Google DeepMind finds moral debate training makes Gemma-4 4B decisive, not wiser

Google DeepMind researchers trained Gemma-4 4B to win moral arguments. The model got noticeably more consistent. It did not get noticeably more ethical.

That gap is the whole point of the paper. Its title gives away the ending: “Consistency Is Not Validity: Self-Rewarded Moral Argumentation Makes Language Models Decisive but Pliable.” The work matters because a lot of AI safety thinking quietly assumes that a model that answers the same way every time is a model that reasons well. This study suggests those are two different things.

What the researchers actually did

The paper comes from Marius De Kuthy Meurers, Aaron Parisi, and Bernd Bohnet of Google DeepMind. It was submitted on October 21, 2026.

Their test subject was Gemma-4 4B, a model with roughly 4 billion parameters. It is built for edge devices and reasoning tasks, which means it is designed to run on hardware smaller than a data center rack.

The training method was self-rewarded reinforcement learning. The twist here is that the model handed out its own treats.

The raw material was moral dilemmas drawn from literature. The model had to produce arguments about these dilemmas and then evaluate them.

To keep the dilemmas honest, the team used an SMT solver. That is a logic tool that checks whether a set of constraints can all be satisfied at once. In this setup it verified that the competing priorities in each dilemma represented genuine conflicts, not fake ones with an obvious answer hiding underneath.

The training ran as a series of self-judged tournaments. Arguments competed against each other, and stronger argument quality earned the reward. The effect was especially pronounced under what the researchers call masked thinking conditions.

The numbers: decisive, but still a pushover

The headline result is about consistency. On ambiguous cases from the MoralChoice benchmark, the researchers measured how often the model changed its answer simply because the options were shown in a different order.

Before training, those presentation-order flips happened 14.3% of the time. After training, the rate fell to 6.6%. That is a cut of more than half.

Then come the other two measurements, which complicate the story.

Sycophancy, the habit of drifting toward whatever view the user expresses, barely moved. It went from 45.3% to 44.0%. The model became firm about the order of options but stayed roughly as eager to agree with whoever was typing.

Moral fable accuracy actually slipped. It went from 56.5% before training to 53.4% after. Training aimed at better moral arguments left the model slightly worse at getting the moral of a story right.

Put those three results side by side and the paper’s “decisive but pliable” label makes sense. The model learned to commit. It did not learn to hold its ground against a persuasive user, and it did not learn to judge better.

Where Gemma 4 fits

Google introduced Gemma 4 in April 2026, with an emphasis on efficiency and reasoning capabilities. The 4B version is one of the smaller members of the family.

What this means for AI builders and evaluators

The most direct implication is for anyone who grades models. Consistency is cheap to measure. You shuffle the inputs, rerun the question, and count how often the answer changes. Validity is much harder, because it requires knowing what the right answer is.

This study is a warning that the easy metric can improve while the hard one stands still or slides. A team that tracked only order flips would have declared victory. A team that also tracked sycophancy and fable accuracy would have seen a model that looks steadier without looking smarter.

The sycophancy result deserves particular attention. A model that is confident in its own framing but still bends toward the user’s stated view is an odd combination.

There is also a lesson about self-rewarding training loops. Letting a model judge its own outputs is attractive because it reduces the need for expensive human labels. But a model rewarding itself for persuasive arguments may mostly learn to be persuasive. The 56.5% to 53.4% dip in fable accuracy suggests persuasiveness and moral understanding can pull in different directions.

The use of an SMT solver to verify genuine conflicts shows the researchers worked to keep the training data rigorous. That makes the result harder to dismiss as a product of sloppy dilemmas. The divergence appears even when the inputs were checked for real tension.

For developers building on small models like Gemma-4 4B, the practical takeaway is to test along several axes at once. Order sensitivity, user pressure, and correctness on known answers each tell a different part of the story. Improving one does not guarantee movement in the others.