By quantifying exploration via Effective Rank in semantic hidden-state space, the work shows near-zero correlation with exploitation dynamics and introduces VERL to improve RLVR performance on LLM reasoning tasks.
fpER finalpmqq “fpm Kqfor allmPR K.(50) Fix any constantcPR. Consider the affine subspace Ac :“ tmPR K :m K “cu.(51) On this subspace,ER finalpmq
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Semantic-Space Exploration and Exploitation in RLVR for LLM Reasoning
By quantifying exploration via Effective Rank in semantic hidden-state space, the work shows near-zero correlation with exploitation dynamics and introduces VERL to improve RLVR performance on LLM reasoning tasks.