Renormalizing probability scores destroys calibration, dev warns in model training debate
spikedoanz · x · 2026-09-17
In a reply to a proposal to simply renormalize scores (adding just a few ms), spikedoanz argues this is no way to build a coherent prediction model: renormalization turns (0.09, 0.01) into exactly (0.9, 0.1), when the model should really signal uncertainty (0.5, 0.5). Training with KL on such independently sampled scalars would likely drive the model insane.
Related event: Developers Warn Against Training on Renormalized Probability Scores(2 posts)→
More from Research
- OpenAI's next model "Doug" nears solving the Hodge Conjecture, a Millennium Prize Problem — kimmonismus · 2026-09-18
- Experiment calibrates 48 attention heads down to 12, swapping the rest for band-diagonal sparse attention — ostrisai · 2026-09-18
- NVIDIA open-sources NeMo Data Designer: declarative synthetic data pipeline lifts Nemotron Nano v3 from 80.2% to 86.9% — dair_ai · 2026-09-18
- New blog finds data issues in a frontier benchmark, ports Agents' Last Exam CLI subset to Verifiers v1 — dejavucoder · 2026-09-18
- Hybrid Mamba-Transformer skips RoPE: Nemotron-style arch fixes Mamba's ICL and long-context weaknesses with few attention layers — gordic_aleksa · 2026-09-18
- FOOM is only plausible if P=NP, argues one skeptic of the intelligence explosion — jessi_cata · 2026-09-18