Today we're launching a multi-lab open collaboration, the ARBOR project, to accelerate AI interpretability research for reasoning models. Please join us! github.com/ARBORproject... (ARBOR = Analysis of Reasoning Behavior through Open Research) - ThreadSky

About ThreadSky

wattenberg.bsky.social • 11 days ago

Today we're launching a multi-lab open collaboration, the ARBOR project, to accelerate AI interpretability research for reasoning models. Please join us!

https://github.com/ARBORproject/arborproject.github.io

(ARBOR = Analysis of Reasoning Behavior through Open Research)

Comments

wattenberg.bsky.social•11 days ago

Take a look at some initial research projects, and see if there's one you'd like to work on:
https://github.com/ARBORproject/arborproject.github.io/discussions
Or propose your own idea! There are many ways to contribute, and we welcome all of them.

ngutten.bsky.social•11 days ago

Is there a good place for partial results, small experiments without intent to expand, etc here? E.g. 'I trained Qwen2.5B for a week on tsumego problems on one GPU and it failed to learn chain of thought' kinds of things?

wattenberg.bsky.social•11 days ago

Excellent idea! I just added an "observations" index page on the wiki for things like that.
https://github.com/ARBORproject/arborproject.github.io/wiki/Observations

ngutten.bsky.social•9 days ago

Alright, I've put a small writeup of the tsumego experiment, hope that's okay. This was done in collaboration with Crosslabs, but I'm not sure how you would like that sort of credit to be noted - let me know and I'll adjust it.

wattenberg.bsky.social•9 days ago

Thank you! That's a great write-up, and this is definitely an interesting experiment. The distinction between how the model might do parsing vs. solving is very much worth thinking about. I added a few thoughts on the wiki page. https://github.com/ARBORproject/arborproject.github.io/wiki/Chain-of-Thought-for-Tsumego-(Go-Life-or-Death)-Problems

ngutten.bsky.social•9 days ago

Definitely been thinking about whether there's a way to work in a natively 2d space for this, e.g. to get chains of thought made up of image sequences. That would reduce parsing difficulty. But I'm not sure I can finetune a good enough multimodal model on one GPU.

Posting Rules

Be respectful to others
No spam or self-promotion
Stay on topic
Follow Bluesky's terms of service

Comments

Posting Rules

Reply