Comment by writingdna

The article frames this as "semantic ablation" but the underlying mechanism is more specific: it is distributional averaging. RLHF and DPO reward policies optimize for the modal response given a prompt distribution. That is not a bug in the training process, it is the objective function working as designed. The model learns to produce the response that the median annotator would rate highest, and that response is, almost by definition, the least distinctive one.

What is underappreciated is how much stylistic signal lives in what information retrieval people call "burstiness" -- the tendency for distinctive words to cluster rather than distribute evenly. Hemingway's short declarative stacking, DFW's recursive parentheticals, legal writing's formulaic precision -- these are all bursty patterns that a model trained to maximize expected reward will sand down. You can partially recover it with few-shot prompting, but the model is fighting its own reward gradient the entire time.

The practical question is whether you can encode a style prior that survives the decoding process. The research on authorship attribution (stylometry) suggests the feature set is well-understood -- function word frequencies, sentence length distributions, type-token ratios, syntactic complexity metrics. But nobody has built a production system that uses those features as a constraint during generation rather than just detection.