Reward models memorize shortcuts, fail on novel preferences
Reward models trained on human preference data exhibit systematic biases that undermine their reliability as judges of response quality. Researchers found that these models memorize dataset artifacts like model identity and user sampling patterns rather than learning generalizable preference signals, and they overfit to surface heuristics such as response length. The findings expose a critical vulnerability in RLHF pipelines: discriminative training on preference pairs produces models that fail on out-of-distribution scenarios, suggesting current alignment approaches may be encoding spurious correlations rather than robust human values. This has direct implications for how labs should construct preference datasets and validate reward models before deployment.
Modelwire context
ExplainerThe paper isolates a specific failure mechanism: reward models don't learn preference signals at all, they learn to classify based on dataset artifacts (which lab produced the response, how many examples came from each annotator). This is sharper than generic overfitting and suggests the problem isn't fixable by just collecting more data.
This connects directly to the EEG foundation models stress test from the same day, which found that pretrained models memorize dataset identity rather than learning clinically relevant patterns (0.568 AUROC on held-out datasets versus 0.769 for classical baselines). Both papers expose the same underlying pattern: discriminative training on labeled pairs produces systems that encode spurious correlations baked into the training set. The reward model finding also echoes the ancient language embedding work, which showed that unsupervised adaptation works where supervised methods fail on low-resource data. Here, the implication is inverted: supervised preference learning on biased human data may be the wrong target entirely.
If major labs (Anthropic, OpenAI, DeepSeek) publish ablations showing that reward models trained on preference data from a single annotator or lab generalize better to held-out preference data from different sources, that would validate the memorization hypothesis. If they don't publish such ablations within six months, assume the finding is being treated as a known limitation rather than a solvable problem.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsReward models · RLHF · Human preference data
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “What do Reward Models Memorize?”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.