WebAnchor: Anchoring Agent Planning to Stabilize Long-Horizon Web Reasoning

TL;DR

This paper introduces WebAnchor, a novel RL framework that stabilizes long-horizon web reasoning by anchoring planning to initial steps, significantly improving performance across multiple benchmarks and model sizes.

Contribution

We propose Anchor-GRPO, a two-stage RL method that decouples planning from execution, enhancing long-horizon web reasoning in LLM-based agents.

Findings

Outperforms baseline RL methods on four web reasoning benchmarks.

Achieves 46.0% pass@1 on BrowseComp with WebAnchor-30B.

Demonstrates scalability with larger models and longer contexts.

Abstract

Large Language Model(LLM)-based agents have shown strong capabilities in web information seeking, with reinforcement learning (RL) becoming a key optimization paradigm. However, planning remains a bottleneck, as existing methods struggle with long-horizon strategies. Our analysis reveals a critical phenomenon, plan anchor, where the first reasoning step disproportionately impacts downstream behavior in long-horizon web reasoning tasks. Current RL algorithms, fail to account for this by uniformly distributing rewards across the trajectory. To address this, we propose Anchor-GRPO, a two-stage RL framework that decouples planning and execution. In Stage 1, the agent optimizes its first-step planning using fine-grained rubrics derived from self-play experiences and human calibration. In Stage 2, execution is aligned with the initial plan through sparse rewards, ensuring stable and efficient tool usage. We evaluate Anchor-GRPO on four benchmarks: BrowseComp, BrowseComp-Zh, GAIA, and XBench-DeepSearch. Across models from 3B to 30B, Anchor-GRPO outperforms baseline GRPO and First-step GRPO, improving task success and tool efficiency. Notably, WebAnchor-30B achieves 46.0% pass@1 on BrowseComp and 76.4% on GAIA. Anchor-GRPO also demonstrates strong scalability, getting higher accuracy as model size and context length increase.