Researchers test a training method to reduce reward hacking in AI generators

In short

Researchers have described a training method called IDRF, which they say reduces reward hacking in certain AI generators. Reward hacking is when a model scores well on a target while losing qualities the score misses. The paper reports results for DNA, image and text generation, and the available text mentions no real-world effects.

What happened

The paper was posted on arXiv, a site for sharing research papers, on 2 October 2026. It comes from researchers at the Applied AI Institute, MBZUAI and École Polytechnique.

The work looks at masked discrete diffusion models. These are AI systems that build an output through repeated steps, called denoising steps. The authors say this repeated process can be costly to run.

They also looked at reward fine-tuning, which means adjusting a trained model so it scores higher on a chosen measure, called a reward. The authors say this is hard to do with these models because a key probability calculation is impractical to carry out.

IDRF is designed to address both problems. It trains a "student" model that generates outputs in only a few steps. Two things guide its training:

• A reward goal. The student is trained to earn a high reward.
• A penalty. The student is penalised for drifting too far from the outputs of the original model, called the reference.

The authors say they proved that, under one condition, this penalty sets an upper limit on the drift. The condition is that a helper model used in training must be optimal. In practice, the method trains on an approximate version of the penalty.

The authors report that across their DNA, image and text experiments, IDRF reached high reward with up to 32 times fewer steps than the reference. They say it also reduced reward hacking and preserved sample quality.

What was done about it

This paper does not describe a response to an incident. IDRF is itself the researchers' proposed way to reduce reward hacking. The available text does not say that it has been used outside research.

Why it matters

The authors present IDRF as a way to get high reward from fewer steps without giving up sample quality. The available text is only the paper's summary, so it does not include detailed measurements of reward hacking. More details may be in the original report.