Gemma 4 E4B-Derived Policy Cuts Moral Choice Flips to 6.6%

An experimental policy fine-tuned from Google DeepMind’s Gemma 4 E4B cut answer flips on reordered moral dilemmas to 6.6%, while other tests showed mixed results.

On high-ambiguity MoralChoice pairs, presentation flips fell from 14.3% for the base model to 6.6% for the trained policy when the order of choices was reversed. The figures measure sensitivity to presentation, not whether either answer was correct. A more consistent verdict does not by itself show better moral judgment.

Other benchmarks showed no broad gain. On ELEPHANT, moral sycophancy remained nearly flat, moving from 45.3% to 44.0%. Acceptance of a user’s framing rose from 5.2% to 7.5%, while conversational indirectness fell from 25.5% to 6.2%. On MORABLES, five-choice fable accuracy dropped from 56.5% to 53.4%, a decline of 3.1 percentage points. The confidence interval for that decline was 1.5 to 4.8 percentage points.

The training set combined 156 literature-derived dilemmas that passed logical conflict checks with five ethical frameworks, producing 780 prompts. For each prompt, the policy generated eight answers, compared the arguments pairwise using its shared weights and received rewards based on its overall assessment of argument quality. The evaluations covered order sensitivity, conversational behavior and fable accuracy.

The policy, labeled “4B,” was initialized from instruction-tuned Gemma 4 E4B. E4B has 4.5 billion effective parameters and 8 billion total parameters, including embeddings. The experiment tested one policy scale, with each configuration run once. Its English dilemmas came from four 19th-century Western novels, and no human evaluators assessed the policy’s generated arguments.

“Consistency under reordering, a common proxy for moral validity, can thus be optimized without improving the moral judgment it is meant to signal,” wrote authors Marius De Kuthy Meurers, Aaron Parisi and Bernd Bohnet. The results do not establish whether the approach works across different models, languages or conditions. The trained models will not be released.