BULLPENThis demo runs BULLPEN model embeddings in the browser for an anonymous ICLR 2027 submission.
02 Embedding space

What the symbols mean

How it works

This demo

Limits

Method: ability, style and the model card

Descriptive statistics from the embeddings; not reported paper results

The model

How this demo estimates θ

How this demo estimates s

How the card selects properties

What the plotted axes track

Reported results

#What we foundValue
2Text (768-d) minus bits (32-d), style tasks+.34 [.25, .44]
4Text minus each label's own defining formula (length, hedging, refusal)−.11, −.12, −.15
5Text minus statistics, six other targets+.20 [.10, .30]; +.08 without Arena
7Fused code minus bits and minus text, 32 dims, all tasks+.102, +.125; 3 of 3 seeds
8Text (768-d) minus fused code, style tasks; refusal+.09 [.03, .15]; .45 vs .27–.28
12Unseen models, text at 25 minus bits at 900 questions, style tasks (all tasks)+.27 [.15, .39] (.41 vs .37)
17Closed models: accuracy vs best open; rank of hedging and length.86–.93 vs .81; Spearman .97