Training data must be valid under all searched policies, eval data doesn't
JoshPurtell · x · 2026-11-02
The author draws a distinction: training data must be valid under all conditions/policies that could be searched during training, while eval data only needs to be valid for the particular set of finished models produced by a particular procedure—ideally ones trained in ignorance of the evals. He adds that optimization pressure goes beyond reward hacking: "benchmaxxing" was a major problem before the RL-data-ification of evals, precisely because eval data has a much narrower validity mandate than training data.
More from Research
- J-Lens is just a first-order Taylor expansion, researcher points out — burny_tech · 2026-11-02
- Mollick: expert-level work benchmarks are brutally hard, and ARC-AGI, GDPval, METR are saturated — emollick · 2026-11-02
- Rebuilding a Protein Domain Distribution Pipeline After a Backup Failure — AryHHAry · 2026-11-02
- NVIDIA's Elastic Latent Recurrence lets one 4B model trade loop count for accuracy: 18→29 on AA Index — burny_tech · 2026-11-02
- IBM's STAIR retriever cuts invalid references 65× and hits 82.6% accuracy by using document structure — solyarisoftware · 2026-11-02
- ReLift Hits SOTA Dense 3D Tracking on WorldTrack With 2.7x Faster Pipeline — Technion · 2026-11-02