Introducing Reinforcement Learning from Collective Rewards (RLCR).
RLCR is a recipe that trains language models as agent organisations, which share a message board and workspace. We use it to train Avalon-M1-Mini (120B) and study its behaviour.
We find a number of interesting results, including:
🛠️ Learnt organisational hierarchies and divisions of labour
🕵️ Increased collective verification, through hires of verifiers and testers
👥 The rise of collective thinking, where agents think more about other agents and how to help them
We believe RLCR can provide a laboratory for studying collective intelligence and multi-agent safety.
Blog post and technical report 👇
What happens when AI learns to organise itself?
Last Friday, I spoke at the Autonomous Organisation (AO) Summit in London, co-hosted by @AO_Commons and @ARIA_research’s Scaling Trust programme.
At @GenReasoning, we built Avalon, where AI agents learn to organise and collaborate autonomously. I believe collective intelligence is the next frontier of AI.
But greater autonomy raises a bigger question: How do humans stay in control? And more fundamentally, who decides what's worth pursuing?
To me, AI isn’t just a technical challenge, but a question of human identity.
Thanks to @ObadiaAlex, @anke_g_liu, @samccarthy27 and the team for bringing together thoughtful minds.
🐝 How to Train a Swarm
In our first blog post, we explain how latency pressure can incentivise parallel work, the communication primitives agents need to co-ordinate, and outstanding challenges in training large agent collectives.
gr.inc/blog/how-to-train-a-s…
Hear the story of how we developed Avalon at the London AO Summit hosted by @ARIA_research.
We’ll be talking not only about the results, but the research process of how we got there, and what our future roadmap looks like.
General Reasoning CEO @rosstaylor90 explains how their new RLCR experiment taught AI agents to organize themselves, hire teammates, and decide when to DM each other:
"The idea of RLCR is, instead of just trying to optimize single agents, how do you optimize collective structures? Teams, research communities, and eventually economies, markets."
"We optimize an organization in a Slack-like harness. They learn who they should hire, what roles they should hire for, when they should be pinging someone annoyingly in their direct messages versus making a public channel."
"It’s not just another model, it’s a whole new ball game. It’s a whole new direction."
@GenReasoning
Introducing Reinforcement Learning from Collective Rewards (RLCR).
RLCR is a recipe that trains language models as agent organisations, which share a message board and workspace. We use it to train Avalon-M1-Mini (120B) and study its behaviour.
We find a number of interesting results, including:
🛠️ Learnt organisational hierarchies and divisions of labour
🕵️ Increased collective verification, through hires of verifiers and testers
👥 The rise of collective thinking, where agents think more about other agents and how to help them
We believe RLCR can provide a laboratory for studying collective intelligence and multi-agent safety.
Blog post and technical report 👇
Read below for a summary of our findings and our technical report.
This is still under revision as we train Avalon-M1-Mini. We are a small team, so please be patient as we get back to you on any questions, or make updates.
gr.inc/releases/avalon-m1
makes experiments way more reproducible, its a nice way to guarantee that solution on the live internet won’t leak your evals, and prevents hacking.
it’s not a big uplift if you already have your own search engine.
General Reasoning folks are even offering it as a service:
🔍 Introducing BackSearch.
LLMs are increasingly asked to predict the future, but a good backtest requires a snapshot of the internet at a point in time.
BackSearch allows LLMs to search the web as it was on a particular date. It’s great for:
🔮 Forecasting and prediction markets.
📈 Quantitative finance.
🌍 RL environments that simulate the world.
🥶 Freezing websearch for benchmark reproducibility
We’re releasing a narrow slice of our index to begin with, focused on the news domain for 2026. Based on feedback we’ll open up more of our index in subsequent releases.
👇 Try BackSearch out with the link below.
🔍 Introducing BackSearch.
LLMs are increasingly asked to predict the future, but a good backtest requires a snapshot of the internet at a point in time.
BackSearch allows LLMs to search the web as it was on a particular date. It’s great for:
🔮 Forecasting and prediction markets.
📈 Quantitative finance.
🌍 RL environments that simulate the world.
🥶 Freezing websearch for benchmark reproducibility
We’re releasing a narrow slice of our index to begin with, focused on the news domain for 2026. Based on feedback we’ll open up more of our index in subsequent releases.
👇 Try BackSearch out with the link below.
Because of backtest variance, we also record a process-based rubric measure called "sophistication", which we track over time. This uses a human expert rubric to judge the sophistication of the strategies employed.
GLM-5.2 shows impressive sophistication compared to other open models, although it still does not surpass the closed SoTA on this metric for any time this year.
You can see the full leaderboard and more analysis here:
gr.inc/releases/introducing-…
We evaluated recent open models on KellyBench.
Here is what we found:
🏆 GLM 5.2 is new open source SoTA, but still loses -30% on average over 5 runs.
📈 We estimate GLM 5.2 is 6+ months behind the frontier based on KellyBench and internal quant evaluations. (Note: we have not evaluated Fable)
🌗 Kimi K2.6 slightly improves on Kimi K2.5 but still struggles at -60% average RoI.
🐈 Recent Mistral models struggle, obtaining mean RoIs of -78% and -99% respectively.
Leaderboard link and more graphs below.
We've updated the leaderboard with GPT-5.5 results:
gr.inc/releases/introducing-…
TLDR: 30% strategy sophistication (new SoTA), much more efficient than GPT-5.4, but still in the red and systematically underperforming human approaches.
New models on KellyBench, the benchmark for long-horizon sequential decision making.
🎆 Claude Opus 4.7 - new state-of-the-art, with the highest strategy sophistication recorded on the benchmark (28.5%). But still loses -3.7% on average over five seeds.
🐋 DeepSeek V4 Pro - shows strong feature development capabilities, but scores poorly on overfitting rubrics which harms it in deployment (-47%).
Link and more graphs below.