Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Tell me why! Explanations support learning relational and causal structure
Andrew K Lampinen, Nicholas Roy, Ishita Dasgupta, Stephanie Cy Chan, Allison Tam, James Mcclelland, Chen Yan, Adam Santoro, Neil C Rabinowitz, Jane Wang, Felix Hill
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 property and reward explanation prediction targets | IMPALA RL agent · Perceptual odd-one-out 2D | performance (% correct): 61.9\pm 2.2 -> 91.3 \pm 0.7 | Improved |
| Add property and reward explanation prediction targets | IMPALA RL agent · Perceptual odd-one-out 3D | performance (% correct): 29.5\pm 0.7 -> 92.7 \pm 1.4 | Improved |
| Add single-dimension color explanations | IMPALA RL agent · Deconfounding: choose color | performance (% correct): 55.4 \pm 2.6 -> 95.5\pm 0.9 | Improved |
| Add single-dimension shape explanations | IMPALA RL agent · Deconfounding: choose shape | performance (% correct): 24.2\pm 7.6 -> 87.5\pm 2.9 | Improved |
| Add single-dimension texture explanations | IMPALA RL agent · Deconfounding: choose texture | performance (% correct): 15.4\pm 6.7 -> 86.2\pm 0.9 | Improved |
| Add behavior- and context-relevant explanation prediction | IMPALA RL agent · Meta-learning easy levels | performance (% correct): 24.0\pm 0.6 -> 96.9 \pm 0.3 | Improved |
| Add behavior- and context-relevant explanation prediction | IMPALA RL agent · Meta-learning hard levels | performance (% correct): 24.6\pm 1.2 -> 90.5 \pm 1.6 | Improved |
| Replace prediction targets with explanations provided as inputs (ablation) | IMPALA RL agent · Odd-one-out tasks | performance | Hurt |
| Add both property and reward explanations instead of using either type alone (ablation) | IMPALA RL agent · Perceptual odd-one-out 2D | learning speed | Improved |
| Use reward explanations rather than property explanations alone (ablation) | IMPALA RL agent · Meta-learning hard levels | learning | Improved |
| Use behavior- and context-relevant explanations instead of behavior-irrelevant explanations (ablation) | IMPALA RL agent · Meta-learning easy and hard levels | learning | Improved |
| Use behavior- and context-relevant explanations instead of context-irrelevant explanations (ablation) | IMPALA RL agent · Odd-one-out tasks | learning | Improved |
Closest papers by meaning
- Explainable Reinforcement Learning through a Causal Lens (AAAI 2020 · similarity 0.89)
- Inherently Explainable Reinforcement Learning in Natural Language (NeurIPS 2022 · similarity 0.89)
- Ground-Compose-Reinforce: Grounding Language in Agentic Behaviours using Limited Data (NeurIPS 2025 · similarity 0.86)
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models (ICML 2026 · similarity 0.86)
- Language Agents Meet Causality -- Bridging LLMs and Causal World Models (ICLR 2025 · similarity 0.86)
- Learning to Model the World With Language (ICML 2024 · similarity 0.85)
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.
