A new benchmark stress-tests vision-language segmentation models against rain, fog, snow, and night, exposing systematic failures tied to weather type and severity.
A new benchmark tests whether frontier models can spot dark jokes across English, Arabic, images, and video, and the results expose a cultural reasoning gap.
A new label-free RL method trains a single model as both generator and verifier, boosting math reasoning by 4-6% while doubling self-verification accuracy.
A new post-training quantization method scores every token individually using integrated gradients, closing the gap to full precision to just 1.33 percent at 3-bit weights.
A new reinforcement learning framework uses an inverse dynamics model as a reward signal to close the gap between pretty videos and executable robot motions.
A new expert-annotated benchmark of 246 real dental cases exposes how top MLLMs fail at safety-critical clinical routing when compared to junior dentists.
Lakr233 open-sources vphone-cli, a Swift CLI that automates the multi-step process of booting a patched iOS 26 VM on Apple Silicon Macs using Apple's own Virtualization.framework.
OpenDataLoader PDF hits #1 on extraction benchmarks with 0.907 accuracy, runs fully local, and is the first open-source tool to auto-tag PDFs for accessibility compliance
Mamba-3 redesigns state space models from the ground up for inference, beating Transformers on latency while matching or exceeding them on accuracy at the 1.5B scale.
K-Dense open-sources Claude Scientific Writer, an agentic tool that researches, writes, and formats publication-ready papers, grants, and clinical docs from your own data.
Apple open-sources Embedding Atlas, a WebGPU-powered tool for interactively exploring millions of embedding vectors with auto-clustering, cross-filtering, and MCP agent access.
T3 Code is an open source control plane that orchestrates Claude Code, Codex, Cursor, Grok, and OpenCode from one interface using your existing subscriptions.