rlxp-reward-review

Reviewing reinforcement-learning reward changes involves checking proposed parameter tweaks and code edits for task alignment, privileged signals, and saturation risks before running experiments to prevent reward hacking and unstable training behavior.

Reward hacking in RL experiment workflows occurs when a reward function exploits unintended signals or guardrails, causing misaligned objectives; reviewing reward changes before approval flags these privileged signals and discontinuities to maintain task alignment.

Assessing approval readiness for a reward function change requires evaluating task alignment, scale compatibility, and signal availability against the task objective and metric spec, producing a risk assessment and suggested fixes before validation.

Yes, checking task alignment for reward parameter tweaks before training requires the study objective and expected success criteria to verify the proposed reward change matches the intended task outcome and avoids misaligned objectives.

When editing reward code, look for guardrail conflicts involving privileged training-time signals, discontinuities, and saturation that threaten safety and stability, using available training-time signals and reward reflection context to screen for risks.

Unstable training behavior after a reward change often stems from scale incompatibility or discontinuities in the reward function, requiring a conservative review of signal availability and task alignment to identify and fix design risks.