August 11, 2026

A new benchmark of 170 rigorously curated code-refactoring tasks spanning seven programming languages, built to fix quality and test-suite issues found in prior SWE-Bench variants. The best evaluated coding agent resolves only 41.2% of tasks, showing current agents still struggle with large-scale multilingual refactoring.

Mind Lab presents an open agent-model family that keeps learning after deployment via a Mixture-of-LoRA architecture: a frozen base model composes specialist LoRA adapters (chat, coding, agent, GenUI) selected per turn, with usage data distilled back into the adapters over time.

Introduces a 1,253-example, expert-annotated benchmark across 60 subjects in four scientific disciplines for testing whether video generation models get scientific facts and causal dynamics right, not just visual realism. Finds models vary widely on scientific correctness with a pronounced gap between proprietary and open-source systems.

Technical report for Motif 3, a 314B-parameter (13.2B active) decoder-only Mixture-of-Experts model built on Grouped Differential Latent Attention plus manifold-constrained hyper-connections and a multi-token-prediction head, with competitive results on reasoning, coding, and long-context tasks.

Proposes a training-free framework that distills a large teacher agent's successful trajectories into hierarchical memory (workflow, subtask, and function-level) for small student agents, injecting workflow/subtask memory proactively and function memory reactively on tool-call errors. Tested with 4B-8B student models against a GPT-5-mini teacher, giving accuracy gains up to 27.2 points over baselines on tool-use benchmarks.