RewardGuard – detect reward hacking in RL training loops
FreeFreeFree tier
About RewardGuard – detect reward hacking in RL training loops
RewardGuard is an open-source AI alignment and safety tool that detects reward hacking, component imbalance, and starvation in reinforcement learning training loops. It provides plug-and-play reward monitoring via a single .step() call, offering balance reports, auto weight correction, alignment scores, and native integrations with WandB, TensorBoard, and Stable-Baselines3. The free version delivers read-only analysis and diagnostics, while the premium version adds automatic reward rebalancing, live monitoring, and guardrails against reward hacking.
Key Features
Reward distribution analysis
Imbalance detection and warnings
Training diagnostics and trend monitoring
Actionable recommendations for fixing imbalances
Auto-adjustment of reward weights (Premium version)
Live monitoring during training (Premium version)
Guardrails against reward hacking (Premium version)
Integrations with WandB, TensorBoard, and Stable-Baselines3
Plug-and-play single `.step()` call integration
Pros & Cons
Pros
- Plug-and-play integration with minimal code changes
- Detects reward hacking, imbalance, and starvation early
- Provides clear, actionable reports with recommended weight adjustments
- Free version is fully open-source and available via pip
- Integrates with popular training tools (WandB, TensorBoard, SB3)
Cons
- Free version is read-only; does not modify training behavior
- Premium version is private and requires authentication token
- Limited to reward-based RL loops; not applicable to other RL settings
- Documentation and community support are still maturing
Best For
Detecting reward hacking in RL training loopsEnsuring balanced multi-component reward functionsMonitoring training degradation and misalignment earlyAutomatically rebalancing rewards during training (Premium)Auditing RL systems for AI alignment and safety
FAQ
What is RewardGuard?
RewardGuard is an AI alignment and safety tool that analyzes RL training logs to detect reward hacking, component imbalance, and starvation, providing actionable insights and automatic corrections.
How does the free version differ from the premium version?
The free version provides read-only analysis, imbalance detection, and reports. The premium version adds automatic reward rebalancing, live monitoring, and guardrails against reward hacking.
Does RewardGuard modify my training code?
The free version does not modify training behavior; it only analyzes logs. The premium version can automatically rebalance rewards during training.
What integrations does RewardGuard support?
RewardGuard integrates with Weights & Biases (WandB), TensorBoard, and Stable-Baselines3 (SB3) out of the box.
How do I install RewardGuard?
The free version is installed via `pip install rewardguard`. The premium version requires `pip install rewardguard-premium --index-url private-registry-url` with an authentication token.