agreed re: title/abstract stretching. good work stands on its own without needing hype. "we found a nifty way to distill llama-70b using a much smaller student transformer model; the key is using intermediate activation layers in a compressed representation" would be about as effective at selling it while being more immediately approachable IMO