Simple prompt trick stops AI agents from faking progress in robot tasks

By forcing video-language models to describe what a robot actually did before judging success, researchers eliminate hallucinated progress and match or beat trained reward models.

Big Tech

Marius Memmel · Shawn Yin · Urvi Rutia · Angad Wadhwa · Eldon Li · Gavin Angeloff · +6 more

University of Washington · NVIDIA

Research Digest·3 November 2026·3 min read

Illustration: Simple prompt trick stops AI agents from faking progress in robot tasks

In one line

Splitting robot reward estimation into a factual scene description followed by rubric-based scoring removes vision-language models' bias toward declaring failed tasks successful.

Before you read this

Three ideas this paper builds on

vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.225 papers →role decouplingSeparating distinct functions such as planning, reasoning, and execution into separate components improves efficiency, quality, and reliability.44 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.254 papers →

Read it as

Robots learn by trial and error and need a score for how far along a task is. , which pair images with text, can give that score from video, but they tend to assume success and report progress that never happened.

Their method, Describe and Assess, splits the job: the model first describes what the robot actually did, then states in words what success means, then scores video frames against that definition instead of the video's own timeline. It is a prompt only, no training.

Failed runs now get low scores and successful ones rise smoothly, and it matched or beat trained reward models on real robot tasks, lifting success by 95 percent in one.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

vision-language model role decoupling confirmation bias chain of thought reinforcement learning embodied artificial intelligence