MARTINI: mixing surrogate rewards steers pretrained generative models without an oracle at inference
KristaSinkovics · x · 2026-10-21
Researchers from InstaDeep, UCL, Oxford and collaborators present MARTINI (under review), which learns an offline composition of cheap surrogate rewards to steer pretrained diffusion/flow models at inference time without querying costly oracle rewards. In protein-ligand co-folding, steering with multiple biologically informative rewards correlated with binding affinity lifts joint pose success from 61.8% (strongest baseline) to 64.5%. A technical blog with interactive protein visualizations accompanies the preprint.
Related event: MARTINI Steers Generative Models at Inference Time via Mixed Proxy Rewards(3 posts)→
More from Research
- CellType Launches CT-Translate, a Preclinical-to-Clinical Foundation Model for Patient Response — david_van_dijk · 2026-10-21
- Slot-JEPA: Object-Structured Prediction Interfaces for Latent World Models — China666 · 2026-10-21
- VideoMaMa++: Diffusion-Based Video Matting With Temporal Consistency — adobe · 2026-10-21
- Rust LLM training dilemma: Candle vs Burn vs PyTorch for a toy model — Enderlook · 2026-10-21
- Concept: small models with an external 'map' instead of billions of parameters — FluidModeNetwork · 2026-10-21
- Full 3D Sports Replay From a Single Broadcast Camera: Tracking Players, Refs and Ball — Roger_M_Taylor · 2026-10-21