Answer-token decision outputs are uncalibrated, engineer cautions
LysandreJik · x · 2026-10-12
Follow-up in Lysandre's decision-model thread.
- Recap: decisions become answer tokens; after one forward pass, softmax over only the A/B/C logits yields the distribution (billing 0.03 / technical 0.94 / sales 0.03).
- Key caveat: unlike dedicated decision models, this output is not calibrated, so logits are not necessarily interpretable.
More from Research
- Ian Osband to talk at Oxford on 'The Price of Learning': policy gradient can be a terrible way to learn — IanOsband · 2026-10-12
- Jina AI to Host 4th Embeddings & Search BoF at EMNLP2026 in Budapest — JinaAI_ · 2026-10-12
- Yoav Goldberg defends OpenAI's FFT result: critique TCS culture, not the paper — yoavgo · 2026-10-12
- DeepMind's Ian Osband to argue policy gradient 'can be a terrible way to learn' in Oxford talk — IanOsband · 2026-10-12
- Scott Aaronson: the singularity has already begun for math — ChrSzegedy · 2026-10-12
- UN green chemistry brief highlights AI-designed enzymes beyond directed evolution — AnimaAnandkumar · 2026-10-12