Research map / Multi-agent systems and games / Human-AI and goal-conditioned
State2Explanation: Concept-Based Explanations to Benefit Agent Learning and User Understanding
Devleena Das, Sonia Chernova, Been Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add S2E concept-based reward shaping to sparse-reward MuZero | MuZero · Connect 4 | training samples to similar Win%: ~200k less training samples | Improved |
| replace expert-defined reward shaping with S2E-informed shaping | MuZero · Lunar Lander | Reward and Win% | No clear effect |
| 13.9% incorrect explanation retrievals vs expert shaping upper bound (ablation) | MuZero+S2E-RS · Connect 4 | learning rate: 13.9\% | No clear effect |
| show S2E concept explanation instead of no explanation | Connect 4 users · ATS improvement | t-test: t(70) = 3.08 | Improved |
| show S2E concept explanation instead of action explanation | Connect 4 users · ATS improvement | t-test: t(70)=-3.25 | Improved |
| show S2E concept explanation instead of value explanation | Connect 4 users · ATS improvement | t-test: t(70)=4.03 | Improved |
| S2E-retrieved explanation vs ground-truth concept explanation (ablation) | Connect 4 users · ATS improvement | ATS | No clear effect |
| add InF and TeG abstraction to S2E concept explanations vs none | Lunar Lander users · ATS improvement | t-test: t(98) = 3.15 | Improved |
| add InF and TeG abstraction vs unabstracted concept explanation (ablation) | Lunar Lander users · ATS improvement | t-test: t(98)=-3.19 | Improved |
| add InF and TeG vs value-based explanation | Lunar Lander users · ATS improvement | t-test: t(98)=3.59 | Improved |
| use unabstracted S2E concept explanation in high-frequency Lunar Lander (ablation) | Lunar Lander users · ATS improvement | ATS | No clear effect |
| InF thresholds x-position 0.15, tilt 0.01 and -0.05 sensitivity (ablation) | InF · Lunar Lander | fraction of concepts filtered | No clear effect |
Closest papers by meaning
- Explaining Reinforcement Learning Agents through Counterfactual Action Outcomes (AAAI 2024 · similarity 0.88)
- Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.87)
- Understanding Individual Agent Importance in Multi-Agent System via Counterfactual Reasoning (AAAI 2025 · similarity 0.87)
- Utilizing Human Behavior Modeling to Manipulate Explanations in AI-Assisted Decision Making: The Good, the Bad, and the Scary (NeurIPS 2024 · similarity 0.86)
- Interpreting Emergent Planning in Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.86)
- LICORICE: Label-Efficient Concept-Based Interpretable Reinforcement Learning (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
