Spreading a fixed rollout budget evenly across prompts wastes it, and allocating by hit utility improved Pass@K over compute-matched GRPO at no extra cost
AspireAI community · arXiv
Group-based policy optimization methods such as GRPO typically hand every prompt the same fixed number of rollouts, which over-allocates compute to prompts whose sampled groups are already saturated while under-exploring prompts where more samples might reveal a correct trajectory. This paper defines hit utility as the posterior probability that at least one rollout in a proposed additional allocation for a prompt will be correct, and builds a learning-free allocation policy on top of it that maximizes total posterior hit utility within each allocation batch while leaving the downstream reward evaluation and the group-based advantage estimator untouched. Across two training sets, four mathematical reasoning benchmarks and three model families at multiple scales, the method preserves comparable Pass@1 and improves Pass@K over compute-matched GRPO, and it outperforms related allocation baselines at large K; it is also drop-in compatible with other group-based estimators such as RLOO. The ablations are the reassuring part: the uniform prior the policy falls back on is competitive with five prompt-conditioned learned-prior alternatives, which means the gain does not depend on training another model. The limitation is that it targets group-based RLVR, where a correct trajectory can actually be sampled, and the gains show up mainly at large K rather than in the single-sample Pass@1 most pipelines report.
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.