SIRI Trains AI Agents to Absorb Skills Then Run Without Them

Training LLM agents to handle long, multi-step tasks is hard enough. Keeping them fast and cheap at inference time is a separate battle entirely. A new paper introduces SIRI (Self-Internalizing Reinforcement Learning with Intrinsic Skills), a three-phase framework that lets agents discover and absorb reusable behavioral skills during training, then discard the scaffolding entirely before deployment. The result: a leaner agent that runs on the original prompt alone, with no external lookups, no skill banks, no added latency.

The skill bank trap

LLM agents are increasingly deployed on long-horizon, open-ended tasks where the ability to accumulate and reuse procedural knowledge across episodes determines downstream performance. A widely adopted abstraction is the skill bank: a retrievable collection of distilled, reusable textual or executable artifacts that encode how the agent performs recurring tasks. Think of it like a recipe book the agent consults mid-task.

The problem is that this recipe book has to live somewhere. One line of work equips the agent with a persistent memory of past interaction traces, retrieved at inference time as references for similar future tasks. That means every deployed agent needs an external retrieval system running alongside it, inflating context length, engineering complexity, and response latency. Skill retrieval itself is now a major bottleneck in realistic tool ecosystems.

The deeper issue is credit assignment. Scalability to long-horizon LLM agent training remains limited. Unlike static tasks, agent-environment interactions unfold over many steps and often yield sparse or delayed rewards, making credit assignment across individual steps significantly more challenging. You only find out at the end of a 50-step episode whether the agent succeeded, so figuring out which early actions actually mattered is genuinely hard.