AI Writer: Beam Search Story Generator

Decoder Settings

Beam width K 3

Candidates per node 4

Temperature (exploration) 0.8

Diversity penalty 0.6

Best Beam

Once

Cum. log-prob

0.00

Perplexity

—

Active beams

Originality

—

All Beams

How it works

A real autoregressive AI writer picks each next word by scoring candidates with a probability P(word | history), then decoding a full sentence step by step. Greedy decoding always keeps the single best word and can get stuck in a mediocre sentence — beam search instead tracks the K best partial sentences ("beams") at once:

score(path) = Σ log P(w_t | w_<t)
keep = top-K paths by score(path), each step
perplexity = exp( -score(path) / length )

This simulator stands in a small deterministic scoring function for a trained language model: every candidate word gets a pseudo-affinity log-probability with the previous word (repeating a word is penalized like a real repetition penalty), plus a temperature-scaled random term that models sampling noise / model uncertainty.

Beam width K — how many candidate sentences are tracked in parallel; wider beams explore more of the tree before committing.
Candidates per node — how many next-word options are generated from each beam at every step (the model's local vocabulary fan-out).
Temperature — scales random noise added to each candidate's score; low = the decoder trusts the scoring function and stays coherent, high = more erratic exploration.
Diversity penalty — subtracted from a candidate's score for each already-selected beam this round that chose the same word, so beams are pushed toward different words (diverse beam search) — a direct probe of the AI writer's originality: raise it and the beams stop converging on the same phrase.

The 3D tree grows away from the camera one depth layer per step: node height reflects cumulative score, kept branches glow purple, pruned candidates fade to grey and are removed. Originality below is the fraction of unique words used across all beams' final sentences — a low value means the "writer" is repeating itself, a high value means genuinely divergent phrasing.

Kept beam node Pruned candidate

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.