# Iluvatar Labs

> Reality-inspired AI architectures for the AGI era. We build closed-loop autonomous systems for post-training, scientific research, and bioinformatics. Founded in 2025, based in the United States.

Iluvatar Labs is an AI research lab building tools that close the loop between model training and downstream behavior. Our core thesis is that today's AI stack is largely open-loop — models are trained, evaluated, and deployed without continuous correction — and that the next decade of progress requires closed-loop systems with continuous monitoring, automatic adjustment, and rich context tracking. Our products and research apply this thesis across post-training, autonomous scientific research, and computational biology.

## Products

- [Marvin](https://iluvatarlabs.com/marvin/): An autonomous research agent for ML and scientific work. Marvin reviews literature, generates and tests hypotheses, runs full research loops end-to-end, and maintains rich persistent context (Mithril) across sessions. Cross-domain: frontier AI, computational biology, materials science.
- [Actuator](https://iluvatarlabs.com/actuator/): A patent-pending closed-loop control layer for model transformation. Streams reward and eval signals through live training, adapts pressure when capability drifts, and halts before damage. Drops into existing post-training stacks. Use cases include distillation, compression, and reinforcement learning with reduced alignment tax.

## Research

- [Divergent-Convergent Attention (DCA)](https://iluvatarlabs.com/blog/2026/03/divergent-convergent-attention/): A transformer attention primitive that maintains parallel attention streams at different window sizes and reconciles them through learned periodic consensus. Improves long-context reasoning and recall.
- [Elastic Speculation](https://iluvatarlabs.com/blog/2025/11/elastic-speculation/): Speculative decoding with adaptive depth. Delivers 30-50% lower inference latency and ~50% fewer speculative KV writes while preserving output quality.

## Announcements

- [Meet Marvin](https://iluvatarlabs.com/blog/2026/03/introducing-marvin/): Why we built an autonomous research agent and how it differs from existing options.
- [Introducing Actuator](https://iluvatarlabs.com/blog/2026/04/introducing-actuator/): Why post-training is the primary differentiation lever for AI labs in 2026 and how Actuator closes the loop.

## More

- [Discover (full blog index)](https://iluvatarlabs.com/blog/): Product announcements, research notes, field reports.

## Technical resources

- [Sitemap](https://iluvatarlabs.com/sitemap.xml)
- [RSS / Atom feed](https://iluvatarlabs.com/feed.xml)

## Contact

- Email: ben@iluvatarlabs.com
- X / Twitter: [@IluvatarLabs](https://x.com/IluvatarLabs)
- LinkedIn: [Iluvatar Labs](https://www.linkedin.com/company/iluvatarlabs/)