bf16 Rounding Silently Discards Most RL Policy Updates: A Hidden Training-Inference Mismatch and an Unbiased One-Line Fix
cs.LG, cs.AI
Submitted: 2026-10-29
Updated: 2026-10-29
Related papers
- Power Distribution Bridges Sampling, Self-Reward RL, and Self-Distillation
- Train Separately, Merge Together: Modular Post-Training with Mixture-of-Experts
- Democratizing Tool Learning with Environments Fully Simulated by a Free 8B Language Model
- D-STEER: Two-Way Control of DPO-Induced Safety Behavior in Activation Space
- Optimal hypersurface decision trees
- Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation