ProgressCompass: Embodied Progress Reward Models Are Lost Without the Right Context

ProgressCompass. Embodied Progress Reward Models Are Lost Without the Right Context. By Jianshu Zhang and eight others, from Northwestern University, the University of California, Santa Barbara, and two other institutions. Published on the twenty-ninth of September. No code repository is linked to it.

Embodied agents now take on ever longer tasks, and for a long task, knowing only whether it finally succeeds or fails says little. The steps along the way matter. Progress reward models, or PRMs, address that by scoring how far a task has come at every step, which lets them serve as dense rewards, verifiers and monitors. The trouble is that in a long task, the current frame alone often cannot show how far things have come, because progress depends on what happened before. The authors call this context-dependent progress estimation, and their claim is the one in the title. Embodied PRMs are not incapable of estimating progress. They are lost without the right context.

Existing benchmarks for progress estimation mostly use short tasks, whose progress can be read off the current observation. To test the context-dependent case, the authors build ContextProgress-Bench, with twenty-four manipulation tasks across one hundred twenty episodes. It covers three settings. In State Recall, information needed to judge progress appeared earlier and is not in the current frame. In Sequence Tracking, steps follow a fixed order, so progress means knowing which steps are done and which comes next. In Recurrence Disambiguation, frames that look alike sit at very different points of progress.

The diagnosis is paired. Each PRM keeps the same input format in both runs, and in one of the runs its instruction integrates the right context. Without it, even PRMs that read the entire history get lost. With it, the same five models cut their progress error by seventy-seven to eighty-two percent.

That result motivates ProgressCompass, an autonomous agentic loop that reorients an existing PRM, using current general-purpose VLMs to supply the context the PRM needs. Wrapped in the loop, the same frozen PRM cuts its progress error by sixty-three percent and raises its rank agreement by seventy-six percent.

The evidence comes from manipulation tasks, twenty-four of them, and the paired diagnosis covers five models. The largest error reductions come from runs where the right context was integrated into the instruction, a controlled condition built to isolate the cause. ProgressCompass is the step toward supplying that context automatically, and its gains are reported for a frozen PRM inside the loop: smaller than in the paired diagnosis, but obtained without the context being handed over. The claim concerns long tasks where progress depends on history, which is exactly the case existing benchmarks had left underexplored.

The reframing is the contribution. A model that fails at a task may lack the ability, or it may lack the information, and the paired runs are built to tell the two apart. Here the answer is information. That turns a modelling problem into a context problem, and ProgressCompass is one way of solving it, leaving the frozen model as it was and changing only what it is given to work with.