ManIT

rl @ bandr • there is no greater danger than playing it safe

[ManIT @manofrl]

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.

[laxman @llmluthor]

an open recipe for synthetic agent traces plus everything that didn’t work https://t.co/PZTMSnuKdz

[ManIT @manofrl]

@danielrupawalla Hmm that does make sense cuz google has literally watched people search for the last 25 years. that has to account for something right?

[ManIT @manofrl]

@michael_chomsky But like tbh ain't that contrarian and like... impressive if u really think about it

[ManIT @manofrl]

@tokenbender Gillfoyle who the fuck have I been talking to??

[ManIT @manofrl]

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

[General Reasoning @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 👇

[ManIT @manofrl]

@llmluthor generating convincing traces was much easier than showing they helped the trained model.

that gap deserves more attention than only the generation count.

[ManIT @manofrl]

@deepseek_ai yoo lets fucking gooo.

[ManIT @manofrl]

@RRicefan Wait ur telling me argon has taste???

Soemthings wrong imo

[ManIT @manofrl]

@llmluthor more like an autoscaling issue

[ManIT @manofrl]

i used to pray for days like these

[laxman @llmluthor]

had so much traffic, it went down now back again :D

[ManIT @manofrl]

@llmluthor the 🐵 knows. trust me, he's always known.

[ManIT @manofrl]

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 :)

[laxman @llmluthor]

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 @manofrl]

@llmluthor the 🐒 is carrying this entire post

[laxman @llmluthor]

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 🐒

[ManIT @manofrl]

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 🐒

[ManIT @manofrl]

@llmluthor Praise the fool

[ManIT @manofrl]

tbh was named with good faith

[ManIT @manofrl]

lowk cooked with the name

[laxman @llmluthor]

getting the guy behind jev to say "help step-jev" was not on my bingo card 😭

[ManIT @manofrl]

context -> diogo(jev creator) reposted our stepjev tweet

[ManIT @manofrl]

we’re so back (the reward was never stuck)

[Diogo Almeida @CompleteSkeptic]

Help Step-Jev, my reward is stuck (plateauing) 🥵