Reward Weighted Classifier-Free Guidance as Policy Improvement in Autoregressive Models
About
Consider an auto-regressive model that produces outputs x (e.g., answers to questions, molecules) each of which can be summarized by an attribute vector y (e.g., helpfulness vs. harmlessness, or bio-availability vs. lipophilicity). An arbitrary reward function r(y) encodes tradeoffs between these properties. Typically, tilting the model's sampling distribution to increase this reward is done at training time via reinforcement learning. However, if the reward function changes, re-alignment requires re-training. In this paper, we show that a reward weighted classifier-free guidance (RCFG) can act as a policy improvement operator in this setting, approximating tilting the sampling distribution by the Q function. We apply RCFG to molecular generation, demonstrating that it can optimize novel reward functions at test time. Finally, we show that using RCFG as a teacher and distilling into the base policy to serve as a warm start significantly speeds up convergence for standard RL.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Molecular Property Optimization | lipophilic efficiency | Normalized Reward0.6 | 10 | |
| Molecular Property Optimization | drug like | Normalized Reward0.45 | 10 | |
| Molecular Property Optimization | fragment like | Normalized Reward34 | 10 | |
| Molecular Property Optimization | gpcr like | Normalized Reward54 | 10 | |
| Molecular Property Optimization | kinase like | Normalized Reward57 | 10 | |
| Molecular Property Optimization | low mw high potency | Normalized Reward0.45 | 10 | |
| Molecular Property Optimization | QED maximize | Normalized Reward62 | 10 | |
| Molecular Property Optimization | oral bioavailable | Normalized Reward0.36 | 10 | |
| Molecular Property Optimization | soluble permeable | Normalized Reward0.39 | 10 | |
| Molecular Property Optimization | easy synth druglike | Normalized Reward0.62 | 10 |