Curiosity-driven RL for symbolic equation solving