TL;DR
This guide develops Interpreting RL Policies: What the Model Learned vs What You Rewarded from its central engineering problem to practical design choices and limitations. It synthesizes the roadmap topics without claiming new experimental results. Reported findings belong to the cited sources.
Target searches: “interpretability RL”, “understanding what RL model learned”
Why This Topic Matters
Modern LLM and agent training turns familiar reinforcement-learning ideas into systems problems: actions may be tool calls, rewards may come from fallible graders, and trajectories may span many partially observed steps. The classical foundation is safe and constrained RL; this article focuses on what changes in the newer setting.
The Core Ideas
1. The gap between the reward you wrote and the behavior you got
2. Extending your explainable-RL post (saliency, VIPER distillation, reward decomposition) into the LLM setting
3. Reading chains of thought — and the finding that the stated reasoning is not reliably the actual reasoning
4. Probing: detecting when a backdoored model is about to behave badly
5. Emergent attention heads during RL post-training — what actually changes in the network
6. Why this matters commercially, not just for safety: if you don’t know what it learned, you don’t know when it will fail
Practical Checklist
- Define the state, action, termination condition, and reward before choosing an optimizer.
- Log raw trajectories and every reward component, not only the aggregate score.
- Separate training measurements from held-out evaluation and adversarial tests.
- Treat generated code, tool calls, and environment actions as untrusted.
- Record model, framework, benchmark, and grader versions so comparisons remain reproducible.
Limitations
This area changes quickly. Framework support and benchmark leaderboards are time-sensitive, while many recent methods are preprints rather than settled practice. The article describes mechanisms and reported evidence; it does not present an original training run.
Key Learnings
- The gap between the reward you wrote and the behavior you got
- Extending your explainable-RL post (saliency, VIPER distillation, reward decomposition) into the LLM setting
- Reading chains of thought — and the finding that the stated reasoning is not reliably the actual reasoning