skills.rest
junhyekh
junhyekhCOMMUNITY
Explore all skills by this author →
GitHub

rlxp-reward-engineering

Design minimal reward logic patches for reinforcement-learning tasks.

1|Updated May 14, 2026
Search Tags:#reinforcement learning#evidence gating#reward engineering#reward hacking#metric divergence#patch drafting
One-click install
npx skills add https://github.com/junhyekh/rlxp --skill rlxp-reward-engineering
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: rlxp-reward-engineering
Source: https://github.com/junhyekh/rlxp/tree/main/plugins/rl-experiment-assistant/skills/rlxp-reward-engineering
Command: npx skills add https://github.com/junhyekh/rlxp --skill rlxp-reward-engineering

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This Skill helps you design or patch reward logic for one reinforcement-learning task when the current reward terms cannot express the desired behavior and the study contract allows reward-code changes.

Core Features & Use Cases

  • Minimal reward edits: Propose the smallest safe change that adds the missing signal without rewriting the whole objective.
  • Risk-aware review: Check for reward hacking, privileged signals, scale issues, and monitoring gaps before any patch is drafted.
  • Evidence-gated workflow: Use it when reward saturation, sparse behavioral signals, or training-reward divergence suggests scalar tuning is not enough.
  • Use case: A model keeps gaming a proxy reward, so you inspect the reward components, confirm the issue in analyses, and draft a bounded patch plan with sanity checks.

Quick Start

Ask the assistant to review the active study’s reward code, analyses, and contract permissions, then draft the smallest safe reward change for the current RL task.

Frequently Asked Questions about rlxp-reward-engineering

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I patch reinforcement learning rewards when the model keeps gaming a proxy metric?▼

To patch reinforcement learning rewards safely, this Skill inspects existing reward components, confirms the reward hacking issue in your analyses, and drafts the smallest bounded change that adds the missing behavioral signal without rewriting the whole objective.

Why does my training reward diverge from task metrics and how can I fix it?▼

Training reward diverges from task metrics when current reward terms cannot express the required behavior, requiring a minimal reward logic patch to reintroduce the missing signal and correct the objective alignment.

What is the best way to add a missing behavioral signal to an RL reward function?▼

The best way to add a missing behavioral signal is through an evidence-gated workflow that requires your reward code, task card, metric spec, latest analyses, and explicit permission before proposing a minimal safe patch.

How do I check for reward saturation and scale issues before modifying an RL objective?▼

Checking for reward saturation and scale issues requires a risk-aware review of your reward components and training analyses to identify monitoring gaps and privileged signals before drafting any reward logic patches.

Can I use this workflow to completely rewrite a reinforcement learning objective?▼

No, this workflow designs minimal reward edits rather than complete rewrites, proposing the smallest safe change to address reward hacking or sparse signals while maintaining the existing objective structure.

When do I need to provide a documented exception to draft reward patches?▼

You need to provide a documented exception or reward reflection alongside your task card and metric spec when standard scalar tuning is insufficient and your study contract explicitly permits reward-code changes.

Related Skills

Explore More Software Engineering Skills→
junhyekhjunhyekh
1

rlxp-reward-review

Reviews reinforcement-learning reward changes for task alignment and approval readiness.

Community
Intermediate
attentiondotnetattentiondotnet

grpo-rl-training

Implement GRPO/TRL reinforcement learning fine-tuning workflows for language models.

Community
Advanced
rocklambrosrocklambros
2

auditing-rlhf-reward-hacking

Audit RLHF, DPO, and RLAIF checkpoints for reward hacking by comparing reward win-rates with preference win-rates.

Community
Advanced
Simon-Copilot-StudioSimon-Copilot-Studio

grpo-rl-training

Facilitate end-to-end GRPO/RL fine-tuning with TRL for language models.

Community
Advanced