MARTINI Steers Generative Models at Inference Time via Mixed Proxy Rewards
InstaDeep and collaborators propose MARTINI, which mixes cheap proxy rewards to steer pretrained generative models toward an intractable target reward distribution at inference time, without retraining or querying the true reward.
2026-10-21 ~ 2026-10-21 · 3 related posts
- MARTINI: mixing cheap surrogate rewards steers diffusion models without retraining — KristaSinkovics · 2026-10-21
- MARTINI: mixing surrogate rewards steers pretrained generative models without an oracle at inference — KristaSinkovics · 2026-10-21
- MARTINI algorithm mixes surrogate rewards to steer generative models without querying the oracle — mmbronstein · 2026-10-21