GPT-5.5 tops private Kaggle citation benchmark

My read: this looks more like a retrieval-and-attribution win than a pure reasoning breakthrough, and that’s exactly why the 5.4 to 5.5 jump is interesting. The smaller 5.4 mini beating 5.4 in the same setup also hints that the release changed calibration, training mix, or decoding behavior in ways that matter on this kind of memory-heavy eval.

• –Exact-title recovery is vulnerable to training-data proximity and memorization effects, so a strong score does not automatically mean deeper reasoning
• –Private Kaggle benchmarks are useful signals, but without public methodology they are easy to overread and hard to compare across model families
• –If the result holds up, GPT-5.5 may be better at scientific search, citation assistance, and source attribution workflows where precision matters more than eloquence
• –The 5.4 mini vs 5.4 inversion suggests model size alone is not the whole story; post-training and inference tuning likely matter a lot here
• –For developers, the practical takeaway is to still verify citations externally, even when the model appears unusually strong at recall