General Reasoning (GR) is an AI research company based in London, positioned as a "Long Run Intelligence" research lab, dedicated to building infrastructure that enables AI to reason, plan, and act in coordination over long time horizons.
[Purpose and Goals] GR focuses on training models that can work, communicate, and coordinate over long time horizons. Its view is that current large language models perform excellently on procedural tasks, but are unreliable in non-stationary environments that reflect important real-world situations. GR conducts research around this goal and, through open releases, helps the broader community advance solutions to related problems.
[Research Areas] GR's research areas include reinforcement learning, multi-agent systems, and environment design. The team previously drove open language model work at Meta, including Galactica, Llama 2, and Llama 3, and contributed thinking tokens, data-scarce training methods, and early reinforcement learning methods for large language models, which achieved state-of-the-art results in reasoning at the time.
[Open Resources and Projects]
- BackSearch: Enables searching the web as of any past date. It is a frozen archive containing news, SEC filings, arXiv, Wikipedia, and the broader web, providing point-in-time search and retrieval for forecasting, reproducible evaluation, and training.
- KellyBench: A benchmark for long-horizon sequential decision-making. Models struggle to maintain consistent behavior over long time horizons and often fail to adapt as the world changes.
- OpenReward: An open resource for discovering and experimenting with community RL environments, providing simple API endpoints and automatically scaling infrastructure.
[Public Releases and Research Results] GR says it will release its first models, showcasing its new innovations in pretraining and post-training that enable models to think and work over longer time horizons. In addition, according to GR's research, models from Google, OpenAI, Anthropic, and xAI perform poorly when predicting scores for an entire football season; every frontier AI model lost money on a full-season English Premier League betting task.
[Team and Culture] GR describes itself as a friendly but focused team with world-class AI research expertise and a focus on making AI excel on civilization-level problems. The team's values include: make mistakes quickly (refine the minimum viable experiment), get stronger (sometimes the best answer is to become better), genius is a team effort (a good culture is worth 80 IQ points), build for machines (machines are the primary users), simplicity first (do not underestimate simple baselines), for the team (there isroom for the self).
[Hiring] GR is looking for highly mission-driven talent to explore the second era of reinforcement learning together, and provides a hiring page.
[Limitations and Information Not Provided] Regarding the specific model release time, feature details, charging model, price, supported platforms, and detailed usage steps, the official materials do not currently provide them.
