scalable

HFlow is a scalable multimodal data pipeline tool for robotics applications, recently featured on Product Hunt.

Presents STCAD, a scalable framework using BERT-based encoding and CURE clustering to perform trajectory clustering and anomaly detection on terabyte-scale AIS maritime data, demonstrating stable clusters and clear separation of anomalous vessel behavior.

This paper introduces GRACE, a framework that uses LLM-generated semantic descriptions at the attribute-value level to create unified metric spaces for clustering mixed tabular data, achieving scalability comparable to statistical baselines while improving clustering accuracy.

Proposes a scalable subdocument deduplication framework for LLM pretraining that separates duplicate detection from copy retention, using frequency- and length-aware policies. Experiments on FineWeb-Edu and a code web corpus show improved model performance.

DecoupleMix introduces a systematic framework for optimizing pretraining data mixtures for Vision-Language Models by decoupling inter-class and intra-class ratio search, using convex optimization to improve scalability and performance over heuristic baselines.

Molt is a PyTorch-native training framework for agentic reinforcement learning designed to be compact and clean for easy modification, while achieving performance comparable to Megatron-based stacks.

This paper proposes a generative AI-assisted summarization framework using GPT-5 model variants to address input-length limitations in automated essay scoring, demonstrating trade-offs between model capacity, summary fidelity, and computational cost on the ASAP 2.0 dataset.

Proposes a perturbation-based learning rule for online self-supervised learning in echo state networks that avoids reservoir-size-dependent variance growth by decomposing the learning cost and perturbing only the input-dependent component.

This paper proposes SCISE, a scalable unsupervised graph clustering framework that uses community-aware sampling and structural entropy to overcome structural isolation in mini-batch training, achieving state-of-the-art results on benchmark datasets.

This paper proposes single-prover interactive proofs for AI safety verification, avoiding the need for debate between two competing models, and extends the approach to oracle-aided computations.

Memora is a scalable memory system for AI agents that decouples storage from retrieval, enabling long-horizon tasks with up to 98% fewer context tokens while setting new state-of-the-art on benchmarks. The paper is published at ICML 2026.

DataStates-LLM introduces a scalable checkpointing architecture for transformer models using composable state providers, achieving up to 4x higher throughput and reducing training time by 2.2x compared to existing solutions.

This study suggests that AI can make expert-led periodic reanalysis of old medical cases more scalable, helping clinicians revisit cases as medical knowledge advances and potentially bring answers to more cases that previously evaded analysis.

μ_0 is a scalable world model that predicts smooth 3D trajectories for interaction points, enabling embodiment-agnostic robot learning without action labels by using a TraceExtract system for supervision.

PatchSTG introduces a patch-based spatiotemporal graph Transformer for traffic forecasting on irregular sensor networks, achieving near-linear complexity while maintaining competitive performance.

GRASP introduces a geometry-aware, interaction-based method for scalable pretraining data attribution that models subset dynamics, outperforming existing additive approaches by over double the task-level rank correlation while reducing computation costs.

mimalloc is an open-source, high-performance, scalable memory allocator that serves as a drop-in replacement for malloc and free. Designed for modern highly concurrent applications and large memory scales, it is used in major services like Bing and integrated into projects such as NoGIL CPython and Unreal Engine.

Browser Use launches a new browser infrastructure service featuring subsecond cold starts, lower cost at $0.02/h, and unlimited scaling, now live for developers.

This paper presents a distributed approach for constrained multi-agent reinforcement learning that uses state-augmented policy learning and neighbor-to-neighbor consensus over dual variables to satisfy global resource constraints while scaling linearly with the number of agents. Experiments on smart grid demand response demonstrate that consensus coordination is essential for feasibility, scaling to thousands of agents unlike centralized training approaches.

This paper introduces the concept of Access Sets to budget expert reads, enabling scalable weight-space model merging.