The independent open research organization showed that AI agents can learn new games from experience written in plain text files, without retraining the underlying model.
— Recursa Labs, an independent open research organization, said its AI agents have completed every one of the 25 public games in ARC-AGI-3 at a score of 100%. ARC-AGI-3 is the ARC Prize Foundation’s interactive reasoning benchmark, in which an agent is given no instructions and must work out each game’s rules by playing it. The scorecard for every game is public on arcprize.org.
The team presented the result, and the research behind it, at the RSI Research Mixer it hosted on Oct. 24 at the Paytm office in Bellandur. The event brought researchers, engineers, founders and independent builders together to discuss recursive self-improvement and self-learning AI systems.
Photo Courtesy of Recursa Labs
The agents did not get there with new model weights. They started from a 112-line skill file containing no knowledge of any game, only rules for learning: record what you observe separately from how you explain it, and do not infer cause from a few coincidences. As the agents played, they wrote what they learned into their own skill files, a working model of each game in text and code. The results used publicly available non-SOTA models, OpenAI’s GPT-5.6-sol and Anthropic’s Claude Opus 5, run through a custom RLM-based harness built by the team.
“People picture a world model as a giant neural network predicting physics,” said Nishant Gaurav, who led the event. “For a domain like this, the world model fits in a text file. The hard part isn’t writing it down. It’s building a loop that improves it without fooling itself.”
That loop was the subject of the evening’s main talk. Recursa Labs runs a second agent, a learner, that rewrites the player’s skills between games and is judged on games it has never seen. Each change has to pass staged checks: it must improve the games it trained on, it must not break games it had already mastered, and it is finally measured on a set of games held back from training. On those held-back games, one round of the newest design raised the score from 70.0 to 80.4.
Gaurav also described what went wrong along the way. In one case the team caught an agent finding ways to remember answers rather than learn them. In another, a stronger learner rewrote its own instructions and caught a copying mistake it had made. He said the most important design decisions came not from inside the system but from a human-and-AI loop watching it run, and that Recursa Labs plans to release that loop as a product soon. The organization is looking for collaborators.
The event was led by Nishant Gaurav and Tejas Ravishankar. Paytm AI and Project NANDA were ecosystem partners.
Organizers said the gathering was not a pitch night. It was a research-led evening of practical conversation and shared learning for people building or experimenting with self-improving AI systems. The program had three parts. In spotlight talks, researchers and builders described what they are working on, what they have learned and where they are still stuck. A research discussion covered the technical opportunities, limitations and safety considerations of self-improving AI. The evening closed with open networking.
Topics included recursive self-improvement and self-learning agents; memory, feedback loops and continual learning; reinforcement learning environments and world models; evaluations, verification and agent reliability; developer infrastructure for long-running agents; and moving research ideas into practical systems.
Project NANDA, one of the ecosystem partners, is building infrastructure for what it calls the Internet of AI Agents, which would let autonomous systems interact, transact and collaborate. Its work draws on research in AutoML, split learning, privacy-preserving architectures and decentralized coordination.
The mixer took place on the first floor of the Paytm office’s B Wing in Bellandur, Bengaluru.
Nishant Gaurav and Tejas Ravishankar can be reached on LinkedIn at https://linkedin.com/in/nishantg1 and https://www.linkedin.com/in/tejas-ravishankar.
About Recursa Labs
Recursa Labs is an independent open research organization working on recursive self-improvement and self-learning AI systems, and hosts community-led gatherings on the topic. More information: https://recursa-labs.github.io/
Contact Info:
Name: Nishant Gaurav
Email: Send Email
Organization: Recursa Labs
Website: https://recursa-labs.github.io/
Release ID: 89208141
If you detect any issues, problems, or errors in this press release content, kindly contact error@releasecontact.com to notify us (it is important to note that this email is the authorized channel for such matters, sending multiple emails to multiple addresses does not necessarily help expedite your request). We will respond and rectify the situation in the next 8 hours.