A hint-scaffolding RL method that dynamically tunes each problem's hint length to keep rollout accuracy near 50% outperforms static GRPO, SFT, and prior hint-based approaches on math reasoning benchmarks.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Staying in the Sweet Spot: Responsive Reasoning Evolution via Capability-Adaptive Hint Scaffolding
A hint-scaffolding RL method that dynamically tunes each problem's hint length to keep rollout accuracy near 50% outperforms static GRPO, SFT, and prior hint-based approaches on math reasoning benchmarks.