halos, with halo masses down to
2026-04-07
arXiv (prépublication)
doi.org
arxiv.org
Robust Mendelian Randomization Estimation using Weighted Quantile Regression
Julien St-Pierre
Archer Y. Yang
Mireille E. Schnitzer
Marc-André Legault
In Mendelian randomization (MR) studies, genetic variants are used as instrumental variables (IVs) to investigate causal relationships betwe… (voir plus)en exposures and outcomes based on observational data. However, numerous genetic studies have shown the pervasive pleiotropy of genetic variants, meaning that many, if not most, variants are associated with multiple traits, potentially violating the core assumptions of IV estimation. Uncorrelated pleiotropy occurs when genetic variants have a direct effect on the outcome that is not mediated by the exposure, while correlated pleiotropy occurs when genetic variants affect the exposure and outcome via shared heritable confounders. In this work, we propose a novel MR method, called MR-Quantile, based on weighted quantile regression (WQR) that is robust to both correlated and uncorrelated pleiotropy. We propose a procedure for selecting the optimal quantile of the ratio estimates through a likelihood-based formulation of WQR using the asymmetric Laplace distribution. Monte Carlo simulations demonstrate the empirical performance of the proposed method, especially in settings with many invalid IVs with weak pleiotropic effects. Finally, we apply our method to study the causal effect of resting heart rate on atrial fibrillation. Genetic variants associated with heart rate were identified in a genome-wide association study of 425,748 individuals from the VA Million Veteran Program, and used as instruments in a two-sample MR analysis with summary statistics from a genetic meta-analysis of 228,926 AF cases across eight studies.
2026-04-07
arXiv (prépublication)
doi.org
arxiv.org
The Illusion of Stochasticity in LLMs
Xiangming Gu
Soham De
Michalis K. Titsias
Larisa Markeeva
Petar Veličković
Razvan Pascanu
In this work, we demonstrate that reliable stochastic sampling is a fundamental yet unfulfilled requirement for Large Language Models (LLMs)… (voir plus) operating as agents. Agentic systems are frequently required to sample from distributions, often inferred from observed data, a process which needs to be emulated by the LLM. This leads to a distinct failure point: while standard RL agents rely on external sampling mechanisms, LLMs fail to map their internal probability estimates to their stochastic outputs. Through rigorous empirical analysis across multiple model families, model sizes, prompting styles, and distributions, we demonstrate the extent of this failure. Crucially, we show that while powerful frontier models can convert provided random seeds to target distributions, their ability to sample directly from specific distributions is fundamentally flawed.
2026-04-07
arXiv (prépublication)
doi.org
arxiv.org
Uncertainty Assessment in Deep Learning-based Plant Trait Retrievals from Hyperspectral data
Eya Cherif
Teja Kattenborn
Luke A. Brown
Michael Ewald
Katja Berger
Phuong D. Dao
Tobias B. Hank
Étienne Laliberté
Bing Lu
Hannes Feilhauer
Abstract. Large-scale mapping of plant biophysical and biochemical traits is essential for ecological and environmental applications. Given … (voir plus)their finer spectral resolution and unprecedented data availability, hyperspectral data, in concert with machine and particularly deep learning models, have emerged as a promising, non-destructive tool for accurately retrieving these traits. However, when deploying these methods on a large scale, reliably quantifying the associated uncertainty remains a critical challenge, especially when models encounter out-of-domain (OOD) data, i.e., samples that differ substantially from those of the training data, such as unseen geographical regions, species, biomes, data acquisition modalities, or scene components (e.g., clouds and water bodies). Traditional uncertainty quantification methods for deep learning models, including deep ensembles (deterministic and probabilistic) and Monte Carlo dropout, rely on the variance of predictions but often fail to capture uncertainty in OOD scenarios, leading to overly optimistic and possibly misleading uncertainty estimates. To address this limitation, we propose a distance-based uncertainty estimation method (Dis_UN) that quantifies prediction uncertainty by measuring the dissimilarity in the predictor space (spectral inputs) and embedding space (features learned by the deep model) between the training and test data. Dis_UN leverages residuals as a proxy for uncertainty and employs dissimilarity indices in data manifolds to estimate worst-case errors via 95-quantile regression. We evaluate Dis_UN using a pretrained deep learning model to predict multiple plant traits from hyperspectral images, analyzing its performance across OOD data, such as pixels containing spectral variations from urban surfaces, bare ground, water, clouds, or open surface waters. In this study, we target six leaf and canopy traits: leaf mass per area, chlorophylls, carotenoids, nitrogen content, equivalent water thickness, and leaf area index. Compared to scaled variance-based methods, Dis_UN provides (1) a superior estimation of uncertainty in OOD scenarios, achieving 36 % higher contrast (KS distances: 0.648 vs. 0.475) between non-vegetation pixels, particularly under mixed-pixel conditions at medium resolution (30 m); (2) uncertainty quantification without requiring normality or symmetry assumptions, accommodating asymmetric error patterns; (3) enhanced interpretability of uncertainty sources, as uncertainty is directly linked to sample dissimilarity from the training data; and (4) computational efficiency at inference (2.6–7.7× faster), requiring only a single forward pass compared to multiple passes for ensemble-based methods. Challenges remain for traits that are affected by spectral saturation. These findings highlight the advantages of distance-aware uncertainty quantification methods and underscore the necessity of diverse training datasets to minimize sampling biases and enhance model robustness. The proposed framework improves the reliability of uncertainty estimation in vegetation monitoring and offers a promising approach for broader applications.
2026-04-07
Biogeosciences (publié)
doi.org
Evaluation of data-driven kinematic models for autonomous control of continuum robotic in-situ bioprinters
Swen A.T. Groen
David Brenken
Samuel Smocot
James Richard Forbes
Luc Mongeau
Audrey Sedal
Minimally invasive in-situ bioprinting involves the direct deposition of hydrogels within the body to reconstruct tissue defects. These biop… (voir plus)rinters use soft robotic printheads to extrude hydrogels through a hollow channel and nozzle. The accurate control of the nozzle tip position is critical for safety and shape fidelity. Due to a lack of sensing integration, existing control strategies are limited to feedforward models and are design-specific, thereby increasing development cost and complexity. This present study systematically compared three different data-driven modeling strategies for autonomous control of a cable-driven continuum in-situ bioprinter: 1) Polynomial regression, 2) Gaussian process regression, and 3) a neural network. Submillimeter accuracy was achieved for both the Gaussian process and the neural network in static measurements. The Polynomial regression model had a 1.67 mm accuracy. Dynamic trajectory tracking indicated that the performance of the neural network was comparable to that of the polynomial regression model and lower than that of the Gaussian process. Printing of different shaped constructs yielded minor visual deviations across all models from the target shapes. These results support the feasibility of real-time autonomous control for minimally invasive in-situ bioprinters and indicate the advantages of the different models during the design process of novel printing strategies.
2026-04-06
IEEE International Conference on Soft Robotics (publié)
doi.org
Observability-Informed Optimal Sensor Placement for Soft Robots
Samuel Smocot
James R. Forbes
Audrey Sedal
This paper presents the application and experimental evaluation of a systematic method for optimal sensor placement in soft robots. Existing… (voir plus) methods either lack generalizability across different soft robot morphologies or do not account for system dynamics. The applied method uses convex optimization to find the optimal sensor configuration that maximizes an observability Gramian-based metric. The framework is experimentally evaluated using position and strain measurements on a soft continuum arm. Kalman filter state estimates using optimal sensor placements yield lower reconstruction error than a baseline across all sinusoidal input trials, with improvements on the order of millimeters. This case study shows that linear control theory tools can guide optimal sensor placement in soft robots, suggesting an interpretable approach to sensor placement that may extend to other morphologies.
2026-04-06
IEEE International Conference on Soft Robotics (publié)
doi.org
Programmable Membrane Shape via Localized Stiffening of a Granular Suspension
Omar Khater
Karim Saliba
Audrey Sedal
Granular actuators are often implemented with as many vacuum pumps as jamming segments. For highly articulated motion, these segments increa… (voir plus)se mechanical design complexity. In this work, we propose cooperative use of pumps in a monolithic liquid-granular suspension to locally control jamming behavior. As an example of a granular soft robot, we change the shape of a buckled membrane by varying the local concentration of a particle-water suspension. We then develop and validate a numerical model based on localized stiffening to explain the underlying mechanisms behind membrane motion. We finally demonstrate that this scheme enables actuation programmed by pump sequences. This novel actuation may further enable localized jamming in arbitrary internal structures, ultimately leading to simple soft robots with programmable articulated motion.
2026-04-06
IEEE International Conference on Soft Robotics (publié)
doi.org
Understanding Performance Gap Between Parallel and Sequential Sampling in Large Reasoning Models
Xiangming Gu
Soham De
Larisa Markeeva
Petar Veličković
Razvan Pascanu
Large Reasoning Models (LRMs) have shown remarkable performance on challenging questions, such as math and coding. However, to obtain a high… (voir plus) quality solution, one may need to sample more than once. In principal, there are two sampling strategies that can be composed to form more complex processes: sequential sampling and parallel sampling. In this paper, we first compare these two approaches with rigor, and observe, aligned with previous works, that parallel sampling seems to outperform sequential sampling even though the latter should have more representation power. To understand the underline reasons, we make three hypothesis on the reason behind this behavior: (i) parallel sampling outperforms due to the aggregator operator; (ii) sequential sampling is harmed by needing to use longer contexts; (iii) sequential sampling leads to less exploration due to conditioning on previous answers. The empirical evidence on various model families and sizes (Qwen3, DeepSeek-R1 distilled models, Gemini 2.5) and question domains (math and coding) suggests that the aggregation and context length do not seem to be the main culprit behind the performance gap. In contrast, the lack of exploration seems to play a considerably larger role, and we argue that this is one main cause for the performance gap.
2026-04-06
arXiv (prépublication)
doi.org
arxiv.org