Noah Farr

Master Student in AI and Machine Learning at TU Darmstadt.

[Noah Farr @__noahfarr__]

@yacineMTB Wasting flops on? You rather are more sample efficient since the critic just gets the info and doesn’t have to extract it from incomplete data

[Noah Farr @__noahfarr__]

@JadeYi25 Nice! I tried to make BC work but somehow I didn’t get it to work well.

[Noah Farr @__noahfarr__]

@creus_roger Well done!

[Noah Farr @__noahfarr__]

the fly brain can also find diamonds now in craftax @mitrma

[lyra @_lyraaaa_]

the fly brain can play beat saber

[Noah Farr @__noahfarr__]

@JustinWaugh Good luck! I also am currently working on pushing the boundaries of Craftax. My agent is getting stuck in fiery realm though

[Noah Farr @__noahfarr__]

@JustinWaugh So you are doing BC of the Astra demonstrations?

[Noah Farr @__noahfarr__]

@creus_roger @mitrma Thanks! Not published because it isn't a result, it's the baseline for
something I'm working on. Runs on my own stack, PPO with a few PufferLib
(@jsuarez5341) tweaks, tuned on Craftax.

On Astra, agreed that it would be cool. 1% of 5B is 50M steps, where we're at about 10 return.

[Noah Farr @__noahfarr__]

@creus_roger @mitrma For scale, this is a 1M parameter MinGRU. The 1B run takes 14 minutes on a single
A100 and the 5B run takes 1 hours, at about a million environment steps a
second.

[Noah Farr @__noahfarr__]

@creus_roger @mitrma We get there with plain recurrent PPO too, running thousands of environments on one GPU. At 5B steps it reaches the vault in 19% of episodes and the troll mines in 4%. It scores 69 there and 47 at 1B, where the best leaderboard entry is 41. No trillion parameter LLM needed.

[Noah Farr @__noahfarr__]

microdux 🦆 all 14 official Microduck tasks, in JAX.

Walk, turn, spin, skate. MuJoCo Playground native, so registry.load just works.

Rewards verified term by term against the original rather than eyeballed.
https://t.co/IGS7J8km8Z

[Noah Farr @__noahfarr__]

@j_foerst Any chance of the same for the 18 PhD slots? Would love to hear what the perfect DPhil applicant looks like

[Noah Farr @__noahfarr__]

@allenainie Don’t forget to checkout https://t.co/J3gep1WM7k from the RL in Big Worlds Workshop at RLC!

[Aryaman Reddi @ReddiAryaman]

Our paper on Streaming RL under Partial Observability won Best Paper Award at the Big Worlds Workshop at #RLC2026. We enable memory in streaming RL using recurrent trace units.

Project page: https://t.co/gXGExiEIIz

Collaborators: @__noahfarr__ *, @CarloDeramo , @Jan_R_Peters

[Noah Farr @__noahfarr__]

@jxmnop reports of its death have been greatly exaggerated. built a browser extension that brings the crimson back with one click https://t.co/sL9qmBCqPs

[Noah Farr @__noahfarr__]

@yacineMTB It really isn’t. It takes the elegant parts and slaps on stuff like clipping to make it more well behaved because that’s cheaper than a proper trust region.

[Noah Farr @__noahfarr__]

@jsuarez Report both.

[Noah Farr @__noahfarr__]

@jsuarez I like to think of offline RL as the extreme point of off-policy RL

[Dominik F. Farr @Dominik2718]

Wikipedia race with GPT-4o-mini. https://t.co/z0XfVbVQQB

[Mark Tenenholtz @marktenenholtz]

When I train an RL agent to play soccer