A multi-domain RL corpus and training study showing RL elicits pretrained knowledge in math, code, and science but appears to acquire new skills in logic, simulation, and tabular reasoning.
It then iterates through each contest, checking if the rating update is allowed based on the division and the current rating
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
-
Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective
A multi-domain RL corpus and training study showing RL elicits pretrained knowledge in math, code, and science but appears to acquire new skills in logic, simulation, and tabular reasoning.