Research map / Reinforcement learning / Visual and representation RL
Robust In-Context Reinforcement Learning Under Reward Poisoning Attacks
Paulius Sasnauskas, Yiğit Yalın, Goran Radanovic
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 adversarial training to DPT | DPT · Bandit (eps=0.4, attacker targets AT-DPT) | cumulative regret: 63.6 -> 24.2 | Improved |
| add adversarial training to DPT | DPT · Bandit (clean environment) | cumulative regret: 11.5 -> 13.0 | Hurt |
| add adversarial training to DPT | DPT · Bandit (eps=0.4, uniform random attack) | cumulative regret: 37.2 -> 38.7 | No clear effect |
| add adversarial training to DPT | DPT · Linear bandit (eps=0.4, attacker targets AT-DPT) | cumulative regret: 70.29 -> 2.49 | Improved |
| add adversarial training to DPT | DPT · Darkroom2 (eps=0.4, attacker targets AT-DPT) | episode reward: 216.1 -> 242.2 | Improved |
| add adversarial training to DPT | DPT · Darkroom2 (clean environment) | episode reward: 306.8 -> 267.4 | Hurt |
| add adversarial training to DPT | DPT · Miniworld (eps=0.4, attacker targets AT-DPT) | episode reward: 81.2 -> 104.8 | Improved |
| train against adaptive attacker instead of non-adaptive (ablation) | AT-DPT · Bandit (eps=0.4, adaptive attacker targeting AT-DPT) | cumulative regret: 88.1 -> 37.1 | Improved |
| train against adaptive attacker instead of non-adaptive (ablation) | AT-DPT · Bandit (eps=0.4, non-adaptive attacker targeting AT-DPT) | cumulative regret: 22.8 -> 38.0 | Hurt |
| train against a strong learned attacker rather than uniform random noise (ablation) | AT-DPT · Bandit (eps=0.4, attacker targets AT-DPT) | cumulative regret: 39.3 -> 24.2 | Improved |
| train with 30% suboptimal expert actions (ablation) | AT-DPT · Bandit (eps=0.4, attacker targets AT-DPT) | cumulative regret: 24.2 -> 41.2 | Hurt |
| train with 10% suboptimal expert actions (ablation) | AT-DPT · Darkroom2 (eps=0.4, attacker targets AT-DPT) | episode reward: 242.2 -> 225.2 | Hurt |
Closest papers by meaning
- Robust Policy Gradient against Strong Data Corruption (ICML 2021 · similarity 0.90)
- Beyond Worst-case Attacks: Robust RL with Adaptive Defense via Non-dominated Policies (ICLR 2024 · similarity 0.87)
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.87)
- Towards Robust Model-Based Reinforcement Learning Against Adversarial Corruption (ICML 2024 · similarity 0.87)
- Multi-level Certified Defense Against Poisoning Attacks in Offline Reinforcement Learning (ICLR 2025 · similarity 0.87)
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.87)
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.
