Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
Yue et al., 2025, Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (arXiv preprint arXiv:2504.13837)
Subscriber deck — free during early access
1 cards in “3. RLVR’s Effect on Reasoning Capacity Boundary” · whole deck
3. RLVR’s Effect on Reasoning Capacity Boundary
- What is Yue et al.’s main finding about pass@k curves?Across model families, math, code and visual reasoning benchmarks, and RL algorithms, RLVR models win at small \(k\) (e.g. pass@1), but as \(k\) reaches the tens or hundreds the base models catch up and overtake them: RLVR narrows the set of solvable problems rather than widening it.
pass@k curves cross: a simulation of 400 problems, not data from the paper - Yue et al., 2025, Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? (arXiv preprint arXiv:2504.13837), 3. RLVR’s Effect on Reasoning Capacity Boundary
Source
- Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Yang Yue, Shiji Song and Gao Huang. 2025. Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?. In arXiv preprint arXiv:2504.13837. https://arxiv.org/abs/2504.13837