Agyn introduces a novel multi-agent system that models issue resolution in software engineering as a team activity, rather than a monolithic or pipeline process.

Leveraging the capabilities of LLMs, the system assigns specialized agents to roles such as coordination, research, implementation, and review, replicating the structure and processes of a real development team.

Agyn operates autonomously through the entire development lifecycle, from analysis to pull request creation and review, achieving a 72.4% task resolution rate on SWE-bench 500. This highlights the significance of organizational structure and agent infrastructure.

👉 Read the full article on arXiv

BrainFuse is a unified infrastructure that integrates biophysical neural simulation with gradient-based learning, bridging the gap between neuroscience and artificial intelligence.

The system integrates detailed neuronal dynamics into a differentiable learning framework and accelerates customizable ion-channel dynamics by up to 3,000x on GPUs.

Demonstrating capabilities in both neuroscience and AI tasks, BrainFuse can deploy models with approximately 38,000 neurons and 100 million synapses on neuromorphic hardware with low power consumption, accelerating the development of next-generation bio-inspired intelligent systems.

👉 Read the full article on arXiv

MEG-XL is a data-efficient model designed for clinical brain-to-text interfaces, which require minimal training data from paralyzed patients.

The model is pre-trained using 2.5 minutes of MEG (magnetoencephalography) context per sample, significantly longer than prior work, capturing extended neural context equivalent to approximately 191,000 tokens.

MEG-XL achieves supervised performance with a fraction of the data and outperforms existing brain foundation models, demonstrating that long-context pre-training effectively exploits extended neural context for improved word decoding.

👉 Read the full article on arXiv