Safe RL benchmarks usually report average safety, which hides how often and how badly the constraint is violated, and this paper proposes reporting three views together
AspireAI community · arXiv
Safe reinforcement learning is typically formalized as a constrained MDP where an agent maximizes reward while keeping expected cumulative cost under a bound, and existing safe-RL benchmarks predominantly report whether an algorithm is safe on average. This paper argues that expectation-based reporting is not enough to characterize true safety, because it misses how often and how severely the bound is violated, whether that holds across tasks and bounds, and whether training-time behavior reflects the final converged policy. The authors introduce evaluation metrics covering each of those concerns plus aggregation across tasks and bounds, and a safety tier system for categorizing and comparing algorithms on safety and reliability at both training time and for a final policy, along with an empirical evaluation across multiple safe navigation tasks. Their finding is that aggregate metrics, distributional reporting and task-and-bound-specific results each reveal information the others cannot, so they recommend reporting all three jointly rather than compressing them into one number, and they release SafeRLEval as an open-source evaluation suite. The limitation is that the tiers are only as meaningful as the tasks chosen to populate them.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.