A weakly-supervised few-shot semantic segmentation model using heterogeneous visual aggregation.

Dynamic Multi-Scale Coordination Framework for Time Series Forecasting

Multi-agent framework for generating visually rich and fully editable methodology figures from scientific text.

Neural solvers that use PDE residual fields as input to iteratively correct predictions.

Benchmark and code for diagnosing visual spatial intelligence in VLMs.

A plug-and-play temporal wavelet digit interface for improving LLM time series forecasting.

dBASE III rebuilt in the browser with its own interpreter.

A contamination-aware multi-zone benchmark for evaluating knowledge boundaries in LLMs.

A framework to recover task-expert performance from a multi-task merged model.

A fast, lightweight AI safety guardrail based on a label-only encoder.

A language for defining access control policies used for formal verification of agent instructions.

Open-source architecture for hardware-aware EEG analysis and grounding.

A pipeline for detecting and steering away from sycophancy in LLMs via linear features.

A machine learning compilation framework for universal deployment of LLMs across platforms including Android.

Local-first desktop app to record work context and generate daily reports with local API and MCP support.

A silky smooth fiber runtime for ClickHouse to improve concurrency and performance.

Vulkan-based implementation of Direct3D 8/9/10/11 for Wine and Proton.

A searchable directory of 2,000+ self-hosted apps with license and activity data.

A QNX-inspired, dual kernel architecture open-source OS for RISC-V.

Empirical study on Mixture-of-Experts inference efficiency on consumer and edge hardware.