Publications

MosaicLeaks:Privacy Risks in Querying-in-the-Open for Deep Research Agents

Alexander Gurung

Spandana Gella

Alexandre Drouin

Issam H. Laradji

Perouz Taslakian

Rafael Pardinas

Deep research agents increasingly combine private local documents with external tools like web retrieval, creating a privacy risk: an agent'… (voir plus)s external queries may leak sensitive information from its local context. This risk is amplified by the mosaic effect, where individual queries may appear harmless but become revealing in aggregate. We introduce MosaicLeaks, a benchmark of 1,001 multi-hop deep research tasks that chain private enterprise documents and a public web corpus, forcing agents to make external queries that depend on local information. We evaluate leakage with an adversary LLM that observes only the agent's external queries and attempts to infer private information at three levels: the agent's research intent, answers to specific private questions and verifiable claims about the enterprise documents. We find that models across families and sizes frequently leak at all three levels, that zero-shot privacy prompting reduces but does not eliminate leakage and that reinforcement learning for task performance alone worsens leakage. To address this, we propose Privacy-Aware Deep Research (PA-DR), an RL framework that combines situational rewards for task success with a learned privacy classifier to provide dense credit assignment over both per-query and mosaic-level leakage. Training Qwen3-4B-Instruct with PA-DR improves accuracy from 48.7% to 58.7% and reduces answer and full-information leakage from 34.0% to 9.9%.

2026-05-28

arXiv (prépublication)

Orth-Dion: Eliminating Geometric Mismatch in Distributed Low-Rank Spectral Optimization

Tatsuhiro Nakamori

Laura Gomezjurado Gonzalez

Ganesh Talluri

Ansh Tiwari

Hideyuki Kawashima

Ioannis Mitliagkas

Guillaume Rabusseau

Hiroki Naganuma

Low-rank gradient compression reduces communication in distributed training by representing updates with rank-…

2026-05-28

CoLoRAI @ International Conference on Machine Learning (poster)

Quantitative Equational Logic

Giorgio Bacci

Radu Mardare

Prakash Panangaden

Gordon Plotkin

We develop a quantitative analogue of equational reasoning, which we call quantitative equational logic. The quantitative equations use, ins… (voir plus)tead of classical equality, quantitative equalities, which are equalities indexed with nonnegative reals. Thus, s = ε t means that “ s and t are points in a metric space and their distance is less than ε”. Quantitative equalities will be used to encode behavioural distances, with ε being an upper bound on the measure of dissimilarity between two terms. We develop the metatheory of this subject. We define a notion of quantitative algebra, which is the quantitative analogue of universal algebra. We prove a completeness theorem for quantitative equational logic, and we show that we obtain monads on suitable categories of metric spaces. We present a set of examples where the free algebra of a quantitative equational theory corresponds to some well-known structure. These examples are: Hausdorff metrics from quantitative semilattices; p -Wasserstein metrics (hence also the Kantorovich metric), and the total variation metric.

2026-05-28

Journal of the ACM (publié)

Reparametrizing Shampoo and SOAP for Subspace Basis Updates and BFloat16 Storage

Alan Milligan

Zikun Xu

Simon Lacoste-Julien

Felix Dangel

Wu Lin

Shampoo-based methods, such as KL-Shampoo and SOAP, have demonstrated strong performance in training neural networks and rely on QR decompos… (voir plus)ition. Because existing QR implementations require single-precision (FP32) arithmetic and remain computationally expensive, these methods become time- and memory-intensive when their preconditioning matrices are large. Moreover, using BFloat16 (BFP16) storage to reduce memory usage can degrade the performance of Shampoo-based methods. We propose a reparametrization of the preconditioner that supports BFP16 storage and forms a complete basis by combining updated basis vectors with unchanged ones. By updating only part of the basis through QR decomposition in a subspace, our approach reduces computational overhead while mitigating the performance degradation caused by BFP16 storage. Our approach applies broadly to Shampoo-based methods that employ QR decomposition, including KL-Shampoo, SOAP, and KL-SOAP. In particular, it improves the performance of SOAP and KL-SOAP under BFP16 storage, enabling KL-SOAP to match or exceed KL-Shampoo. Overall, our approach makes Shampoo-based methods more memory- and time-efficient.

2026-05-28

HiLD @ International Conference on Machine Learning (poster)

Reusable Low-Rank Subspaces Explain Why Cross-Modal Transfer Adapts with Tiny Updates

Zhenghan Tai

Vasilii Feofanov

Parameter-efficient finetuning methods such as LoRA routinely adapt massive pretrained transformers to new tasks using only tiny low-rank up… (voir plus)dates, but the representational geometry that makes this possible remains unclear. We use cross-modal transfer from a language-pretrained transformer to time-series forecasting as a controlled probe of low-rank adaptation, asking why so few directions are sufficient. Across adaptation regimes, LoRA recovers most of the transfer benefit of full finetuning; effective-rank analyses show that pretrained representations already concentrate on a low-rank subspace that finetuning \emph{redistributes} rather than rebuilds; and a single linear projection over frozen hidden states aligns with realistic time-series trajectories without paired supervision. Randomly initialized models, by contrast, first construct a compressed representation through a uniform layer-wise collapse before they can specialize. These results support a view of cross-modal adaptation as low-rank \emph{direction selection} within reusable pretrained subspaces.

2026-05-28

CoLoRAI @ International Conference on Machine Learning (poster)

Beyond Go/No-Go Decisions: A Regional Selection Framework for Uncertainty-Aware Molecule Screening

Weihua Shi

Yixuan Li

Tian Bai

Yijie Zhang

Kaiqiong Zhao

Marc‐André Legault

Hui Peng

Yue Zhao

Eric D. Kolaczyk

Xiang Yu

Archer Y. Yang

In drug discovery, quantitative structure–activity relationship (QSAR) models are widely used to guide Go/No-Go decisions within the Desig… (voir plus)n–Make–Test–Analyze (DMTA) cycle. However, conventional decision heuristics typically rely on a single cutoff, leading to a rigid binary select/discard paradigm. This approach is particularly ill-suited for borderline compounds near the decision boundary, where screening decisions are especially sensitive to prediction uncertainty and premature choices may either discard viable leads or advance likely failures, thereby increasing downstream assay costs. To address this limitation, we propose Regional Selection (RS), an uncertainty-aware three-way decision framework that partitions compounds into Predicted Pass, Predicted Fail, and Predicted Indeterminate regions. By explicitly reserving high-uncertainty compounds for targeted follow-up, RS avoids the pitfalls of premature binary classification. We formalize this framework through Regional Selection Inference (RSI), which casts region assignment as a multiple-hypothesis testing problem. We develop two imple- mentations of RSI: an empirical calibration-based method (RSI-EC), which thresholds uncertainty-normalized scores via empirical calibration, and a conformal selectionbased method (RSI-CS), which constructs conformal p-values for region assignment. RSI-EC is supported by large-sample calibration arguments, whereas RSI-CS provides finite-sample, distribution-free guarantees under exchangeability. Extensive evaluations across 15 high-dimensional QSAR benchmarks show that both RSI procedures reliably control the false discovery rate while maintaining high screening power. In limited-data regimes, RSI-CS yields particularly stable FDR control, whereas RSI-EC can be slightly less conservative; both perform strongly as sample sizes increase. We further study a cost-aware extension that incorporates asymmetric downstream costs through the score construction while keeping the nominal FDR target fixed. This extension introduces a tuning parameter that can reduce realized downstream cost, with dataset-dependent trade-offs against screening power. Overall, RSI offers a mathematically grounded and resource-aware alternative to single-threshold screening, allowing discovery teams to better balance decision confidence with assay budgets.

2026-05-27

ChemRxiv (accepté)

Croissant Tasks: A Metadata Format for Reproducible Machine Learning Evaluations

Omar Benjelloun

Leonardo Martins Bianco

Isabelle Guyon

Thanh Gia Hieu Khuong

Jonathan Lebensold

Sebastian Lobentanzer

Luis Oala

Benedictus Kent Rachmat

Ihsan Ullah

Peyman Vahidi

Joaquin Vanschoren

Reproducibility is fundamental to the scientific method, yet remains a critical challenge in machine learning. Contributing factors include … (voir plus)underspecified execution details and brittle software environments. Human-centric remedies, such as checklists and manual verification, help but require intensive effort and fail to scale. To address this, we introduce Croissant Tasks: a declarative, machine-actionable metadata format that abstracts low-level implementation details into high-level specifications. This format enables conceptual reproducibility: verifying claims via independent, agent-generated implementations rather than brittle source code replication. We contribute: (1) the Croissant Tasks specification, formally decoupling task problem from solution; (2) an automated LLM pipeline that retrofits existing benchmarks into this format; and (3) empirical validation showing autonomous agents can ingest these specifications to generate functional, accurate reproduction pipelines from scratch. We envision this format as a new foundation for automated and conceptual reproducibility in machine learning.

2026-05-27

arXiv (prépublication)

Does The Way You Plan Matter? An Empirical Study of Planning Representations for LLM Web Agents

Leila Kosseim

Despite recent advances, LLM-based web agents still struggle with limited exploration, omission of critical steps, and sensitivity to task c… (voir plus)onstraints. Prior work suggests that many of these failures stem from weaknesses in planning, yet the impact of alternative natural language plan representation remains unexplored. To address this, we introduce PlanAhead, a static planner-executor framework that evaluates the impact of plan representation in agent performance. We first automatically categorize WebArena tasks into 3 difficulty levels, enabling consistent difficulty grading without human annotation. Then we systematically evaluate 4 different plan representations on the tasks categorized as hard: sequential subgoals, narrative, pseudocode, and checklist; across different families of multimodal LLM powered agents (OpenAI, Alibaba, and Google). To account for stochastic variability, we introduce two novel evaluation metrics: Achievement Rate (AR) and Solved-Task Consistency (STC). Our results show that both, the plan formulation and the underlying LLM generating the plan, significantly influence web-agent robustness and task success.

2026-05-27

arXiv (prépublication)

EASE Configuration Facilitates A Reproducible Science of LLM Social Simulations

Sneheel Sarangi

Maximilian Puelma Touzel

Aurélien Bück-Kaeffer

Zachary Yang

Jean-François Godbout

Reihaneh Rabbany

LLMs are increasingly deployed to simulate social interactions, yet many of the existing simulators remain ad hoc and monolithic. This lack … (voir plus)of architectural standardization prevents reproducible research and complicates downstream evaluation. We advance a rigorous science of LLM-based multi-agent simulation by modularizing core components into Environments, Agents, Simulation engines, and Evaluation metrics (EASE). We demonstrate the utility of EASE configuration by wrapping it in an experimental study schema for orchestrating workflows centered around answering explicit research questions in generated scenarios. We contribute SiliSocS, an open-source, research-ready Silicon Society Sandbox implementing a study-structured EASE configuration to enable highly configurable and reproducible LLM-based social simulations. Using SiliSocS and EASE, we present three case studies, showcasing the system's comprehensive assessment of existing questions, ability to dive deeper into complex questions, and elaboration of existing studies, respectively. Together, these case studies highlight the limitations of current modeling approaches and isolate the impacts of design choices on key results.

2026-05-27

arXiv (prépublication)

GeneZip: Region-Aware Compression for Long Context DNA Modeling

Hongyu Guo

Genomic sequences span billions of base pairs (bp), posing a fundamental challenge for genome-scale foundation models. Existing approaches l… (voir plus)argely sidestep this barrier by either scaling relatively small models to long contexts or relying on heavy multi-GPU parallelism. Here we introduce GeneZip, a DNA compression model that leverages a key biological prior: genomic information is highly imbalanced. Coding regions comprise only a small fraction (about 2 percent) yet are information-dense, whereas most non-coding sequence is comparatively information-sparse. GeneZip couples HNet-style dynamic routing with a region-aware compression-ratio objective, enabling adaptive allocation of representation budget across genomic regions. As a result, GeneZip learns region-aware compression and achieves 137.6x compression with only 0.31 perplexity increase. On downstream long-context benchmarks, GeneZip achieves comparable or better performance on contact map prediction, expression quantitative trait loci prediction, and enhancer-target gene prediction. By reducing effective sequence length, GeneZip unlocks simultaneous scaling of context and capacity: compared to the prior state-of-the-art model JanusDNA, it enables training models 82.6x larger at 1M-bp context, supporting a 636M-parameter GeneZip model at 1M-bp context. All experiments in this paper can be trained on a single A100 80GB GPU.

2026-05-27

FM4LS @ International Conference on Machine Learning (poster)

GFETM: Genome Foundation-based Embedded Topic Model for scATAC-seq Modeling

Yimin Fan

Yu Li

Jun Ding

Yue Li

Single-cell Assay for Transposase-Accessible Chromatin with sequencing (scATAC-seq) enables investigation of open chromatin landscapes at si… (voir plus)ngle-cell resolution, but its analysis remains challenging because of sparsity, noise, and dataset-specific peak vocabularies. Genome Foundation Models (GFMs), pre-trained on large DNA sequence corpora, offer a potential source of transferable sequence information for scATAC-seq modeling. We introduce the Genome Foundation Embedded Topic Model (\model{}), an interpretable framework that combines GFMs with the Embedded Topic Model (ETM) for sequence-informed scATAC-seq analysis. By integrating GFM-derived DNA sequence embeddings into a topic-model decoder, \model{} improves clustering quality on standard benchmarks and captures cell-state-specific transcription factor activity through motif scoring and attention-based interpretation.

2026-05-27

FM4LS @ International Conference on Machine Learning (poster)