Reward Assignment without Language: Exploring Non-Linguistic Intrinsic Motivation in Self-Evolving Agents

Research Question: How effective are non-linguistic, sensorimotor intrinsic rewards (e.g., novelty, surprise, empowerment) at driving the self-attribution and learning efficiency of self-evolving agents compared to LLM-driven reward assignments?

Hypothesis: Appropriately designed non-linguistic intrinsic motivation signals can match or even outperform LLM-based differentiated rewards in certain environments, especially those with high sensory complexity or limited language cues.

Experiment Plan: - Implement an alternative reward pipeline in AgentEvolver using sensorimotor-based intrinsic motivation (e.g., novelty search, empowerment).

• Run parallel experiments in both language-rich and sensorimotor-rich environments.
• Compare exploration efficiency, sample utilization, and adaptability to that of the original AgentEvolver with LLM-based reward assignment.
• Analyze which types of environments benefit most from each reward strategy.

References:

• Zhai, Y. et al. (2025). AgentEvolver: Towards Efficient Self-Evolving Agent System.
• Pternea, M., Singh, P., Chakraborty, A., Oruganti, Y., Milletarí, M., Bapat, S., & Jiang, K. (2024). The RL/LLM Taxonomy Tree: Reviewing Synergies Between Reinforcement Learning and Large Language Models. Journal of Artificial Intelligence Research.