The Ethereum Foundation, in collaboration with Yukon and zkSecurity, launched the better.codes automated research challenge focusing on koalaIRS12 (the Reed–Solomon proximity problem). The goal is to raise the soundness lower bound of hash-based SNARKs to 128 bits through machine-verifiable proofs; all submissions are verified by the Lean kernel, valid proofs are publicly recorded on-chain, and results feed back into the community.
An NVIDIA technical blog post introduces how to collaboratively build a Holoscan real-time endoscopic tool segmentation application using CLI tools, HoloHub development skills, and AI coding agents (Codex + GPT-5.6 sol max mode).
An NVIDIA technical blog post introduces the support capabilities of NVIDIA FLARE 2.8.0 for federated multimodal AI workflows, including large object externalization, PyTorch tensor streaming downloads (Tensor Downloader), and server-side disk-supported aggregation (tensor disk offload).
NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter omni-model (including a 2-billion-parameter Nemotron reasoning module), optimized specifically for running robot control policies on Jetson Thor.
NVIDIA released SkillEvaluator, an open-source tool for evaluating the impact of AI agent skills on performance. It supports static checks and comparative real-task runs (including Codex and Claude Code harnesses).
In a blog post, AI21 noted that the accuracy bottleneck in agentic search is answer selection rather than generation, as majority voting cannot correct prevalent errors; it introduced independent verifiers that re-retrieve each candidate answer individually and reject incorrect ones.
Google has open-sourced the Common Expression Language (CEL) formal verification framework. Based on the Z3 theorem prover, it supports mathematical correctness proofs for CEL expressions and policies; it can detect logical equivalence errors, incomplete guard conditions, and security invariant failures. It achieves zero false positives through three-stage taint tracking and supports automated verification of AI-generated policies in CI/CD.
IBM Research and Hugging Face jointly released the ALTK-Evolve method, which allows agents to distill reusable guidelines from their own trajectories and inject them during inference without weight updates or manual annotation; evaluated on the AppWorld benchmark across eight models (including DeepSeek-V3.2.
NVIDIA's official technical blog introduces how the ALCHEMI Toolkit (v0.2.0) provides PyTorch-native, GPU-accelerated MLIP construction modules along with accompanying AI coding agent skill packages.
A team from the University of Sherbrooke in Canada developed the Ice Dart drone, equipped with retractable micro-spiked feet and friction dampers. It can land on ice slopes approaching 60 degrees at a speed of 3 meters per second, achieving a 100% success rate on 58-degree slopes with wind speeds of 30 kilometers per hour. Field tests have been completed on Fjallsjökull Glacier and iceberg lakes in Iceland. This capability supports long-duration silent observation and improves iceberg tracking accuracy.
Cursor has released a new Git storage system named Continuity. It adopts a write-ahead log (WAL) architecture with S3 as the source of truth, replacing the three-phase commit consensus mechanism used in traditional Spokes. It supports an arbitrary number of replicas, stateless routing, optimistic replication, and WAL-driven compaction. It is optimized for high-frequency agent creation of small repositories and large-scale CI scenarios involving enterprise monorepos.
The official Hugging Face blog announced the Sentence Transformers v6.0 update, which introduces the new MultiVectorEncoder type with native support for ColBERT-style late-interaction retrieval models (such as PyLate.
Dharma AI has released a constraint-aware GPU scheduler that models real-time inference as time-series demand curves and allocates batch jobs by priority across scheduling cycles. This approach increases GPU utilization on identical hardware by up to 33 percentage points and boosts priority-weighted output by up to 105%. The scheduler operates with latencies of 1–15 milliseconds in scenarios ranging from 8 to 64 GPUs, and includes open-source models and an interactive demo.
NVIDIA released the Nemotron 3.5 Lightning NVFP4 checkpoint, achieving near-full-precision accuracy at a size of 22GB (compared to 66GB for the original BF16) through Quantization-Aware Distillation (QAD).
Persona AI has selected shipyard welding as the first deployment task for its humanoid robots due to characteristics such as a shortage of skilled welders, task rhythms suitable for robots (centimeter-per-second speeds, millimeter-level repetitive actions), and environments requiring bipedal movement (crossing beams, working on floors/overheads/portholes). It has established partnerships with HD Hyundai and POSCO, aiming to deploy hundreds of robots at a single site, emphasizing value creation through increased capacity rather than cost reduction.
The Axiom Math team used their AI system AxiomProver to automatically formally verify the 246 theorem on prime gaps for the first time—namely, that there exist infinitely many pairs of primes differing by 246. This work established a library of prime gap results reusable for other mathematical formalization tasks, aiming to extend formal methods to the security verification of AI-generated code.
Google Research released a study proposing the PhotoScan deep learning framework, which can estimate 3D body composition metrics (body fat percentage BF%, Android-to-Gynoid fat ratio A/G) from standard 2D smartphone photos.
Ollama v0.32.12 has been released, adding support for the Qwen 3.8 27B model, which shows improved performance in coding, professional work, research, and long-horizon agentic tasks; it also optimizes performance and output quality for Apple Silicon devices, providing a dedicated qwen3.8:27b-mlx tag.
Cursor has been officially acquired by SpaceX, completing the acquisition process initiated in April. The two parties had previously collaborated with SpaceXAI to accelerate model training. Cursor will receive support from SpaceX's largest global GPU cluster to provide stronger models at lower costs; Grok 4.6, released on Wednesday, is an initial result of this integration.
Zhipu released GLM-5.3, which retains the same base model but significantly enhances the intelligence ceiling through extreme post-training scaling. Scores improved from 4.6 to 28.3 on Terminal-Bench 3.0 and from 46.2 to 66.9 on DeepSWE v1.1.