A new preprint on arXiv proposes a different way to look at LLM agent behavior: instead of treating a rollout as a sequence of steps, the authors model it as a path through a shared embedding space that includes both the agent's actions and the environment's observations. This reframing allows them to compare trajectories geometrically rather than only textually.

The central finding is that for a given task, successful and failed rollouts separate in this space. The separation is clear enough that the representation becomes outcome-predictive—meaning the trajectory's shape or position can indicate whether the agent is on a path to success or failure. The abstract notes the separation holds with simple methods, though the full details are not yet available in the public summary.

This work could lead to better monitoring and steering of LLM agents during execution, as well as new ways to learn from failed trajectories. The authors frame it as a step toward understanding why some agentic paths diverge from successful ones, potentially enabling more robust decision-making in complex environments.