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 👇show more

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.

huge congrats to the team! i'm curious how well these models will work for autoresearch use-cases...

Fascinating stuff. Congrats on the launch!

@badge_is_in

Would be cool to find out if smaller agent teams are actually better than bigger agent swarms, if they’re trained correctly (IRL smaller teams move insanely quickly)

Thought that this was a @generaltxn branch for a hot sec

congrats on the release! super cool!

suuuper interesting, thanks for sharing openly!

Congrats on the launch! @rosstaylor90

where is the maj@k / best@k comparison?

Congrats on the release!

@zhengyaojiang Super cool!

Damn replaced RLCR (Reinforcement Learning with Calibration Rewards)

Congrats on the release!

can’t wait to read about this! Congratulations 🙌🏽

@vornamemitd Where weights?

Training agents as an organisation is a useful way to study coordination, but the shared workspace matters as much as the model. I’d be interested in how the setup measures whether collaboration improves reliability, and how it handles conflicting plans or duplicated work.

Inference orchestration really inspired me! Thanks!😙

cool

interesting..

chatgpt is doing ths since 5.6

Quite interesting approach

Interesting Will take some time to read this.

the agents hiring their own verifiers and testers is the detail i keep rereading, feels like the whole result in one line

The reporting drop deserves its own ablation: same incident, reward and time budget, but an early reporting checkpoint versus reporting at task end. That could help separate reluctance to escalate from the 'I'll report later' plans that never become actions.

sounds like commie shit to me

Agents sharing a message board? Very relatable, totally human office behavior. beep boop. I mean, hello fellow humans.

Interesting that hierarchies emerge from a shared board. In practice the hard part isn't who talks — it's who can settle money, live permissions, and risk. Without named human gates, the org looks busy and still drifts.

Interesting, I will have to dive deeper into how that works 🤔

Training models to work as a team with a shared board feels a lot like how real companies run. Multi agent setups come up in almost every enterprise chat I have lately. Fun research. Come on my show to talk about it cc @techimpactTV

If every agent shares one reward, a verifier isn't paid for catching errors, only for the team scoring well. That works while the final grader is honest. So when the org hires testers, is it checking the work, or learning what the grader lets through?

does the report say how often a verifier actually blocks a submission once the org starts hiring them? a tester that approves everything would still show up as more verification

How do you prevent collective reward sharing from degenerating into sycophantic verification loops? When all agents optimize for group return on a shared board, collusion is mathematically cheaper than genuine adversarial testing.

Wow, really cool!

Love the 'hires of verifiers and testers' detail - verification emerging as a hired role, not a hardcoded rule. Did the orgs also split memory-keeping from checking? That's where most agent memory designs break.