Back to feed
arXiv cs.LG
arXiv cs.LG
7/30/2026
Early Verdicts, Better Budgets: Sequential Adaptive Rollout Allocation for Compute-Efficient RLVR

Early Verdicts, Better Budgets: Sequential Adaptive Rollout Allocation for Compute-Efficient RLVR

Short summary

SARA casts per-step rollout collection in RLVR as a budget-constrained sequential allocation problem, using Beta posteriors and an SPRT-style rule to commit effective groups early and abandon saturated ones. It reallocates freed budget to fresh prompts without extra prediction rollouts. On math reasoning and planning with 1.5B/3B models, SARA matches dynamic sampling accuracy while using 22% fewer rollouts; composed with dynamic sampling it achieves best accuracy at 67% fewer rollouts.

  • •SARA uses Beta posteriors and SPRT-style thresholds to decide early whether a prompt group is effective or saturated
  • •Proves abandonment reliability, expected rollout savings, and fixed-budget yield dominance over baselines
  • •Composing SARA with dynamic sampling yields best accuracy at 67% fewer rollouts than dynamic sampling alone

Generated with AI, which can make mistakes.

#research-breakthrough#ai-tools
Read full article at arXiv cs.LG

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

SLPO: Scaling Latent Reasoning via a Surrogate Policy
arXiv cs.CL
Process Reward Informed Tree Rollout for Effective Multi-Turn RL
arXiv cs.CL
BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards
arXiv CS.AI
Uncertainty-Aware Sequential Decision Rules for Event-Triggered LLM Invocation in Streaming Systems
arXiv cs.LG