Precision-weighted LLM training wins blind tests

Interesting result, but still firmly in replication-needed territory: the signal is real enough to take seriously, yet the study combines two mechanisms, uses one seed, and leans on a judge panel that is broad but not fully controlled.

• –The method is cheap and optimizer-agnostic, so if the effect holds, it could slot into existing pretraining loops without extra wall-clock overhead.
• –The strongest claim is the loss/quality decoupling: identical smoothed validation loss, but a consistent preference edge in blind A/B tests that survives multiple sensitivity filters.
• –The main weakness is causal attribution: token gain and layer gain were not ablated separately at scale, so it is unclear which part matters most.
• –The evaluation is more convincing than a lone anecdote because it mixes humans and foundation-model judges, and both groups point in the same direction.
• –The obvious next step is a multi-seed reproduction with separate ablations and a larger model or longer-context benchmark to see whether the effect survives outside short-form preference tests.