# Interactive Explanations: Diagnosis and Repair of Reinforcement Learning Based Agent Behaviors

arXiv: https://arxiv.org/abs/2105.12938 · HTML: https://mlsanity.com/paper/2105.12938/

## Key facts

- Cited by 8 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 27 May 2021.

## Built on by (most cited first, 8 of 8)

- [Model-Agnostic Policy Explanations with Large Language Models](https://mlsanity.com/paper/2504.05625/) (2025)
- [Explaining Agent Behavior with Large Language Models](https://mlsanity.com/paper/2309.10346/) (2023)
- [Understanding Your Agent: Leveraging Large Language Models for Behavior Explanation](https://mlsanity.com/paper/2311.18062/) (2023)
- [Towards Interactive Reinforcement Learning with Intrinsic Feedback](https://mlsanity.com/paper/2112.01575/) (2021)
- [Interactive Explanations for Reinforcement-Learning Agents](https://mlsanity.com/paper/2504.05393/) (2025)
- [Optimists at Heart: Why Do We Research Game AI? (Extended Version)](https://mlsanity.com/paper/2206.03261/) (2022)
- [ASQ-IT: Interactive Explanations for Reinforcement-Learning Agents](https://mlsanity.com/paper/2301.09941/) (2023)
- [Evaluating RL Explainability Methods by How Much They Help Fix Bugs in Agents](https://mlsanity.com/paper/2608.17524/) (2026)

## Abstract (arXiv)

Reinforcement learning techniques successfully generate convincing agent behaviors, but it is still difficult to tailor the behavior to align with a user's specific preferences. What is missing is a communication method for the system to explain the behavior and for the user to repair it. In this paper, we present a novel interaction method that uses interactive explanations using templates of natural language as a communication method. The main advantage of this interaction method is that it enables a two-way communication channel between users and the agent; the bot can explain its thinking procedure to the users, and the users can communicate their behavior preferences to the bot using the same interactive explanations. In this manner, the thinking procedure of the bot is transparent, and users can provide corrections to the bot that include a suggested action to take, a goal to achieve, and the reasons behind these decisions. We tested our proposed method in a clone of the video game named textitSuper Mario Bros., and the results demonstrate that our interactive explanation approach is effective at diagnosing and repairing bot behaviors.