What they found
The authors build WATERFALL, a loop that creates many candidate reward programs, using different "personas" to get variety. Vision-language models (SI models that read images and text) watch video of what an agent does and judge it in a Swiss-system tournament, a pairing format where similar performers meet, to pick the best candidates. The history of past judgments then guides which candidates to change, push further, or drop. The authors test it on fully observable and partially observable environments, from discrete to continuous. They report that it beats sparse-reward RL, intrinsic-motivation and foundation-model reward-synthesis baselines, even though those baselines keep their usual information advantages, with the clearest gains in the partially observable case. Ablations show that visual evidence, tournament selection and iterative refinement all matter together.
Why it matters
Designing rewards, the signals that tell an RL agent what counts as success, is a major bottleneck, especially when rewards are rare and tasks are long. Many recent methods quietly rely on task descriptions, demonstrations or ground-truth metrics. This paper tries to remove those aids and keep the setting strict. If it holds, agents could discover useful goals through interaction instead of depending on human-written rewards. The abstract gives no specific numbers, so the size of the gains is not clear from it.
How it changes the race
If the result holds, it could cut the manual reward work needed to train agents in hard, partly hidden environments. That would lower the expertise needed to build such systems. The abstract reports only benchmark environments, so real-world reach is unknown.
Safety and control
The system writes its own reward programs and judges them with vision-language models, so the rewards it learns may not match what humans want. The abstract does not discuss this, so oversight of automatically discovered rewards is an open question.
The abstract, by the authors
Reward design remains a central bottleneck in reinforcement learning (RL), particularly in sparse-reward, long-horizon and partially observable settings requiring dependency chaining, memory, and deceptive affordance disambiguation. While recent foundation-model approaches reduce manual reward engineering, most assume goal-conditioning, task descriptions, expert demonstrations, or ground-truth metrics. These assumptions are problematic in partially observable environments, where task-relevant objectives, affordances, and dynamics must be discovered through interaction rather than disclosed a priori. We formalise this problem setting by introducing WATERFALL, an iterative population-based workflow for automated reward discovery under the strict observability constraints inherent in POMDP environments. WATERFALL achieves this by (i) synthesising diverse programmatic reward candidates via persona-conditioned generation, (ii) evaluating candidate behaviour from raw visual rollouts utilising a Swiss-system evolutionary tournament judged by vision-language models to identify elites, and (iii) leveraging longitudinal assessment histories to iteratively mutate, escalate, or prune candidates. We evaluate WATERFALL across fully observable and reveal-gated partially observable environments ranging from discrete to continuous, comparing against sparse-reward RL, intrinsic-motivation and foundation-model reward-synthesis baselines. Our results indicate that WATERFALL discovers and refines reward programs without access to privileged information, outperforming baselines that retain their native information advantages, most clearly in the POMDP regime. Extensive ablations substantiate that visual evidence, tournament-based selection, and iterative refinement are jointly important. Enforcing this contract is what separates autonomous reward discovery from benchmark-privileged reward shaping.
Quoted from arXiv:2611.22854. The research and its words belong to its authors.
