ForkPilot improves long-task AI decisions cutting token use nearly in half
ForkPilot: Self-Evolving Policy for Retrospective Search in Long-Horizon Agents
Artificial Intelligence
Summary
Sometimes AI agents make mistakes when doing tasks that take many steps, because early errors build up later. The authors found that looking back at past decisions to fix mistakes is tricky because it’s hard to tell which earlier steps caused problems, and past experience can become outdated. They created ForkPilot, a new method that learns when to look back and fix earlier steps by training itself on past attempts and improving over time. In tests across many different AI systems and tasks, ForkPilot performed as well as top methods but used much less computing effort.
What this means in practice
- •For ai system engineers: Reduce computational costs by making smarter search decisions when AI agents perform long tasks, improving efficiency without losing accuracy.
- •For customer service platform developers: Improve AI assistants that handle complex customer requests by enabling them to correct errors during multi-step conversations more effectively.
Authors
Xinyue Zeng, Shivam Shandilya, Guilherme Potje, Leonardo Nunes, Rakshanda Agarwal, Ranveer Chandra, Emre Kiciman, Dawei Zhou, Tusher Chakraborty
Abstract
Interactive language-model agents increasingly solve complex tasks through long-horizon, multi-call reasoning, where errors in beliefs or actions can compound across tool interactions. Retrospective search can recover from such failures but is prone to misallocation. Delayed outcomes obscure the contribution of intermediate search decisions, leading to Attribution Complexity, while evolving execution evidence leads to Adaptation Complexity, where previously learned estimates become stale. To address these challenges, we first introduce Search Value Dynamics (SVD), which characterizes the evolving trade-off between the gain and cost of retrospective search. Building on SVD, we propose ForkPilot, a self-evolving two-stage policy-learning framework. In the first stage, ForkPilot learns a search-value policy offline from completed trajectories through automatically constructed outcome comparisons. In the second stage, it makes search decisions based on current observations and then self-evolves by incorporating newly completed trajectories into subsequent policy updates. We evaluate ForkPilot across 6 diverse benchmarks and 7 widely used LLM backbone families, including four open-source families, GPT-5.6 Sol, and Opus 4.8 in a production agentic system, against 9 competitive baselines, including a real-world harness deployment used by hundreds of thousands of paid users. ForkPilot achieves comparable state-of-the-art performance while reducing token usage by up to 59.2%, demonstrating its efficacy.