PIVOT: New framework boosts LLM agent planning by 94% with 5x fewer tokens

Large language model (LLM)-based agents can generate plans that look coherent but fail in execution due to infeasible actions, constraint violations, or compounding errors over long horizons. A new paper from researchers including Tuo Zhang and Dimitrios Dimitriadis introduces PIVOT (Plan-Inspect-eVOlve Trajectories), a framework designed to bridge this plan-execution gap. PIVOT refines agent trajectories iteratively through a four-stage loop: PLAN generates candidate trajectories, INSPECT executes them and computes structured losses with textual gradients, EVOLVE applies those signals to produce improved trajectories, and VERIFY performs a final global check against task constraints. A monotonic acceptance process ensures solution quality never degrades.

Evaluated on DeepPlanning and GAIA benchmarks, PIVOT sets new state-of-the-art results. With human-in-the-loop (HITL) feedback, it delivers up to 94% relative improvement in constraint satisfaction. Its fully autonomous variant still yields substantial gains, proving the trajectory-refinement mechanism works without external supervision. Importantly, PIVOT is computationally efficient, requiring 3x to 5x fewer tokens than competing refinement methods. This efficiency makes it practical for real-world deployment. The findings establish that feedback-based trajectory optimization—whether from humans or self-supervision—is a principled methodology for making LLM agents more reliable in autonomous systems.