| name | multi-modal-multi-environment-machine-teaching-robust-reward |
| description | Hierarchical machine teaching algorithm for robust reward learning across multiple MDPs. Demonstrates comparisons impose stronger constraints than demonstrations in unlimited-data regime. Greedily selects informative environments then queries low-cost feedback. Activation: machine teaching, reward learning, inverse reinforcement learning, multi-environment, robust reward, feedback modalities. |
| metadata | {"arxiv_id":"2607.08647","published":"2026-07-09","authors":"Ali Larian, Qian Lin, Chang Zong Wu, Daniel S. Brown","tags":["machine-teaching","reward-learning","inverse-reinforcement-learning","multi-environment","robust-reward"]} |
Multi-Modal, Multi-Environment Machine Teaching for Robust Reward Learning
Overview
As autonomous agents are increasingly deployed across diverse operational contexts, aligning their behavior with human intent demands reward functions that remain robust to environmental changes. This paper introduces a hierarchical machine teaching algorithm for reward learning that operates across multiple MDPs, addressing the limitation that demonstrations in one MDP entangle reward information with environment-specific structure.
Key Innovations
Theoretical Analysis of Feedback Modalities
- Shows that in the unlimited-data regime, comparisons impose strictly stronger global constraints than other modalities
- Different feedback modalities constrain rewards differently
- Demonstrations in one MDP entangle reward with environment structure, causing generalization failure
Hierarchical Machine Teaching
- First selects informative environments that expose complementary reward constraints
- Then strategically queries low-cost feedback within those environments
- Greedy environment selection based on complementary information
Multi-Environment Robustness
- Rewards learned across multiple environments generalize better than single-environment rewards
- Substantially lower regret on held-out environments under identical feedback budgets
- Demonstrates importance of multi-environment teaching for dynamics-robust reward functions
Methodology
- Environment Selection: Greedily select MDPs exposing complementary reward constraints
- Feedback Querying: Strategically query low-cost feedback within selected environments
- Reward Learning: Learn reward functions that generalize across environments
- Evaluation: Compare against uniform teaching baselines on held-out environments
Implications
- Multi-environment teaching is essential for robust reward learning
- Feedback modality choice significantly impacts reward constraint quality
- Hierarchical approach enables efficient use of feedback budgets
- Applicable to autonomous agent deployment across diverse contexts
Pitfalls
- Requires access to multiple environments during training
- Greedy environment selection may miss globally optimal combinations
- Low-cost feedback may not always be informative
- Theoretical results assume unlimited-data regime which may not hold in practice
Activation Keywords
machine teaching, reward learning, inverse reinforcement learning, multi-environment, robust reward functions, feedback modalities, hierarchical teaching, MDP generalization
Paper Reference
arXiv:2607.08647 - "Multi-Modal, Multi-Environment Machine Teaching for Robust Reward Learning" (Jul 2026)