Softmax probabilities don't match real-world frequencies, breaking cross-company calibration
HanchungLee · x · 2026-09-17
The author argues that while softmax outputs a probability distribution, it does not guarantee that distribution matches real-world frequencies — and frequencies for next-token prediction differ from classification. As a result, a system-one model calibrated for one company won't transfer usefully to another.
More from Research
- Common Crawl puts crawl archives on Hugging Face Storage Bucket, with a getting-started guide — vanstriendaniel · 2026-09-17
- TMLR editor interviewed authors of low-quality submissions: they had no idea what their papers said — TuhinChakr · 2026-09-17
- Schmidhuber: I published the first concrete RSI algorithms in 1987, now RSI is driving AI's future — SchmidhuberAI · 2026-09-17
- FLYPITCH uses a fruit fly connectome to generate delightfully insane WIRED story pitches — willknight · 2026-09-17
- Protein Models ESMC and ESMFold2 Land on HuggingFace with NVIDIA-Built Fused Triton Kernels — AllThingsApx · 2026-09-17
- Evals shouldn't be skewed by infra: why turn limits break on sub-agents — langstonnashold · 2026-09-17