Pinned Tweet
when generating synthetic agent traces, matching the distribution in your real data is only the first step.
if your filters accept successful tasks more often than cases where the agent fails or asks for help, your training data still ends up skewed.
ensure you measure the distribution you keep, not just the one you ask the generator for.
collective rewards are really cool but like who gets the credit?
one agent does the work and five of them share the reward (lowk communism).
or does the CEO take all the reward and if shit goes bad all the blame too?
feels like step-level rewards matter even more here
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 👇
we built Edgecase to turn expert decisions into reusable training recipes.
it learns what makes a decision correct, varies the situation and checks whether the new answer holds up.
live today. Code + recipe are open source :)
your agent's traces only cover what already broke.
synthetic data should not imitate them it should extrapolate and generate the failures that haven't happened yet.
early results on real production failure diagnosis:
- trained on ZERO real traces, it matches a model trained on real expert traces (82% vs 82%)
- few-shot post-training stops being a lottery (every run 80–86%)
- the model sees failure types no expert ever labelled
introducing edgecase🐵a platform to extrapolate your agent's traces into training data.
recipe and code are fully open source.
ManIT retweeted
every team post-training agents has at most a handful of labelled traces, which they also use for eval.
nobody has thousands and the few you have only show the failures you have already seen.
so most synthetic data just remixes them. more of the same.
we went the other way:- extrapolate from your traces and generate the failures that haven't happened yet.
then we trained a model on zero real traces and put it head to head with one trained on real expert data.
the result surprised us.
launching tomorrow. open source + live on the platform 🐒
imitation gets you to the average expert. extrapolation gets you to the edge case.
imitation = rewrite your few expert-labelled traces. more of the same.
extrapolation = generate the failures that haven't happened yet.
we trained on zero real traces and put it head to head with real expert data. result surprised us. open source + live tomorrow 🐒
"machine autonomy emerges not from data abundance but from strategic curation of high-quality agentic demonstrations"
same way taste works. you don't build it from doing everything, you get it from doing few high priority things and execute those really well
we gave the agent an AI judge and it learned to do performative research instead of actually doing research
the judge loved it tho 😭
🤖 Made with AI
Excited to release Step-Jev 🔥
step-by-step rewards for agent RL, even when there’s no answer key
After DeepSeek R1, everyone said step rewards get hacked, so just reward the final answer.
We found the opposite. With an AI judge grading only the final answer, our agent learned to fool it and accuracy dropped every single update until it was worse than no training at all.
Step rewards fixed it. Everything is open!
been tinkering with this for a while now, would love your thoughts on the results
@chem_safety @danielrupawalla @ypatil125
damn so claude science is really an RLM, honestly makes sense.
does this mean the next step is RLM+CLM ?
@RulinShao @a1zhang
the goat gave his thoughts
xitter.cf/a1zhang/status/2105409…
few thoughts, I've been sent this paper many times today, and it's cool!
1. it's a v clever idea, and I'd say even more extreme than RLMs on the spectrum of ReAct-style vs. pure context offloading (so no they're not the same)
2. im hopeful to see harness designs that use this principle, and they somewhat go hand-in-hand with model + RLM progress as well
3. the KV issues scare me admittedly, and the proposed fix is a bit hacky despite it seeming to work well for their results. that being said, I suspect different model shapes won't have this issue, so it's fixable! another + for this direction
I think that the function of making it for any startup is truly connected to how fast they adapt to what they see + hypothesize about what will happen in the coming few months
ManIT retweeted
Introducing Noma: an open decision model.
State in. Calibrated answer out. No generated text.
16 ms per decision
~40x faster than Jev
$0.023 per 1,000 decisions
100% on JevBench easy
Open weights, MPL-2.0.
pip install blackdrome-noma
ngl the part that got me wasn’t even beating jev
we trained it on another verifier’s labels and it basically learnt the judgement itself
feels kinda stupid to keep paying a huge model to make the same decision forever once you have enough traces lol