Knowledge-enhanced approach proposed for single-cell foundation models

A new arXiv paper examines how single-cell foundation models depend on large transcriptomic pretraining datasets, noting that adding more data brings diminishing returns at rising computational cost. The authors' data scaling analysis suggests incorporating structured biological knowledge could improve efficiency instead of relying on scale alone. The work points toward knowledge-enhanced pretraining as an alternative direction for the field.

arXiv paper argues foundation models should move toward open-ended discovery

A new arXiv preprint proposes "Discovery Foundation Models," framing open-ended discovery as the next stage for AI systems. The authors argue that models have moved from recalling and reasoning over existing knowledge to acting with tools and learning from outcomes, and that the next step is generating genuinely new findings. The paper is a position piece rather than an experimental release.

Language-Guided Multimodal Foundation Model Targets Brain Signal Analysis

Researchers posted a preprint describing a multimodal foundation model that uses language guidance to handle brain signal analysis without task-specific retraining. The approach is presented as supporting zero-shot and multi-task settings, addressing limited generalization in existing end-to-end and pre-trained models. The work appears on arXiv under cs.AI and cs.LG.

Adaptation Interfaces for In-Context Tabular Foundation Models in Time-to-Event Prediction

A new arXiv paper examines how tabular foundation models, which perform well on standard classification and regression tasks, can be adapted to time-to-event (survival) prediction. The authors propose adaptation interfaces to address the difficulty that censored outcomes pose for in-context learning. The work is listed under both machine learning and AI categories on arXiv.

CAL-MOS Uses Layer Adapters to Improve Speech Quality Prediction Across Foundation Models

A new arXiv paper introduces CAL-MOS, a method that uses adapters to combine representations from multiple layers of speech foundation models for non-intrusive speech quality assessment. The approach aims to make mean opinion score (MOS) prediction more robust when the underlying foundation model varies. The abstract notes that selecting which layer's representations to use remains an open question in this area.

EdgeHAR: Compact Sensor Foundation Model for On-Device Human Activity Recognition

Researchers present EdgeHAR, a compact foundation model for sensor-based human activity recognition that is built to run on edge devices rather than in the cloud. The work targets common real-world problems such as shifts in sensing conditions, including unfamiliar users and devices. It is described as an edge-native approach to wearable and ubiquitous computing.

EEG-Xplain framework targets interpretability of EEG foundation models

A new arXiv paper proposes EEG-Xplain, a unified attribution framework intended to make EEG foundation models such as BIOT, LaBraM, and EEGMamba more interpretable. The authors argue that the black-box nature of these models hinders clinical trust and neuroscientific validation. The work aims to provide a common approach for attributing model outputs to neural signal inputs.

Adaptive 3D-RoPE: Physics-Aligned Positional Encoding for Wireless Foundation Models

A revised arXiv paper proposes Adaptive 3D-RoPE, a rotary positional encoding scheme designed to match the physical structure of wireless channel data. The method targets wireless foundation models used for unified channel state information (CSI) acquisition in 6G networks, where such models already outperform task-specific baselines. The authors argue that aligning positional encoding with physics improves how these models generalize across tasks.

FlowTSFM: Turning Encoder Depth into Quantile Transport

A new arXiv preprint introduces FlowTSFM, an approach for encoder-based time series foundation models that assigns a predictive role to intermediate Transformer layers instead of supervising only the final forecast. The method recasts encoder depth as a form of quantile transport, according to the abstract. The announcement provides only the opening portion of the paper's abstract, so full details of the architecture and evaluation are not yet available in this report.

Tabby: Open Pretraining Recipe Released for Time Series Foundation Models

Researchers introduce Tabby, a long-context probabilistic foundation model designed for time series data, built on an encoder-only patch Transformer architecture. The release includes a fully open account of the pretraining process, covering the decisions and components behind the model's construction. The work aims to make time series foundation model development more reproducible and accessible.

Uncertainty Quantification Method Proposed for Earth System Foundation Models

A new arXiv preprint addresses how to quantify uncertainty in spatiotemporal foundation models used for Earth system modeling. The authors frame the problem around decision-making and risk control rather than pure prediction accuracy. The work aims to make such models more reliable when their outputs inform real-world choices.

S-CEReBrO Architecture Targets Memory Limits in Continuous EEG Monitoring

A new arXiv paper introduces S-CEReBrO, an approach aimed at overcoming the memory constraints that arise when applying Transformer-based foundation models to long-running EEG recordings. Global attention scales poorly with signal length, which limits how far such models can be used for continuous brain-monitoring data. The work proposes a redesign intended to make long-sequence EEG analysis more tractable while retaining the generalization benefits of large pretrained models.

Physically Typed and Geometry-Aware Representations for Earth Foundation Models

A new arXiv paper proposes combining semantic geospatial embeddings from Earth-observation foundation models with the spherical operators and Earth-specific geometry used in weather and climate modeling. The authors argue for representations that carry explicit physical types and geometry awareness, rather than purely semantic features, to better serve downstream Earth science tasks.

arXiv Paper Examines Whether Tabular Foundation Models Still Require Feature Engineering

A new arXiv preprint in machine learning asks whether manual feature engineering remains necessary now that tabular foundation models exist. These models are pretrained across many tabular datasets and applied through in-context learning, which may reduce reliance on hand-crafted features. The abstract frames the question as an open issue for tabular machine learning research.

Real-time foundation model for endoscopy supports task-specific fine-tuning

Researchers present woma, a foundation model trained without labels on roughly one million gastrointestinal endoscopy frames. Task-specific models are then fine-tuned from this base, and the authors describe a systematic design intended for production deployment, including requirements and performance targets. The work targets real-time use in clinical endoscopy workflows.

Paper examines distilling synthetic data for time series foundation models

A new arXiv preprint looks at how time series foundation models are pretrained on artificially generated trajectories where the underlying data-generating process is known. The work focuses on distillation methods rather than the conventional loss-based pretraining objectives that compare model outputs against targets. It aims to improve how these models learn from synthetic time series data.

MANAS-2: Constrained Reconstruction Approach for EEG Foundation Models

A new arXiv paper introduces MANAS-2, an EEG foundation model that reframes masked reconstruction as a constrained problem. The authors argue that optimizing reconstruction on low-SNR brain waveforms does not reliably yield the most useful latent representations. The model is presented as a step toward EEG pretraining objectives that better match downstream usefulness.

Dreaming in Code: Curriculum Learning for Open-Ended Worlds

A revised arXiv paper proposes a method for generating training environments as code, aiming to support curriculum learning in open-ended settings where agents face a continually expanding space of tasks. It builds on prior work that uses foundation models to programmatically produce diverse environments, extending it toward structured curricula. The work is a research preprint and has not been peer-reviewed.

Multimodal Foundation Model Pretrained for Lunar Remote Sensing

Researchers introduce a multimodal, multiresolution foundation model trained from scratch for lunar remote sensing. It was pretrained on SomBench, a geographically partitioned dataset of roughly two million co-registered tile bundles covering 11 sensor modalities at two spatial resolutions of 1 m/pixel. The work targets general-purpose representation learning for planetary surface analysis.

InfoAtlas: Foundation Model for Zero-Shot Statistical Dependence Estimation

A new arXiv paper introduces InfoAtlas, a foundation model designed to estimate statistical dependence between high-dimensional random variables without per-task training. The authors target the high computational cost of existing neural mutual information estimators, which usually rely on iterative optimization. The method aims to deliver zero-shot dependence estimates, potentially removing the need for task-specific tuning.

TechCrunch Disrupt 2026 session to focus on foundation model impact on AI startups

TechCrunch will host an interactive panel at its Disrupt 2026 conference aimed at founders of AI companies. The session centers on how startups can keep delivering value as foundation models keep improving. The event listing frames the core question as whether a builder's own roadmap can survive advances made by upstream model providers.

Can foundation models moderate online content? Comparing instruction- and example-driven policies

A new arXiv paper investigates whether foundation models can apply complex content moderation policies reliably and consistently. The study compares two ways of translating moderation rules into model behavior: conveying them through explicit instructions versus through illustrative examples. The findings are relevant to platforms seeking scalable, automated moderation of online content.

A Taxonomy of Architecture Options for Foundation Model-based Agents: Analysis and Decision Model

A paper on arXiv lays out a taxonomy of the main architecture options available for building agent systems powered by foundation models. It pairs this classification with a decision model designed to help practitioners analyze trade-offs and select appropriate designs when developing or operating such systems.

Infra-Bench CLS: open-source benchmark for critical infrastructure classification

A new arXiv paper introduces Infra-Bench CLS, a global, open-source benchmark for evaluating how well Earth observation foundation models can classify critical infrastructure. The work addresses the scarcity of complete infrastructure location data worldwide, a gap that is particularly acute in developing regions. The benchmark is intended to support more efficient mapping and monitoring of the built environment at a global scale.

Researchers probe whether speech foundation models truly learn words

A new arXiv study investigates self-supervised speech foundation models, which are widely deployed for speech recognition and to supply tokens for speech-capable language models. The authors analyze what these models' internal representations encode, testing whether they capture genuine word-level linguistic structure rather than only acoustic patterns. The results bear on how such models should be interpreted and used in downstream speech applications.

Time-Series Foundation Model Benchmarks Still Reflect Pretraining Familiarity on Later Hold-Outs

A new study questions whether time-series foundation models can be fairly evaluated using test data collected after their pretraining cutoff. It finds that even a temporally later, contamination-free hold-out does not fully isolate genuine generalization, as familiarity with the underlying data distribution absorbed during pretraining persists. The result suggests the field needs evaluation practices that go beyond simply withholding recent data.

Ling 2.0: open reasoning-focused language models scale to 1 trillion parameters

A new technical report introduces Ling 2.0, a family of reasoning-oriented foundation models built on a unified Mixture-of-Experts architecture that spans from tens of billions up to one trillion parameters. The series is released as an open language foundation, with the stated goal of strengthening general reasoning ability across all model sizes.

SIRF: Spec-Internalized Risk Foundation Model for Industrial Content Risk Control

Researchers introduce SIRF, a foundation model designed for content risk control in industrial settings. The work argues that deployment success depends less on average accuracy and more on how much risky content can be automatically handled while maintaining high precision and sub-second response times. The model internalizes specification rules rather than relying on external filtering logic.

arXiv Paper Explores Whether Foundation Models Can Reason About Topological Relations

A new preprint introduces MindTopo, a study on whether foundation models can handle topological spatial relations such as connectivity and containment, which stay stable under continuous deformation. The authors note that cognitive science treats these relations as a basic part of spatial understanding, distinct from metric cues like distance and angle. The work examines how well current models capture this kind of reasoning.

IBM releases Granite Time Series PatchTST-FM-r2 model with commercial-friendly license

IBM has published a revised edition of its Granite time series foundation model, PatchTST-FM-r2, on Hugging Face. Built on the Patch Time Series Transformer architecture, the model targets forecasting workloads and is offered under licensing terms that permit commercial use.

WHY IT MATTERS ↘A commercially licensed, openly available forecasting foundation model lowers the cost and legal friction of adopting time-series AI in production, an area where enterprises have mostly faced proprietary or research-restricted options. The rapid revision also signals that vendors are competing on maintained, enterprise-ready time-series models rather than one-off releases.