BRISA Reading Group 03 - TOPReward

​Before our official launch this fall, we’re continuing our pre-launch series with another reading group: sitting down with a paper, understanding it properly, and asking what it would take to build it ourselves.

​Our third paper is TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics (Chen et al., 2026), which asks a surprisingly simple question:

​Can a pretrained Vision-Language Model already tell us how well a robot is doing — without training a separate reward model?

​In robot learning, knowing whether a task was completed is relatively easy. Knowing how much progress the robot is making along the way is much harder.

​TOPReward approaches this without manually designing rewards or training another model.

​Instead of asking a VLM to explicitly output a progress score, it asks the model whether a task has been completed and looks directly at the probabilities it assigns to the possible answer tokens.

​Those token probabilities become a zero-shot reward signal that can be used to measure task progress, detect success, and reweight robot demonstrations.

​What makes the idea especially interesting is that the reward model itself was never explicitly trained to be a reward model.

​The paper therefore raises a broader question:

​How much useful supervision is already hidden inside foundation models if we know where to look for it?

​TOPReward also entered our exploration through MolmoAct2, where it is used as a reward signal inside a larger robot-learning pipeline.

​During the session, we’ll focus on three things: how TOPReward actually works, how it is applied to a relatively simple visuomotor policy like ManiFlow — which is not a VLA — and what happens to policy performance when demonstrations are reweighted using this signal.

​That gives us an interesting question beyond TOPReward itself:

​Can knowledge hidden inside a VLM improve robot policies that do not use a VLM at inference time at all?

​And, as always, we’ll look beyond the headline results: how convincing is the reward signal, what does it actually capture, where might it fail, and how difficult would it be to reproduce the idea ourselves?

​No preparation needed. Skim the abstract if you have ten minutes. Everyone is welcome — whatever you study, wherever you study and however much robotics you have done so far.

​Come with questions, disagreements, or your own take on how foundation models should be used in robot learning.

​🎟️ There are only a limited number of spots, so we can keep the session personal and interactive. Please register early and only attend if you've received a ticket.

​📍 Hosted at the Merantix AI Campus. Doors open 18:30, we start at 18:45.

​👋 New here? Join the community group — everything else gets announced there first.