Research map / Reinforcement learning / Visual and representation RL
Exploratory Not Explanatory: Counterfactual Analysis of Saliency Maps for Deep Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Translate the Breakout brick configuration horizontally | A2C agent saliency in Breakout · Breakout | saliency on the tunnel | Mixed |
| Move the ball, paddle, and bricks to the left together | A2C agent saliency in Breakout · Breakout | saliency pattern around the tunnel | No clear effect |
| Intervene on the displayed score with non-ordered score changes | A2C agent in Amidar · Amidar | reward | Hurt |
| Intervene on displayed score using fixed or decremented scores | A2C agent in Amidar · Amidar | object saliency intensity | No clear effect |
| Change enemy distance from the player | Object saliency in the Amidar A2C agent · Amidar | saliency: -0.001 slope; -0.001 r | No clear effect |
| Change enemy distance from the player | Perturbation saliency in the Amidar A2C agent · Amidar | saliency: -0.374 slope; -0.126 r | No clear effect |
| Change enemy distance from the player | Jacobian saliency in the Amidar A2C agent · Amidar | saliency: 0 slope; 0 r | No clear effect |
| Use score interventions that alter rewards and performance | Object saliency in the Amidar A2C agent · Amidar | correlation between reward and saliency differences: 0.274 (intermittent_reset); 0.041 (fixed) | No clear effect |
| Use score interventions that alter rewards and performance | Perturbation saliency in the Amidar A2C agent · Amidar | correlation between reward and saliency differences: -0.10 (intermittent_reset); -0.02 (random_varying) | No clear effect |
| Use score interventions that alter rewards and performance | Jacobian saliency in the Amidar A2C agent · Amidar | correlation between reward and saliency differences: 0.149 (intermittent_reset); 0.281 (random_varying) | No clear effect |
Closest papers by meaning
- Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- Towards Efficient Online Tuning of VLM Agents via Counterfactual Soft Reinforcement Learning (ICML 2025 · similarity 0.85)
- Explaining Reinforcement Learning Agents through Counterfactual Action Outcomes (AAAI 2024 · similarity 0.85)
- Saliency-R1: Enforcing Interpretable and Faithful Vision-language Reasoning via Saliency-map Alignment Reward (CVPR 2026 · similarity 0.85)
- A Simple Framework for Generalization in Visual RL under Dynamic Scene Perturbations (NeurIPS 2024 · similarity 0.85)
- Synthesizing Multimodal Verifiable Game Data to Boost VLMs' General Reasoning (ICLR 2026 · similarity 0.84)
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.