This proposal is intended to aid in finishing the required experiments for two papers to be submitted to ICLR 2026 by the same three authors. Both papers are on the subject of exploration in Reinforcement Learning. As a doctoral student, my primary supervisor and third author on both papers is Amy Loutfi. The two methods in question are GMC and CALOR whose preliminary abstracts are outlined below GMC - Measuring learning progress is essential for curiosity-driven exploration in reinforcement learning, but widely used signals such as prediction error often fail to distinguish meaningful, learnable patterns from random noise. This paper proposes Gradient-Momentum Coupling (GMC), a parameter-space proxy for learning progress that scores each sample's contribution to ongoing parameter changes through its per-parameter normalized absolute product with the momentum from previous gradients. By leveraging momentum's natural filtering of noise and oscillations, GMC identifies samples that contribute to ongoing parameter updates. Controlled experiments demonstrate noise robustness and emergent curriculum learning, with the signal prioritizing tasks by learning speed rather than difficulty. On four unmodified MiniGrid MultiRoom navigation tasks, computing ICM's intrinsic reward with GMC instead of prediction error recovers exploration that ICM loses to unpredictable observations, and it is the only method that solves the largest task. CALOR - We study reward-free reinforcement learning on Craftax-Classic, which matches the original Crafter mechanics, with all environment rewards hidden from the agent and used only for evaluation. We introduce CALOR (CALibrated ORdering), an intrinsic-motivation method that rewards the agent for visiting states that are rare under its own current visitation distribution. CALOR starts from the observation that maximizing state entropy calls for a reward of the form -log d, where d is the visitation density, but that estimating d in a learned latent space is unreliable because the encoder warps the geometry in which the density would be measured. CALOR therefore discards the magnitude of latent distances and keeps only their ordering. A predictor network is trained to rank observations by their within-batch nearest-neighbor distance, and simultaneously to place its own outputs on a calibrated Pareto scale, so that the resulting reward is a proper negative log tail mass rather than an arbitrary function of the encoder metric. Because the calibration is defined against the agent's current experience, the intrinsic reward keeps a fixed scale and never decays as behavior is mastered. On unsupervised Craftax-Classic, CALOR climbs the achievement tech tree well past the point where prior intrinsic-motivation baselines plateau.