BaNEL: Exploration Posteriors for Generative Modeling Using Only Negative Rewards
This addresses a critical bottleneck in generative modeling for sparse-reward tasks, offering a novel approach to post-training without successful samples, which is incremental but impactful for specific domains like reinforcement learning or robotics.
The paper tackles the problem of improving generative models when reward signals are near-zero and reward evaluations are expensive, by proposing BaNEL, which uses only negative rewards from failed attempts to steer generation away from failures, achieving up to several orders of magnitude higher success rates than existing methods.
Today's generative models thrive with large amounts of supervised data and informative reward functions characterizing the quality of the generation. They work under the assumptions that the supervised data provides knowledge to pre-train the model, and the reward function provides dense information about how to further improve the generation quality and correctness. However, in the hardest instances of important problems, two problems arise: (1) the base generative model attains a near-zero reward signal, and (2) calls to the reward oracle are expensive. This setting poses a fundamentally different learning challenge than standard reward-based post-training. To address this, we propose BaNEL (Bayesian Negative Evidence Learning), an algorithm that post-trains the model using failed attempts only, while minimizing the number of reward evaluations (NREs). Our method is based on the idea that the problem of learning regularities underlying failures can be cast as another, in-loop generative modeling problem. We then leverage this model to assess whether new data resembles previously seen failures and steer the generation away from them. We show that BaNEL can improve model performance without observing a single successful sample on several sparse-reward tasks, outperforming existing novelty-bonus approaches by up to several orders of magnitude in success rate, while using fewer reward evaluations.