ICLR 2026 Workshop first day

Kyunghyun Cho

Sergey Levine

• robot foundation models

rlpd

• noneed soft actor critic

• rl for specialist policies

rldg

• distilling rl into openvla

• training on data from rl specialists gives us a better generalist

• can the speicalists also benefit from the generalist
• adding rlt rl tokens to robot foundation models

• cfg and ppo is almost same?

• cfgrl with o optimality variable
• score function

• good/bad data and data during all rl training or just final policy

• condition the policy on optimality indicator

• rl is the key to lifelong long learning of robot foundationmodels

• directly training end to end rl is hard

David Bau

• transplant

1. where does this related - patching - where or when does it infer

2. what is being influenced - probing - what concepts does it use its decision

3. how can we manipulate - steering - can we directly change its thoughts

But why

• to debug audit and improve the models
• are all the last layers of sequence reasoning needed

• to customize the models beyond their trianing
• can we interpolate conditions

• to discover new science/discovery
• what does the ai know that scientist/human do not yet know

Yoshua Bengio

Azalia Mirhoseini

• test time scaling