POST-TRAINING / CONTROLLED COMPARISONS
Compare the named base checkpoint with its post-trained variant. Keep the evaluation items, score scale and treatment visible.
Explore training methods →Published base-to-treatment comparisons
ErdősBench · 139-item controlled slice
ArxivNet SFT + Verified Reasoning V-SAO
+0.500 mean points / 4
139 evaluation items · average judgment
ErdősBench · 139-item controlled slice
Verified Reasoning V-RLVR
+0.064 mean points / 4
139 evaluation items · average judgment
ErdősBench · 139-item controlled slice
ArxivNet SFT + Verified Reasoning V-RLVR
+0.409 mean points / 4
139 evaluation items · average judgment
SimoBench · 126 proofs
OlympiadNet O-RLVR
+0.206 mean points / 7
126 evaluation items · mean proof credit
369 → 395 proof points out of 882. The means above divide these totals by 126.
Ulam powers RLTasks. These are publisher-reported results, not independent replications. Ulam AI · Frontier Data for Frontier Models — Training Lift Scorecards ↗
Publisher-reported comparisons against the named untouched base on the same evaluation items. Exact checkpoint hashes, complete compute ledgers and uncertainty estimates are not supplied on the reviewed scorecard page. The 139-item ErdősBench slice is separate from the 226-item model leaderboard.
These comparisons study particular training routes and evaluation slices. A lift on one slice does not establish improvement on every task or isolate each stage of a combined treatment.
Use the per-model benchmark pages to study broader capability, and use a controlled base-to-treatment comparison to ask what a training change contributed.
Design a matched evaluation →