Truth as a Compression Artifact in Language Model Training

When transformers train on contradictory data, the same problem with both correct and incorrect solutions, which answer do they prefer? We hypothesize that next-token prediction, as a compression process, favors whichever answer cluster has lower description length; truth benefits only when errors lack internal structure. We test this by training transformers (3.5M-1B parameters) from scratch on controlled corpora, systematically varying the structure of errors. We find that (a) when errors are random, models develop a correctness preference scaling from 65% to 85% with model size; (b) when errors follow a single coherent alternative rule, this preference vanishes (~45-51%); (c) two competing wrong rules suffice to restore it (47% to 78%). The pattern reproduces on Wikipedia paragraphs with entity substitution (71% vs 46%) and at 1B scale on a mixed natural-text corpus (77% vs 47%). These results are consistent with the hypothesis that, in controlled contradictory corpora, model preference tracks the relative compressibility of competing answer systems rather than truth per se.