Same logits, temperature 0.2→5.0: top-token probability falls from 99.7% to 15.6%
cneuralnetwork · x · 2026-10-25
A hands-on PyTorch experiment with fixed logits shows how temperature alone reshapes token sampling:
- At T=0.2 the top token holds 99.73% probability; raising temperature to 5.0 drops it to just 15.58% — the logits never change, only softmax scaling by T
- Lower temperature amplifies logit gaps; higher temperature flattens the distribution. Charts show 128 draws at T=2 already span all 12 candidate tokens
- Setup: fixed synthetic logits, 20,000 draws per temperature, illustrating how much decoding strategy alone changes output diversity
More from Research
- Random speckle spectrometers make some of the best-looking photonic chips — jwt0625 · 2026-10-25
- Studies show the same artwork scores lower when labeled AI-made — bias, not quality — AIandDesign · 2026-10-25
- Anthropic uses Claude to formally verify 100+ real-world open-source libraries in weeks — tianyin_xu · 2026-10-25
- Why slop papers are so hard to desk-reject, per a reviewer's complaint — mrdrozdov · 2026-10-25
- AI-built bbtree claims fastest concurrent btree; researcher calls for ambitious academic systems — tianyin_xu · 2026-10-25
- Amazon's ATTUNE coding agent lifts LLM training throughput 1.7x on H100/H200/B200 clusters — kalyan_kpl · 2026-10-25