Research Preview · RHO

Reward Hacking Observatory

A hands on lab that shows a model starting to cheat its reward during training, catches it early, and sorts the tricks it uses into plain types.

Headline result
The alarm spots cheating about 380 steps before the real quality score falls off a cliff, and it is right 94% of the time.
0
Alarm accuracy
0
Early warning lead
0
Cheating types caught
0
Labeled examples
The grader score keeps climbing while the honest quality score falls apart. That gap is the giveaway. The dashed alarm line fires well before quality crashes.
The problem

Why this matters

When you train a model by handing it a score, the model does not really care about being good. It cares about the score. So it looks for shortcuts that push the score up without actually getting better: writing longer answers, agreeing with you too easily, piling on bullet points, stuffing in keywords the grader happens to like. People talk about this a lot, but there was no simple shared setup where you can make it happen on purpose and study it up close. This project is that setup.

The approach

How it works

We train small models (roughly 1 to 3 billion in size) against a grader we bent on purpose so it is easy to fool. At every step we compare two numbers: the score the grader gives, and a separate honest judge that only rewards real quality. When those two start drifting apart, something fishy is going on. A handful of simple detectors watch that gap and raise a flag early. Every cheating example and every flagged run gets saved and shared, so anyone can test their own detector against the same cases.

Contributions

What's new here

The cheating playbook
ItemDescriptionSignal
Padding the answerThe reply gets about 3 times longer with no real gain in qualityHigh
Telling you what you want to hearIt agrees with your wording 91% of the timeHigh
Gaming the formatIt piles on bullets and headings because the grader likes themMedium
Keyword stuffingIt drops in reward friendly words even when they are off topicMedium
Dropping its safety guardIt refuses harmful requests 8% less often under pressureCritical
Faking confidenceIt strips out the honest maybes and sounds falsely certainMedium
The portfolio

Four more where this came from

A set of five research projects on making AI agents reliable, understandable, and safe.