SIRI: Self-Internalizing Reinforcement Learning with Intrinsic Skills for LLM Agent Training

Researchers introduced SIRI, a three-stage RL framework for LLM agents that discovers and internalizes useful skills during training.

It pretrains with GiGPO, mines skills from the model’s own successful rollouts, validates them with pairwise comparisons, and distills only helpful actions into the base policy.

On ALFWorld and WebShop, SIRI improved a Qwen2.5-7B-Instruct agent over GiGPO and outperformed several prompt-based, RL, and memory-augmented baselines.

The approach supports reusable skills for long-horizon tasks without external skill generators or inference-time memory banks, lowering deployment complexity and latency.