Reduced Text-to-SQL agent execution time by up to 5x and decreased tool calls from 8 to 1.
A user test revealed that Claude's 'Use account memory' toggle is enabled by default for new projects and was found enabled on an existing untouched project, allowing non-project chats to access project-specific memory files.
Routing transplant experiments on OLMoE and DeepSeek models show that forcing morality-associated expert paths significantly alters malicious code detection accuracy, independent of the model's readable workspace.
A 300M-parameter transformer trained on synthetic recurrent sequences achieves 0.9–2.4 bits per byte on real languages after reading 1 million bytes.
A flaw has been identified in Aurora MySQL versions 3.08.0 to below 3.14 where index DDL operations clear the Reader buffer pool.
GPT-6 Astra scored 55.0% on Ironclad research tasks, a 32% improvement over GPT-5.6 Sol, while reducing estimated time per attempt by 48%.
A Carnegie Endowment study found that for every 30 Chinese AI researchers in the US, only one moved to China in 2025.
Ergo is an early alpha open-source tool designed to provide auditability and explainability for Rego policies.
GLM-5.3 Flash scored 85% on one attempt, compared to GPT-5.6 Luna's 81%, with scores dropping to 50% and 45% respectively on the hard subset of the benchmark.
The dataset features over 1 million unique tasks collected in 8 weeks, significantly outperforming existing public UMI benchmarks in diversity and scale.
The new 7B model achieves 84.83% on the Alyah benchmark and significantly outperforms larger general-purpose models in dialect fidelity and cultural understanding.
Data integration across three subsidiaries including GS Power enables approximately 690 users to access data via catalog subscriptions
Analysis of 1,250 arXiv papers establishes a taxonomy for self-improvement loops at deployment and training time.
Researchers released an open dataset and security harness showing that trivial configuration mistakes, not kernel exploits, drove most container breakouts in autonomous AI agents.
Introduction of filter tabs reduced detail page access time by over 20% and achieved 82% admin feature coverage
The open-source tool achieves 89.8% Recall@1 on LongMemEval-S while running on minimal hardware like an Intel Celeron N4020.
The platform introduces a structure where AI generates test flows, executes them on real devices, and performs self-recovery upon failure.
The model runs entirely on a single 80 GB GPU without external services.
OpenRig allows developers to define and boot multi-agent coding teams using YAML specs, managing sessions via tmux and a local daemon.
The integration is available on all paid Claude plans and includes new connectors for editing files directly from Claude.
OpenWAM is an open framework for composable world-action models that achieves 98.6% success on the LIBERO benchmark using a Mixture-of-Transformers architecture trained on 14.64k hours of robot video data.
Codemode allows LLMs to execute JavaScript within a sandboxed harness to compose tool calls, manage state, and access internal APIs without consuming context tokens.
The Jev 1.13.0 model achieved a 76.2% pass rate on SREGym-Lite benchmarks, performing close to GPT-5.6 Sol's performance while being approximately 7 times faster and 200 times cheaper.
RISED uses LLM-generated rubrics to guide data selection and self-distillation, achieving the highest mean pass rate across diverse interactive environments.
AI Briefing