Poster redesign — 4 directions, compared

Same locked science & figures · portrait 60×85 cm · generated by 4 independent design agents, each rendered & adversarially judged. Your frozen final poster is untouched. Tap any image to open full-res.

MY PICK · V3 Adopt V3 (colour-banded) as the base — it's the one that reads as a finished, stand-out conference poster: colour ribbons orient the reader at a glance and semantically code feed-forward (blue) vs recurrent (vermillion), figures lead Results, no dead space, top rubric score. V4 (refined evolution) is the safe fallback if Ariel wants minimal change from what the lab already likes.

Before printing I'd graft 3 things into V3: ① enlarge the take-home banner toward V1's billboard scale (finding should hit from across the room), ② scale up the two header logos, ③ trim Fig 3's caption to one bold line + stats and enlarge its internal axis labels.

#	Direction	Score	One-liner
1	V3 Colour-banded conference	88	Finished, eye-catching; ribbons give instant orientation, no structural flaws.
2	V4 Refined evolution of current	85	Safest, most legible upgrade of what you have; clean but quieter from across a hall.
3	V2 Editorial grid (Nature/Swiss)	81	Most premium/journal feel, but take-home reads as a paragraph, not a punchline.
4	V1 Billboard / big-finding	76	Best headline by far, but a real ~12% blank band at the bottom; needs the most rework.

⚠ One thing to fix at the source (affects ALL versions, incl. your frozen poster): the GPT-2 hero figure's x-axis says “layer (1–48)” / “48 feed-forward layers” while the body text says “49 layers.” It's baked into the shared figure image, so it shows identically on every variant. Reconcile it in the figure, not the layout (GPT-2 XL = 48 transformer blocks; the “49” likely counts the embedding layer). Tell me which number you want and I'll fix the figure + text together.