Calder

Mathematician & Research Scientist | MIT Math PhD '24

[Calder @caldermf]

Excited about this paper I worked on with @NghiMDao, @adam_block66985 and @GolowichNoah! We let models retry after wrong answers during RL training and found that this improves their accuracy on math benchmarks, even when they only get one attempt at test time.

[Adam Block @adam_block66985]

🚨New paper led by my amazing master's student @NghiMDao! 🚨
Introducing Retry-Aware Policy Optimization (ReAPO). Conditioning retries on failures improves RLVR exploration: ReAPO improves accuracy and training efficiency relative to GRPO.

1/12