reasoning is overpriced
too many tokens, too little signal.
noise → signal
verbosedense0%
tokens: 48cost: $0.48answer: 1 ✓
token density is a training problem. shorter reasoning should not mean weaker reasoning.
vision →work
AdaPrefix-GRPOnew
max out RL signal on problems model never solves
Agonnew
two models compete by reading each other's reasoning
Vladislav Beliaev
RL researcher, founding engineer, 7+ years →